Received 4 February 2024, accepted 7 April 2024, date of publication 10 April 2024, date of current version 29 April 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3386826

A Comprehensive Systematic Review of YOLO

for Medical Object Detection (2018 to 2023)
MOHAMMED GAMAL RAGAB 1,2 , SAID JADID ABDULKADIR 1,2 , (Senior Member, IEEE),
AMGAD MUNEER 1 , ALAWI ALQUSHAIBI 1,2 , EBRAHIM HAMID SUMIEA 1,2 ,
RIZWAN QURESHI 3 , (Senior Member, IEEE), SAFWAN MAHMOOD AL-SELWI 1,2 ,
AND HITHAM ALHUSSIAN 1,2
1 Department of Computer and Information Sciences, Universiti Teknologi PETRONAS, Seri Iskandar 32610, Malaysia
2 Centre for Research in Data Science, Universiti Teknologi PETRONAS, Seri Iskandar 32610, Malaysia
3 Fast School of Computing, National University of Computer and Emerging Sciences, Karachi 75030, Pakistan

Corresponding author: Rizwan Qureshi ([email protected])

This work was supported by the Ministry of Higher Education (MOHE), Malaysia for providing financial assistance under Fundamental
Research Grant Scheme (FRGS/1/2022/ICT02/UTP/02/4) and Universiti Teknologi PETRONAS under the Yayasan Universiti Teknologi
PETRONAS (YUTP-FRG/015LC0-308) for providing the required facilities to conduct this research work.

ABSTRACT YOLO (You Only Look Once) is an extensively utilized object detection algorithm that
has found applications in various medical object detection tasks. This has been accompanied by the
emergence of numerous novel variants in recent years, such as YOLOv7 and YOLOv8. This study
encompasses a systematic exploration of the PubMed database to identify peer-reviewed articles published
between 2018 and 2023. The search procedure found 124 relevant studies that employed YOLO for
diverse tasks including lesion detection, skin lesion classification, retinal abnormality identification, cardiac
abnormality detection, brain tumor segmentation, and personal protective equipment detection. The findings
demonstrated the effectiveness of YOLO in outperforming alternative existing methods for these tasks.
However, the review also unveiled certain limitations, such as well-balanced and annotated datasets, and the
high computational demands. To conclude, the review highlights the identified research gaps and proposes
future directions for leveraging the potential of YOLO for medical object detection.

INDEX TERMS YOLO, healthcare applications, artificial intelligence, medical object detection, medical
imaging, systematic review.

I. INTRODUCTION process [7]. In recent years, artificial intelligence methods

Object detection is an essential task in computer vision, with have shown great promise for healthcare applications [8],
numerous applications in various domains, including medical [9], [10], [11], [12]. Deep learning-based object detection
imaging [1], [2], surgical procedures [3], and personal protec- algorithms have shown exceptional performance in real time
tive equipment detection [4]. It plays a crucial role in medical object detection and localization [13], [14], [15]. Accurate
imaging by enabling the identification and localization of and efficient object detection algorithms are essential for
abnormalities or objects of interest within medical images [5], assisting healthcare professionals in diagnosing and treating
[6]. Medical diagnosis relies heavily on the accurate detection various medical conditions [16], [17], [18]. One such object
and localization of abnormalities in medical images. The detection algorithm is You Only Look Once (YOLO) [19],
traditional approach to object detection in medical imaging [20], [21], [22], which has gained significant attention.
involves manual annotation and segmentation of the regions YOLO is a state-of-the-art, real time, end-to-end object
of interest, which is a time-consuming and error-prone detection algorithm that has gained significant attention in
the computer vision community [23]. YOLO is a single-stage
The associate editor coordinating the review of this manuscript and detector, which means that it can detect all the objects in an
approving it for publication was Claudio Loconsole . image in a single forward pass through a convolutional neural
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 57815
 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

network (CNN). It uses a single neural network to predict the • Evaluate the performance of YOLO in medical applica-
bounding boxes and class probabilities of the objects present tions by synthesizing its accuracy, precision, recall, and
in an image [24]. This makes YOLO very fast, and it can other relevant metrics as reported in the selected studies.
achieve real time speeds on even a modest GPU [25]. • Identify common challenges, limitations, and gaps in
The application of YOLO in the medical domain has gar- the existing literature on the use of YOLO in medical
nered interest due to its ability to detect and localize anatom- imaging.
ical structures [26], [27], lesions [28], [29], tumors [30], • Provide insights and recommendations for future
[31], [32], and other clinically relevant medical objects [33], research directions, improvements, and potential appli-
[34]. It can detect and localize abnormalities in medical cations of YOLO in the medical domain.
images, which can aid in the early detection and diagnosis The rest of this section contains the following sub-sections:
of various diseases, including breast cancer, lung cancer, You Only Look Once (I-A), YOLO algorithm and architec-
narrowing of blood vessels [35], brain atrophy [36], and ture (I-B), YOLO in action (I-C), image annotation (I-D),
abnormal protein deposits [37], cardiovascular diseases [38], how YOLO operates (I-E), and advantages and drawbacks of
and neurological disorders [39]. The adoption of YOLO YOLO (I-F).
in medical applications has the potential to improve the
accuracy and efficiency of medical diagnosis, which can have A. YOU ONLY LOOK ONCE (YOLO)
a significant impact on patient outcomes. This sub-section provides a brief description of YOLO,
The real time performance of YOLO makes it particu- its versions, structure, and how it works. YOLO, proposed
larly appealing for time-sensitive medical procedures and by Redmon et al. [45], is an object detection algorithm
clinical decision-making [40]. By accurately and efficiently that uses convolutional neural networks (CNN) [46], [47]
identifying objects of interest, YOLO can potentially aid in to detect objects in real time [25]. It is a single-stage
early disease detection, treatment planning, and monitoring method that can achieve real time performance on a standard
of disease progression. However, as the adoption of YOLO GPU [48]. It divides the image into a grid of cells, and
in medical imaging increases, it is essential to evaluate its each cell is responsible for detecting objects within a certain
performance, strengths, limitations, and the specific medical area [49], which allows for faster object detection compared
domains in which it has been applied [41]. Therefore, a sys- to traditional two-stage methods and is particularly useful for
tematic literature review (SLR) may provide a comprehensive real time applications. It has evolved over multiple versions,
and rigorous approach to analyze the existing literature on each offering improvements in speed, detection accuracy, and
YOLO in medical applications. By systematically collecting, capability to detect objects of varying sizes [50].
evaluating, and synthesizing the available evidence, this YOLO exhibits a high level of generalizability, making
review aims to identify the strengths, limitations, and it less prone to failure when applied to novel domains
potential of YOLO in medical applications. The findings of or unexpected situations [23]. Unlike previous approaches
this review will assist researchers, healthcare professionals, that repurpose classifiers for detection, YOLO is a versatile
and developers in understanding the performance of YOLO detector that learns to detect various objects. It acquires
and its suitability for different medical object detection tasks. generalized representations of objects, enabling it to surpass
Survey Motivation: There are already existing review leading detection methods like Deformable Parts Model
articles on YOLO, such as algorithmic developments in (DPM) [51] and Region-based Convolutional Neural Net-
YOLO [25], challenges and architectural developments for work (R-CNN) [52], [53] by a good margin. However, YOLO
object detection using YOLO in [23], and a review on object has some problems with detecting small objects and will
detection techniques in [42]. A survey on object detection for do worse with scenes with many overlapping objects. The
medical images using deep learning techniques was published main advantage of YOLO lies in its real time object detection
in [43] and a comprehensive analysis of applying object capability, which is crucial in time-sensitive applications.
detection methods for medical image analysis in [44]. In this The YOLO architecture has evolved significantly from
paper, we focused on the YOLO architecture, its evolution, its inception in v1 to the cutting-edge advancements in
and applications for three key medical applications; medical v8 as shown in Figure 1. With v1, the initial foundation
images, personal protective equipment detection, and surgical was laid, introducing a groundbreaking concept of real time
procedures. To the best of our knowledge, it is the first article object detection through a single network pass in 2015.
to discuss three key medical applications using the YOLO YOLOv2 (2016): Improves YOLOv5 by using a larger
series architecture. input size, more anchor boxes, and a new loss function.
The main contributions of this paper are: YOLOv3 (2018): Introduces a new network architecture
• Identify and select relevant peer-reviewed articles called Darknet-53, which is deeper and more accurate than
published between 2018 and 2023 that focus on the the previous architectures used in YOLO.
application of YOLO in medical imaging. YOLOv4 (2019): Improves upon YOLOv3 by using a
• Analyze and summarize the characteristics of the new training method called Mosaic data augmentation, which
selected studies, including the medical domains, helps to improve the model’s robustness to different object
datasets, evaluation metrics, and findings. scales and orientations. YOLOv5 (2020): Introduces a new

57816 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

pooling (SPP) module and the path aggregation network

(PAN). The neck’s function is to combine the feature
maps from various layers of the backbone network and
forward them to the head. In YOLO, popular neck
options include Spatial Pyramid Pooling (SPP), Feature
Pyramid Network (FPN), NAS-FPN, and Rep-PAN.
❐ Head: The head component is responsible for handling
the combined features and making predictions regarding
bounding boxes, objectness scores, and classification
scores. In YOLO, a one-stage object detection approach,
like YOLOv3, is employed as the detection head. The
head’s primary role is to generate the ultimate output of
the network, which includes predicted bounding boxes
and class labels. YOLO utilizes various popular heads,
such as Efficient decoupled, Multi-scale, and Anchor-
based detection heads.
The YOLOv1 architecture, Figure 3, is inspired by
FIGURE 1. Timeline of YOLO versions. GoogLeNet and it replaces inception modules with 1 × 1 and
3×3 convolutional layers. This architecture utilizes two fully
network architecture called CSPDarknet, which is more connected layers on the convoluted feature map, outputting a
efficient than Darknet-53. YOLOv6 (2021): Improves upon final prediction grid of size S x S x (5B + K) [48].
YOLOv5 by using a new loss function called GIoU, which This initial YOLO model prioritized speed by utilizing
helps to improve the model’s accuracy. YOLOv7 (2022): a single CNN [58] to directly predict object locations and
Introduces a new network architecture called Panoptic classes in real time. However, this approach led to decreased
YOLO, which can detect both objects and semantic segments accuracy, particularly for small objects or those with overlap-
in an image. ping bounding boxes [59]. The detection architecture of the
YOLOv8 (2023): The latest version of YOLO, which original YOLO model performed a single pass on the image
introduces a number of new features, including a new network to predict object locations and class labels [60]. Subsequent
architecture called BiFPN, a new loss function called CIoU, versions of YOLO introduced improvements to address these
and a new training method called Cross-GPU training. limitations.
YOLOv2, for instance, introduced batch normalization,
B. YOLO ALGORITHM AND ARCHITECTURE anchor boxes, and passthrough layers to enhance object
The YOLO architecture consists of three main components: localization. Additionally, it incorporated multiscale training
(i) the backbone; (ii) the neck; (iii) the head [23], [48], [50], and achieved a processing speed of 40-90 frames per
[54]. The architecture of the backbone, neck, and head can Second (FPS) [50]. These refinements aimed to enhance
vary between different versions of YOLO, and improvements both accuracy and speed in object detection tasks. The
in these components have led to significant improvements in performance of YOLOv3 was improved by incorporating a
the overall accuracy and speed of the YOLO network [25], multi-scale feature extraction architecture. New backbone
[55]. The choice of backbone, neck, and head can affect the network, feature pyramid network, and more anchor boxes.
speed and accuracy of the YOLO model and depends on It allowed a tradeoff between speed and accuracy [25].
the specific application and the desired trade-off [49]. More YOLOv4 and YOLOv5 introduced new network back-
recent versions of YOLO have introduced improvements in bones, improved data augmentation, and optimized training
all three components to achieve better performance. Figure 2 strategies. This enhanced accuracy without severely impact-
represents YOLO Architecture. ing real time performance [61], [62]. YOLO framework
❐ Backbone: The primary role of the backbone is to for object detection consistently evolved to balance speed
extract valuable characteristics from the initial image. and accuracy in detection tasks. PP-YOLO [63], developed
Usually, a convolutional neural network is employed by Alibaba Group, provided further improvements with a
as the backbone, which has been trained on extensive new backbone network, spatial attention module, and path
datasets like ImageNet. The backbone functions as the aggregation network, making it faster and more accurate than
network responsible for extracting features and gener- YOLOv5 [64].
ating feature maps from the input images. In YOLO, YOLOv6 implemented a variant of the EfficientNet archi-
some commonly utilized backbones include VGG16, tecture named EfficientNet-L2, surpassing the EfficientDet
ResNet50, CSPDarknet53, and EfficientRep. architecture used in YOLOv5 in terms of efficiency. It also
❐ Neck: The neck serves as the intermediary between the introduced a design called EfficientRep Backbone and Rep-
backbone and the head in the architecture. It consists PAN Neck, leading to faster and more accurate results. The
of two main components, namely the spatial pyramid head of the network was decoupled, adding layers, and

VOLUME 12, 2024 57817

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

FIGURE 2. The architecture of YOLO consists of a backbone, neck, and head. The backbone, neck, and head
vary in different versions of YOLO. For backbone, normally Darknet, VGG16, or Resnet are used; for neck
feature pyramid network (FPN) [56], and for neck Densenet [57] or sparsenet are used.

FIGURE 3. YOLOv1 network architecture [48].

separating these features from the final head, which improved C. YOLO IN ACTION
performance [65]. YOLO revolutionized the process of object detection by
YOLOv7 improved accuracy without raising inference simultaneously detecting all bounding boxes within an S × S
costs, reducing parameters and computation by 40% and region using grids. It predicts B bounding boxes for each
50% respectively, compared to other leading real time object class, accompanied by confidence scores for C different
detectors [66]. It had a faster, stronger network architecture, classes per grid element [48]. Each bounding box prediction
more accurate detection performance, a more robust loss comprises of five values: Pc , bx , by , bh , bw . Here, Pc repre-
function, and enhanced label assignment and model training sents the confidence score, reflecting the model’s confidence
efficiency. It also required cheaper computing hardware and in the presence and accuracy of the object within the box. The
could be trained faster on small datasets without pre-trained coordinates bx and by denote the box center relative to the
weights [50], [67]. grid cell, while bh and bw indicate the box height and width
YOLOv8 [68], the most advanced model at the time relative to the entire image. The output of YOLO is a tensor of
of writing, had better feature aggregation and a mish size S × S × (B × 5 + C), which may undergo non-maximum
activation function that improved detection accuracy and suppression (NMS) to eliminate duplicate detections. These
processing speed. It is an anchor-free model, predicting grid cells facilitate operations related to bounding box
object centers directly without known anchor boxes. YOLO- estimation and class probabilities [50]. Consequently, YOLO
NAS [69], created by Deci AI, outperformed its predecessors estimates the likelihood of the detection element’s bounding
(especially YOLOv6 and YOLOv8) by achieving higher box center being located within the grid cell, as formulated
mAP values on the COCO dataset while maintaining by Equation 2.
lower latency. It also performed best on the Roboflow C(P) = Prob(p) × IoU (prediction, target) (1)
100 dataset benchmark, indicating its ease of fine-tuning
on custom datasets. Thus, the YOLO family of object where: C(P) is the confidence of prediction P, Prob(p) is
detection models has consistently evolved to optimize both the probability of presence of object p, and IoU (prediction,
speed and accuracy, providing a variety of models to target) is the Intersection over Union between the predicted
cater to diverse applications and hardware requirements. and target bounding boxes.
Table 1 summarizes the key features of each version B ∩ Bgt
of YOLO. IoU = (2)
B ∪ Bgt

57818 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

TABLE 1. Key features of each version of YOLO.

TABLE 2. YOLO included studies categorized in the Oncology domain.

In the context of the YOLO algorithm, the target box is

denoted as Bgt , while the predicted box is represented as
B. The probability (p) signifies whether the object exists
within the detected bounding box. The IoU metric, defined
by Equation 2, calculates the intersection area between the
ground truth and predicted bounding boxes. It determines
an acceptable area for each detected object in the input
image and makes decisions based on it. To obtain the most
suitable bounding box, the confidence value is applied after
the estimation. The process of computation of the IoU can be
illustrated as shown in Figure 4.

D. IMAGE ANNOTATION
Image annotation [70] is a vital process in computer vision FIGURE 4. Computing the Intersection over Union: (a) poor detection
performance, (b) good detection performance, (c) excellent detection
and machine learning. It is the process of labeling or marking performance.
specific objects or regions of interest within an image [71].
It involves adding metadata or annotations to images to
provide additional information about the objects or features or semantic segmentations around those objects. Accurate
present in the image. The purpose of image annotation annotations are crucial for training models to accurately
is to create a labeled dataset that serves as training data detect, recognize, and segment objects. They provide ground
for learning algorithms, particularly for tasks like object truth data, enable object localization, ensure model accuracy
detection, object recognition, and image segmentation [72]. and performance, and facilitate diverse and domain-specific
By annotating images, human annotators or data scientists datasets. Annotations also aid in model evaluation and serve
manually outline or mark the objects of interest within as a valuable resource for transfer learning. In summary,
the image, often by drawing bounding boxes, polygons, image annotation is a fundamental step that underpins

VOLUME 12, 2024 57819

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

TABLE 3. YOLO included studies categorized in the Pathology domain.

the development of reliable and effective computer vision In the above matrix, each line represents a single object
systems in various industries and applications [41]. annotation where c represents the class or label of the
There are several popular tools available for annotating object being annotated. It is usually represented by an
images. The choice of tool often depends on personal integer index corresponding to the class label defined in the
preference, project requirements, and the specific desired fea- YOLO configurations. x and y represents the normalized x, y
tures. Commonly used tools include Visual Object Tagging coordinate of the bounding box’s center point. The value
Tool (VoTT) [73], VGG Image Annotator (VIA) [74], and is relative to the width of the image, ranging from 0 to 1.
Roboflow [75] which is a popular platform for managing, w and h represent the normalized width and height of the
preprocessing, and annotating datasets for computer vision bounding box. The value is relative to the width of the image,
tasks. It provides a comprehensive end-to-end solution for ranging from 0 to 1. Each annotation line corresponds to one
dataset management, annotation, and preprocessing, offering annotated object in the image. Multiple lines can be present
a range of features that can help streamline your object in the annotation file, each representing a different object.
detection workflow. When it comes to annotating datasets for
YOLO, there are a few commonly used annotation formats
E. HOW YOLO OPERATES
that work well with YOLO-based object detection models.
During the process of predicting bounding boxes, YOLO
The most popular annotation format for YOLO datasets is
employs ‘‘dynamic anchor boxes’’ utilizing a clustering
the Darknet format, which is the native format used by the
algorithm. This algorithm groups the ground truth bounding
Darknet framework, the original implementation of YOLO.
  boxes into clusters and utilizes the centroids of these clusters
ci x y w h as anchor boxes [76]. By doing so, the anchor boxes become
 ci x y w h  better aligned with the size and shape of the detected objects.
 
· · · · · · · · · · · · · · · (3) However, the primary source of error in YOLO arises from
 
 ci x y w h  localization. This is due to the fact that the bounding box
ci x y w h ratios are entirely learned from the data, causing YOLO to

57820 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

TABLE 4. YOLO included studies categorized in the Radiology domain.

VOLUME 12, 2024 57821

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

TABLE 4. (Continued.) YOLO included studies categorized in the Radiology domain.

TABLE 5. YOLO included studies categorized in the Surgical Procedures domain.

struggle with atypical ratio bounding boxes [77]. YOLO defines a threshold value for the confidence score,
where predictions below this threshold are discarded. Non-
bx = σ (tx ) + cx
maximum suppression is then applied to generate the final
by = σ ty + cy


positions for the detected bounding boxes. Finally, a loss
bw = (pw ) ∗ etw (4) function is computed for the detected bounding boxes in the
bh = Ph ∗ e th last stage. Figure 5 provides a clear understanding of the work
of YOLO.
In the YOLO framework, the bounding box coordinates are
denoted as bx , by , bw , and bh , while the center coordinates
are represented by x, y, and the width and height are F. ADVANTAGES AND DRAWBACKS OF YOLO FOR
given by bw and ph respectively. Each bounding box has MEDICAL OBJECT DETECTION
estimated coordinates tx , ty , tw , and th . The values cx and YOLO is also highly generalized and can recognize a wide
cy correspond to the upper-left coordinates of the grid cell. range of objects. However, it is important to be aware of the

57822 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

TABLE 6. YOLO included studies categorized in the Personal Protective Equipment Detection domain.

within complex medical images, supporting precise

diagnoses.
• Multi-medical Object Detection: YOLO excels in
detecting multiple objects simultaneously, making it
valuable for identifying various medical anomalies,
tumors, lesions, pathologies factors, or abnormalities
within a single image.
• Adaptability: The YOLO architecture is versatile and
can be fine-tuned for specific medical domains, such
as surgical procedures, personal protective detection
equipment, and medical images.
• Minimal False Positives: YOLO’s ability to incorporate
global context reduces the likelihood of false positive
detections, enhancing the reliability of medical diag-
noses.

2) DRAWBACKS OF YOLO FOR MEDICAL OBJECT

DETECTION
There are some drawbacks of YOLO such as:
• Large Dataset Requirement: One of the primary
drawbacks of YOLO is its need for a substantial dataset
of images for training. The collection of these images
FIGURE 5. YOLO works by dividing the input image into a grid of cells.
Each cell predicts a bounding box, a confidence score, and the class can be both time-consuming and expensive.
probabilities for the objects in that cell. The bounding box is a rectangle • Sensitivity to Object Scale: YOLO’s performance can
that is used to surround the object. A confidence score, between 0 and 1,
indicates how confident the model is that the object is present in the cell. be significantly affected by the scale of objects in
The class probabilities are the probabilities that the object in the cell the input image, leading to potential false positives or
belongs to each of the possible classes.
negatives.
• Difficulty with Small Objects: YOLO often struggles
to detect smaller objects as efficiently as it does larger
advantages and disadvantages of YOLO before using it for a ones. The system divides the image into a grid of cells,
specific application. and small objects might not be large enough to occupy
an entire cell, affecting detection accuracy.
1) ADVANTAGES OF YOLO FOR MEDICAL OBJECT • Issues with Occluded Objects: YOLO tends to falter
DETECTION when tasked with detecting objects obscured by others.
• Real-Time Detection: YOLO’s inherent ability to As it predicts the bounding box of each object,
detect objects in a single pass enables real time an obscured or partially visible object might not
processing of medical images, which is crucial in have a well-defined bounding box, limiting detection
time-sensitive medical scenarios such as surgeries or performance.
emergency diagnostics. • Limited Object Diversity: The system often struggles
• Efficiency: YOLO’s single-pass architecture is com- with a diverse set of object classes. As YOLO is trained
putationally efficient, allowing for faster processing on a finite dataset of images, it may find it challenging
speeds and reduced computational requirements, which to generalize its detection capabilities to unfamiliar
is especially important for medical applications that objects.
demand rapid results. The rest of this research work is structured as follows.
• Accuracy: YOLO’s holistic approach to object detec- Section II provides the methodology for conducting this
tion considers contextual information, leading to accu- study. Section III highlights the results of the literature’s
rate localization and classification of medical objects applicable studies and responds to the previously stated

VOLUME 12, 2024 57823

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

research questions using synthesized data from the included RQ5: How does YOLO compare to other existing
research. Finally, section IV concludes this paper. object detection algorithms in terms of performance,
efficiency, and applicability to medical imaging?
II. METHODS
2) SEARCH STRATEGY
This paper seeks to assemble a comprehensive compilation of
A systematic search of the literature was conducted using
relevant studies focusing on YOLO in the medical domain,
the National Library of Medicine’s PubMed database
covering the period from 2018 to 2023. The aims are to
(https://pubmed.ncbi.nlm.nih.gov, accessed and last searched
explore YOLO’s potential capabilities in medical applica-
on 26 January 2024 ) to identify relevant papers published
tions, diagnosis, and treatment planning. This paper was
between 01/01/2018 and 31/12/2023). The search strategy
conducted using the Preferred Reporting Items for Systematic
employed a combination of keywords, specifically (YOLO
Reviews and Meta-Analyses (PRISMA) statement [78].
AND ((medical application) OR (medical image))), and
The rest of this section has been thoroughly organized
adhered to the PRISMA guidelines [78]. The inclusion
into two subsections. The first subsection (II-A) highlights
criteria focused on original research articles, while review
the evidence acquisition, which explains the aim (II-A1),
papers, abstracts, and reports from meetings were excluded.
search strategy (II-A2), and study selection criteria (II-A3).
Each identified article underwent a thorough evaluation to
Meanwhile, evidence synthesis of this SLR is presented in
determine its eligibility for inclusion in this SLR. Figure 6
the second subsection (II-B).
illustrates the PRISMA flowchart conducted in this study.

A. EVIDENCE ACQUISITION 3) STUDY SELECTION CRITERIA

Evidence acquisition in PRISMA involves systematically The selection criteria were carefully designed to ensure the
searching and selecting relevant studies from databases. identification of research articles that best align with the
Medical applications with YOLO capabilities can aid in objectives of the systematic review while excluding those that
automating the identification and extraction of relevant are not relevant. The following inclusion criteria were applied
evidence from medical images. This integration streamlines to identify the studies to be included in this review:
the evidence acquisition process, improving efficiency and 1) Inclusion Criteria:
accuracy in SLRs. ❐ Peer-reviewed articles published between 2018 and
2023 that focus on the application of YOLO in the
medical domain.
1) AIM
❐ Studies that report original research findings.
The primary goal of this systematic review is to compre-
❐ Studies that primarily focus on medical applications
hensively analyze and evaluate the application of YOLO
of YOLO.
in medical imaging with a focus on assessing its accu-
❐ Studies published in English language.
racy, efficiency, and effectiveness in medical applications. 2) Exclusion Criteria:
Additionally, the review aims to identify medical areas
❐ Studies published before 1 January 2018.
where YOLO has been successfully utilized, examining
❐ Review papers, abstracts, and reports from meetings
the literature to determine its strengths and potential for
or conferences.
significant advancements. Moreover, it provides practical
❐ Studies that do not involve the YOLO algorithm or its
implications for integrating YOLO into medical applications,
application in the medical domain.
considering how its utilization can enhance efficiency,
❐ Studies with insufficient information or incomplete
accuracy, and automation in medical image and diagnosis.
methodology.
To this end, the following questions were established as our
❐ Studies that primarily focus on non-medical applica-
SLR research focus:
tions of YOLO.
RQ1: What are the key medical domains in which ❐ Studies published in languages other than English.
YOLO has been applied for medical object detection?
RQ2: What are the performance metrics used to evaluate B. EVIDENCE SYNTHESIS
the effectiveness of YOLO in medical applications, Initially, 124 articles were obtained from the database search,
and how does YOLO perform in terms of accuracy, and 48 were included in this study. The quality assessment
precision, recall, and other relevant metrics? guidelines employed in this systematic review are designed to
RQ3: What are the specific tasks and objects of reduce bias, enhance transparency, and ensure repeatability.
interest within the medical domain where YOLO has These guidelines focus on the YOLO medical applications
demonstrated strong performance? utilizing medical datasets. The following criteria were
RQ4: What are the limitations and challenges encoun- considered during the assessment process: alignment of the
tered when using YOLO in medical applications, selected studies with the primary objective of the systematic
such as imbalanced datasets, small sample sizes, and review, evaluation of the utilization of performance metrics,
computational requirements? scrutiny of the soundness of the conclusions drawn in the

57824 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

FIGURE 6. PRISMA flowchart.

observable in the graph of the identified publications, starting

with a modest count of 3 identified publications in 2018. This
number remains relatively stable in 2019, suggesting a period
of early academic interest. The subsequent years illustrate
a notable increase in research activity. The year 2020 sees
the number double to 6 identified publications, indicating
a growing recognition of YOLO’s relevance in medical
research. This momentum is maintained and amplified in the
following years, with a steady rise to 18 papers in 2021 and a
sharper ascent to 33 in 2022. The most striking leap occurs
in 2023, where the count of identified publications surges
to 58. This substantial growth over five years signifies a
robust and accelerating interest in the application of YOLO
FIGURE 7. Trend of publications on the YOLO applications in medical within medical imaging. It is worth mentioning that, the
imaging. observed decrease in the number of included studies in
2023 is attributable to the implementation of more strict
inclusion and exclusion criteria, ensuring that only the most
selected studies, and verification of the use of valid and pertinent and high-quality research is selected for review.
reliable datasets. Figure 8 illustrates the distribution of the main versions of
Figure 7 displays the years and the corresponding counts of YOLO detected from the 48 included studies. The pie chart
identified and included publications. A clear upward trend is indicates that YOLOv3 is the most frequently encountered

VOLUME 12, 2024 57825

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

FIGURE 9. Three major domains of YOLO in healthcare; Medical Imaging,

Surgical Procedures, and Personal Protective Equipment Detection.

FIGURE 8. Distribution of YOLO versions in medical imaging.

in chest X-rays, breast masses in mammograms, and brain

version in the literature, with 21 instances, followed by tumors in MRI scans [17], [30], [86], [87]. Its ability to
YOLOv5 in 9 instances, and YOLOv4 in 8 instances. This identify these objects of interest accurately and efficiently
could suggest that these versions have reached a level of enables earlier detection, leading to timely intervention and
maturity and performance that makes them popular among improved patient outcomes. In Figure 10, different medical
researchers and practitioners. YOLOv7 and YOLOv8 are image applications clearly explain the advancement of YOLO
represented as well with 2 instances each, indicating their in the field.
emerging presence in the field and suggesting that more Moreover, YOLO has shown promise in detecting anoma-
research papers on these versions could be yet to be lies in medical images. Anomalies can encompass a wide
published. range of abnormalities, such as fractures, hemorrhages,
or foreign objects. YOLO’s real time performance makes
III. RESULTS it well-suited for tasks such as identifying fractures in X-
This section highlights the results of the literature’s appli- rays, detecting bleeds in brain CT scans, or identifying
cable studies and responds to the previously stated research foreign bodies in radiographic images [87], [88]. By rapidly
questions (Sub-section II-A1) using synthesized data from flagging anomalies, YOLO assists radiologists and clinicians
the included research. in prioritizing critical cases and expediting appropriate
The key medical domains and applications where YOLO treatment.
has been applied for medical object detection, along with In addition, YOLO has been applied successfully in
their performance metrics used to evaluate the YOLO’s segmenting organs such as the heart, liver, kidney, and brain
effectiveness in these applications, with a focus on accuracy, in different imaging modalities, including MRI and CT
precision, recall, and other relevant metrics, are categorized scans [32], [81]. Accurate identification and delineation of
from the 48 included studies into three main domains as organs are crucial for surgical planning, radiation therapy,
shown in Figure 9: I) Medical Imaging (III-A), II) Surgi- and monitoring disease progression [31]. Its high accuracy
cal Procedures (III-B), and Personal Protective Equipment and efficiency streamline the segmentation process and
Detection (III-C). contribute to more precise diagnoses and treatment plans.
In surgical settings, YOLO has shown promise in detecting
A. MEDICAL IMAGING and tracking surgical instruments during procedures [83],
YOLO has exhibited strong performance in various tasks [84]. By analyzing live video feeds, YOLO can identify
and objects of interest within medical imaging [5], including instruments, surgical tools, and other objects of interest in
lesion detection [28], [79], anomaly detection [80], organ real time, providing valuable assistance to surgeons and
segmentation [31], [81], and instrument tracking [82], [83], improving surgical safety. Its ability to handle fast-paced
[84]. These applications highlight YOLO’s potential to environments and track objects accurately makes it a valuable
enhance diagnostic accuracy, improve workflow efficiency, tool in computer-assisted interventions and robotic surgeries.
and ultimately advance patient care in the field of med- Also, YOLO has shown tremendous success in the detec-
ical imaging. One of the primary applications of YOLO tion and classification of breast masses in mammograms [19],
in medical imaging is the detection and localization of [79], [89], [90], [91]. Its use in such applications has proven
lesions [85]. Lesions can include tumors, nodules, masses, to significantly reduce the time, cost, and potential for
or any abnormality that may indicate disease [29]. YOLO has human error inherent in traditional methods of mammogram
demonstrated strong performance in detecting lung nodules evaluation. The YOLO system preprocesses the mammo-

57826 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

FIGURE 10. YOLO in different medical images applications: (a) periodontitis bone loss diagnosis, (b) glomerular
detection, (c) breast cancer detection (d) lung normal and abnormal detection, (e) brain tumor detection, (f) white and
red blood cells detections.

grams and then detects masses in them, distinguishing evaluation process of screening mammograms is a laborious
between malignant and benign lesions without any human task, requiring significant time, cost, and human resources,
intervention [80], [92]. and is prone to errors due to fatigue and the inherent subjec-
The included studies in the medical imagining domain have tivity of human evaluation. However, with the introduction
been further classified into three sub-domains: Oncology of YOLO into this process, an end-to-end computer-aided
(Table 2), Pathology (Table 3), and Radiology (Table 4). diagnosis system has been proposed and implemented [122].
However, while the applications of YOLO in healthcare The described system performs preprocessing on DICOM-
have been fruitful, there are challenges, including the need for format mammograms to convert them into images while
large, diverse, and high-quality datasets for model training. preserving all the data. It is capable of detecting masses
The algorithm’s sensitivity to the scale of objects in images in full-field digital mammograms and can differentiate
is another aspect that needs further improvement. Despite between malignant and benign lesions automatically, without
these challenges, with continuous research and refinement, requiring any human intervention, significantly reducing the
YOLO’s application in medical imaging holds significant potential for human error and streamlining the entire process.
promise for advancing healthcare diagnostics. One of the
most promising areas of application for YOLO is medical B. SURGICAL PROCEDURES
imaging. YOLO has shown promising results in various fields YOLO could also be applied in surgical procedures. It offers
such as radiology, oncology, and pathology [19]. For instance, great potential for surgical procedures, particularly in the
in tumor detection, YOLO can identify and locate abnormal context of computer-assisted and robotic surgery. The
growths in medical images, assisting healthcare professionals algorithm’s ability to detect, classify, and locate objects in
in early disease diagnosis and treatment planning [30], [31], real time can be of significant value in the surgical environ-
[32]. In the context of COVID-19, YOLO has demonstrated ment [82], [83], [84]. For instance, YOLO can be utilized
its value in detecting and quantifying infection patterns in to identify and locate specific surgical instruments within
lung CT scans, contributing to rapid and effective patient the operating field, helping to streamline instrument tracking
management [21]. and potentially reducing surgical errors. Additionally, it could
As healthcare moves towards a more digitized environ- play a role in enhancing the safety and precision of robotic
ment, the volume of medical imaging data is growing. This surgical systems by improving their ability to recognize
data can be effectively analyzed using AI and machine and interact with various surgical elements in real time.
learning tools like YOLO to extract valuable insights that Figure 11 demonstrates the usage of YOLO in surgical
can aid in diagnosis and treatment planning. For instance, procedures.
YOLO has been used to detect and classify breast masses The study by Wang et al. [83] developed an AI model
in mammograms [19], [79], [89], [90], [91]. The traditional based on YOLOv3 to identify parathyroid glands during

VOLUME 12, 2024 57827

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

FIGURE 11. YOLO surgical procedures applications of (a)surgical tool detection in open surgery videos, (b) surgical instruments, (c) real
time instance segmentation of surgical instruments.

endoscopic thyroid surgery. Using 1,700 images from thy- by healthcare professionals and the use of various medical
roidectomy videos, the model outperformed junior surgeons devices. However, these traditional methods can be time-
and was comparable to senior surgeons in identification consuming, costly, and subject to human error.
rates. Amiri Tehrani Zade et al. [84] developed a CNN- For instance, YOLO could be deployed to monitor
based method to enhance needle tracking in ultrasound for patient activity and movements in an in-patient setting,
medical procedures. Using advanced motion estimation and identifying potential falls or other hazardous events before
the YOLOv3 framework, it accurately locates needles in they occur [21]. This real time alert system could greatly
real time ultrasound, outperforming current methods, and enhance patient safety and improve healthcare outcomes.
promising better ultrasound-guided interventions. Table 5 Similarly, for patients with chronic conditions, YOLO could
shows YOLO included studies categorized in the surgical potentially be utilized to monitor medication intake or
procedures domain. adherence to certain therapeutic exercises, promoting better
Surgeons often rely on various imaging technologies, such disease management. YOLO could provide valuable support,
as MRI or CT scans, to guide their procedures. However, identifying significant health events from recorded or live
interpreting these images and applying their insights to the video feeds. Despite the promising prospects, integrating
surgical process can be challenging. With the use of YOLO, YOLO into patient monitoring systems presents challenges,
these images could be analyzed in real time, providing including ensuring patient privacy and dealing with diverse
surgeons with immediate feedback and guidance during and complex real-world scenarios. However, with ongoing
the procedure [82]. This could potentially lead to more research and development, YOLO’s application in patient
precise surgeries, fewer complications, and better patient monitoring can revolutionize care delivery, enhancing patient
outcomes. Furthermore, YOLO could be integrated with safety and health outcomes.
surgical navigation systems to improve real time imaging, Moreover, YOLO has been applied in face mask detec-
enabling surgeons to better visualize the surgical field and tion. Han et al [21] developed an enhanced, lightweight
carry out complex procedures with higher precision. This is YOLOv4-tiny-based detector for real time mask status
particularly relevant for minimally invasive surgeries where detection, offering improved precision and speed with
real time imaging plays a crucial role. Despite these potential fewer parameters, suitable for public health applications.
applications, integrating YOLO into surgical procedures also Loey et al. [4] proposed a deep learning model combining
brings challenges. These include ensuring the algorithm’s ResNet-50 and YOLOv2 to detect medical face masks in
robustness and reliability in a highly variable and complex images, achieving 81% precision and outperforming related
surgical environment, and addressing concerns related to models in accuracy. Table 6 shows YOLO-included studies
patient safety and data privacy. Nevertheless, with continued categorized in the personal protective equipment detection
research and technological refinement, YOLO’s application domain.
in surgical procedures promises to enhance surgical precision With the applications of YOLO, it is possible to build
and patient outcomes [84]. automated patient monitoring systems that can continuously
monitor patients’ vital signs and behavior and alert healthcare
C. PERSONAL PROTECTIVE EQUIPMENT DETECTION professionals to any abnormal patterns or signs of distress.
YOLO has demonstrated significant potential in the field Such systems could potentially lead to faster response times
of patient monitoring as shown in Figure 12. Real-time in emergency situations, more effective use of healthcare
patient monitoring is a critical component of healthcare, resources, and better overall patient outcomes.
providing valuable insights into a patient’s condition and Finally, Table 7 shows the data sources of all the 48
enabling timely interventions. Continuous patient monitoring included studies. From those studies reviewed, no common
is crucial in many healthcare scenarios, from intensive care datasets were identified across the medical applications,
units to home-based care. The ability of YOLO to detect except for a few instances:
and recognize objects in real time can be adapted to monitor • In lung nodule detection: the Lung Nodule Analysis
various aspects of patient care. Traditionally, this monitoring 2016 (LUNA16) dataset was used in 2 papers [86]
has been done through a combination of manual observations and [87].

57828 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

FIGURE 12. Medical personal protective equipment categories: (a) suit, (b) face shield;(c) goggles, (d) mask, and
(e) glove.

• In breast cancer detection: three public benchmark features, it simultaneously escalates the model’s complexity
datasets were utilized. and the computational resources needed.
– INbreast was used in 4 papers [17], [79], [89], Though promising, the adoption of YOLO in healthcare
and [90]. brings with it an array of challenges and limitations that
– Digital Database for Screening Mammography must be acknowledged. However, a prominent drawback
(DDSM) was used in 2 papers [17] and [91]. is its comparative need for more accuracy in detecting
– An enhanced version of DDSM, the Curated Breast small targets, a shortfall with potential ramifications in areas
Imaging Subset of Digital Database for Screen- such as pill recognition or the identification of tiny lesions
ing Mammography (CBIS-DDSM) was used once in medical imaging where the identification of minuscule
by [90]. objects is crucial [125]. It utilizes a deep network structure
• In brain tumor detection: the Tumor Cancer Imaging for feature extraction, enhancing accuracy at the cost of
Archive (TCIA) dataset was utilized by [32] and [81]. considerable computational power. This requirement can be
On the other hand, the majority of the included studies have a limiting factor in healthcare settings where resources are
used different datasets, with many not publicly accessible constrained [126], [127].
due to patient privacy concerns. Legal and ethical guidelines Despite these limitations, YOLO is a robust object
require tight data-sharing controls to protect patient confi- detection algorithm used in various applications. As the
dentiality, highlighting the challenge in medical research to algorithm develops, its accuracy and performance will likely
balance scientific progress with data privacy and protection improve Chaudhary et al. 2023, [128]. The best alternative
norms. to YOLO will depend on the specific application. If speed
is essential, then YOLO may be the best choice [23]. Faster
IV. SUMMARY R-CNN or RetinaNet may be the best choice if accuracy is
This section summarizes this SLR into four sub-sections: critical. It is also essential to consider the size of the objects
the limitations of YOLO in healthcare (IV-A), future that need to be detected. YOLO is not as good at detecting
directions (IV-B), ethical considerations (IV-C), and the final small objects as more significant objects [50]. If small objects
conclusion (IV-D). must be detected, Faster R-CNN or RetinaNet may be better
choices. Finally, it is essential to consider the diversity of the
A. LIMITATIONS OF YOLO IN HEALTHCARE objects that need to be detected [129].
While YOLO has made notable strides in object detection,
it has inherent limitations [123]. The system leverages a one-
stage algorithm that directly predicts object bounding boxes B. FUTURE DIRECTIONS
and class probabilities from images, significantly improv- Developing YOLO variants specifically designed for health-
ing detection speed. YOLO backbones network structure care applications is another promising research direction.
excludes pooling and fully connected layers, instead accom- Such customized systems could cater to the unique demands
plishing image convolutional transformations by modifying of healthcare, resulting in more effective and efficient tools.
the step size of the convolutional core [124]. While this Integrating YOLO with other AI techniques, such as rein-
technique augments the network’s depth and ability to extract forcement learning and transfer learning, could significantly

VOLUME 12, 2024 57829

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

TABLE 7. Data Sources of the selected studies.

57830 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

TABLE 7. (Continued.) Data Sources of the selected studies.

enhance its performance. Reinforcement learning could •Clinical Relevance: The dataset needs to be clinically
enable the system to learn from its errors, thereby continually relevant and accurately represent the challenges that
improving, while transfer learning would allow the applica- medical professionals face in real-world scenarios.
tion of knowledge acquired from one task to related tasks, Addressing these challenges requires collaboration
potentially boosting accuracy and efficiency [130]. Moreover, between medical professionals, data annotators, and machine
it will be interesting to see the potential applications of large learning experts. Rigorous quality control, careful dataset
language models [131] in YOLO or object detection tasks. curation, and domain-specific adaptations of YOLO models
Given YOLO’s success in medical imaging, future studies are essential for successful medical object detection.
are likely to concentrate on enhancing its accuracy for
small target detection and extending its application to other 2) TRANSFER LEARNING FOR MEDICAL OBJECT DETECTION
healthcare areas, such as patient activity monitoring, real time Transfer learning is a valuable technique that can signif-
anomaly detection during surgical procedures, or disease pro- icantly benefit the application of YOLO in the medical
gression prediction based on image data. Moreover, there is imaging domain. Here is how transfer learning can be
a growing interest in integrating YOLO more effectively into leveraged to improve YOLO’s performance:
clinical workflows. This could involve developing interfaces • Pre-Trained Models: Begin by training YOLO on a
for seamless interaction between healthcare professionals large dataset from a related domain, such as natural
and the system or devising protocols to ensure appropriate images. This pre-training imparts general object recog-
communication and utilization of the system’s outputs in nition capabilities to YOLO, capturing low-level fea-
clinical decision-making. This rapidly evolving field will tures that can be valuable for medical object detection.
continue to reveal novel applications, benefits, and limitations • Fine-Tuning: After pre-training, fine-tune the YOLO
of this technology. model using a smaller but domain-specific medical
image dataset. This step adapts the model’s learned
features to the specific characteristics of medical
1) DEVELOPING NEW DATASETS FOR MEDICAL OBJECT images, enhancing its ability to detect medical objects.
DETECTION • Transfer of Knowledge: Transfer learning facilitates
Developing new datasets for medical object detection using the transfer of knowledge from the pre-trained model
YOLO models can be challenging due to several reasons: to the medical domain. This approach jumpstarts the
training process and reduces the amount of labeled
• Data Privacy and Ethics: Medical data is sensitive and medical data required, a critical advantage in medical
protected by strict privacy regulations. imaging where labeled data is often limited.
• Annotating Medical Images: Medical images often • Improved Convergence: Transfer learning allows the
require precise and detailed annotations. Expert knowl- YOLO model to converge faster during fine-tuning,
edge is needed to accurately label abnormalities, making leading to quicker deployment and reducing the risk
annotation time-consuming and labor-intensive. The of overfitting, especially when working with smaller
cost of annotating a large dataset can be significant. medical datasets.
• Limited Data Availability: Unlike general object • Enhanced Feature Extraction: The pre-trained fea-
detection, medical datasets are smaller due to the tures capture valuable information about edges, textures,
limited availability of medical images, especially for and basic shapes. These features can be particularly
rare conditions. This scarcity can affect the model’s beneficial in medical image analysis, aiding in the
performance and generalization. detection of various anomalies.
• Class Imbalance: Medical conditions are often rare,
leading to a class imbalance where certain classes have C. ETHICAL CONSIDERATIONS
very few instances. This can lead to biased models that As YOLO and similar technologies become more prevalent in
perform poorly on underrepresented classes. healthcare, it is important to consider the ethical implications.
• Complexity and Variability: Medical images can Implementing YOLO in healthcare elicits various ethical
exhibit variations due to factors like lighting, equip- and legal quandaries. Issues such as data privacy, informed
ment, patient demographics, and disease progression. consent, and the potential for bias in AI algorithms will need
Capturing this variability in the dataset is crucial for to be addressed. Future research will need to not only focus
robust model performance. on improving the technical aspects of these systems but also

VOLUME 12, 2024 57831

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

ensure that they are used in a way that respects patient rights [2] Z. Li, M. Dong, S. Wen, X. Hu, P. Zhou, and Z. Zeng, ‘‘CLU-
and upholds the principles of medical ethics. CNNs: Object detection for medical images,’’ Neurocomputing, vol. 350,
pp. 53–59, Jul. 2019.
[3] J. Peng, Q. Chen, L. Kang, H. Jie, and Y. Han, ‘‘Autonomous recognition
D. CONCLUSION of multiple surgical instruments tips based on arrow OBB-YOLO
network,’’ IEEE Trans. Instrum. Meas., vol. 71, pp. 1–13, 2022.
To conclude, this SLR offers a comprehensive analysis of
[4] M. Loey, G. Manogaran, M. H. N. Taha, and N. E. M. Khalifa, ‘‘Fighting
the utilization of YOLO in various medical applications, against COVID-19: A novel deep learning model based on YOLO-v2 with
encompassing tumor detection, blood transfusion medicine, ResNet-50 for medical face mask detection,’’ Sustain. Cities Soc., vol. 65,
COVID-19, colorectal cancer, radiology, laryngeal cancer, Feb. 2021, Art. no. 102600.
[5] R. Yang and Y. Yu, ‘‘Artificial convolutional neural network in object
parathyroid surgery, and dorsal hand veins recognition, detection and semantic segmentation for medical imaging analysis,’’
among others. The review incorporated a significant body Frontiers Oncol., vol. 11, Mar. 2021, Art. no. 638182.
of literature, aggregating insights from 124 papers published [6] M. Tsuneki, ‘‘Deep learning models in medical image analysis,’’ J. Oral
Biosci., vol. 64, no. 3, pp. 312–320, Sep. 2022.
between 2018 and 2023. The findings reveal the pivotal role
[7] Y. Zhao, K. Zeng, Y. Zhao, P. Bhatia, M. Ranganath, M. L. Kozhikkavil,
YOLO plays in enhancing the efficiency and accuracy of C. Li, and G. Hermosillo, ‘‘Deep learning solution for medical image
medical diagnoses and procedures. localization and orientation detection,’’ Med. Image Anal., vol. 81,
The study also has a few limitations. In this study, Oct. 2022, Art. no. 102529.
[8] R. Qureshi, M. Irfan, H. Ali, A. Khan, A. S. Nittala, S. Ali, A. Shah,
we only focused on the Pubmed database, other databases T. M. Gondal, F. Sadak, Z. Shah, M. U. Hadi, S. Khan, Q. Al-Tashi,
may have relevant articles as well. However, in the medical J. Wu, A. Bermak, and T. Alam, ‘‘Artificial intelligence and biosensors
domain, Pubmed is considered as a gold standard. Another in healthcare and its clinical relevance: A review,’’ IEEE Access, vol. 11,
pp. 61600–61620, 2023.
limitation is that we considered object detection tasks only in
[9] Q. Al-Tashi, M. B. Saad, A. Muneer, R. Qureshi, S. Mirjalili,
medical images, personal protective equipment, and surgical A. Sheshadri, X. Le, N. I. Vokes, J. Zhang, and J. Wu, ‘‘Machine
procedures. Further medical instruments are not considered learning models for the identification of prognostic and predictive cancer
as medical objects. biomarkers: A systematic review,’’ Int. J. Mol. Sci., vol. 24, no. 9, p. 7781,
Apr. 2023.
By rapidly identifying and localizing ailments ranging [10] R. Qureshi, M. Irfan, T. M. Gondal, S. Khan, J. Wu, M. U. Hadi,
from tumors to various cancers, YOLO has significantly J. Heymach, X. Le, H. Yan, and T. Alam, ‘‘AI in drug discovery and its
improved patient outcomes while reducing diagnosis and clinical relevance,’’ Heliyon, vol. 9, no. 7, Jul. 2023, Art. no. e17575.
[11] T. Panch, H. Mattie, and L. A. Celi, ‘‘The ‘inconvenient truth’ about AI
treatment times. However, despite the remarkable successes in healthcare,’’ NPJ Digit. Med., vol. 2, no. 1, p. 77, 2019.
of YOLO, its deployment is not without challenges. These [12] R. Qureshi, B. Zou, T. Alam, J. Wu, V. H. F. Lee, and H. Yan,
include its sensitivity to object scale, difficulty in detecting ‘‘Computational methods for the analysis and prediction of EGFR-
small or occluded objects, and considerable computational mutated lung cancer drug resistance: Recent advances in drug design,
challenges and future prospects,’’ IEEE/ACM Trans. Comput. Biol.
resource requirements. To harness the full potential of YOLO, Bioinf., vol. 20, no. 1, pp. 238–255, Jan. 2023.
these issues need to be addressed by ongoing and future [13] Z. Zou, K. Chen, Z. Shi, Y. Guo, and J. Ye, ‘‘Object detection in 20 years:
research. A survey,’’ Proc. IEEE, vol. 111, no. 3, pp. 257–276, Mar. 2023.
Also, ethical considerations like data privacy and algorith- [14] M. Nawaz, R. Qureshi, M. A. Teevno, and A. R. Shahid, ‘‘Object
detection and segmentation by composition of fast fuzzy C-mean
mic bias need to be considered in the development of YOLO- clustering based maps,’’ J. Ambient Intell. Humanized Comput., vol. 14,
based systems, particularly in healthcare. In summary, the no. 6, pp. 7173–7188, Jun. 2023.
integration of YOLO into healthcare applications represents [15] Z.-Q. Zhao, P. Zheng, S.-T. Xu, and X. Wu, ‘‘Object detection with deep
learning: A review,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 30,
a significant stride towards a future where AI not only no. 11, pp. 3212–3232, Nov. 2019.
enhances the accuracy and speed of medical processes but [16] I. Krasin, T. Duerig, N. Alldrin, V. Ferrari, S. Abu-El-Haija,
also democratizes access to quality healthcare. Nevertheless, A. Kuznetsova, H. Rom, J. Uijlings, S. Popov, and A. Veit, ‘‘OpenImages:
A public dataset for large-scale multi-label and multi-class image
continued research and development are essential for further classification,’’ Dataset, vol. 2, no. 3, p. 18, 2017. [Online]. Available:
improvements and for the optimal integration of YOLO into https://github.com/openimages
healthcare settings. [17] M. A. Al-antari, S.-M. Han, and T.-S. Kim, ‘‘Evaluation of deep
learning detection and classification towards computer-aided diag-
nosis of breast lesions in digital X-ray mammograms,’’ Comput.
ACKNOWLEDGMENT Methods Programs Biomed., vol. 196, 2020, Art. no. 105584, doi:
The authors would like to thank Ministry of Higher 10.1016/j.cmpb.2020.105584.
Education (MOHE), Malaysia for providing financial [18] Y. E. Almalki, A. I. Din, M. Ramzan, M. Irfan, K. M. Aamir,
A. Almalki, S. Alotaibi, G. Alaglan, H. A. Alshamrani, and S. Rahman,
assistance under Fundamental Research Grant Scheme ‘‘Deep learning models for classification of dental diseases using
(FRGS/1/2022/ICT02/UTP/02/4) and Universiti Teknologi orthopantomography X-ray OPG images,’’ Sensors, vol. 22, no. 19,
PETRONAS under the Yayasan Universiti Teknologi p. 7370, Sep. 2022, doi: 10.3390/s22197370.
[19] A. Baccouche, B. Garcia-Zapirain, Y. Zheng, and A. S. Elmaghraby,
PETRONAS (YUTP-FRG/015LC0-308) for providing the ‘‘Early detection and classification of abnormality in prior mammograms
required facilities to conduct this research work. using image-to-image translation and YOLO techniques,’’ Comput.
Methods Programs Biomed., vol. 221, Jun. 2022, Art. no. 106884, doi:
10.1016/j.cmpb.2022.106884.
REFERENCES [20] M. Dobrovolny, J. Benes, J. Langer, O. Krejcar, and A. Selamat,
[1] A. Esteva, K. Chou, S. Yeung, N. Naik, A. Madani, A. Mottaghi, Y. Liu, ‘‘Study on sperm-cell detection using YOLOv5 architecture with
E. Topol, J. Dean, and R. Socher, ‘‘Deep learning-enabled medical labaled dataset,’’ Genes, vol. 14, no. 2, p. 451, Feb. 2023, doi:
computer vision,’’ NPJ Digit. Med., vol. 4, no. 1, p. 5, Jan. 2021. 10.3390/genes14020451.

57832 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

[21] Z. Han, H. Huang, Q. Fan, Y. Li, Y. Li, and X. Chen, ‘‘SMD-YOLO: [40] I. Pacal, A. Karaman, D. Karaboga, B. Akay, A. Basturk, U. Nalbantoglu,
An efficient and lightweight detection method for mask wearing status and S. Coskun, ‘‘An efficient real-time colonic polyp detection
during the COVID-19 pandemic,’’ Comput. Methods Programs Biomed., with YOLO algorithms trained by using negative samples and large
vol. 221, Jun. 2022, Art. no. 106888, doi: 10.1016/j.cmpb.2022.106888. datasets,’’ Comput. Biol. Med., vol. 141, Feb. 2022, Art. no. 105031, doi:
[22] Z. Huang, Y. Li, T. Zhao, P. Ying, Y. Fan, and J. Li, ‘‘Infusion port 10.1016/j.compbiomed.2021.105031.
level detection for intravenous infusion based on YOLO v3 neural [41] L. Tan, T. Huangfu, L. Wu, and W. Chen, ‘‘Comparison of RetinaNet,
network,’’ Math. Biosciences Eng., vol. 18, no. 4, pp. 3491–3501, 2021, SSD, and YOLO v3 for real-time pill identification,’’ BMC Med.
doi: 10.3934/mbe.2021175. Informat. Decis. Making, vol. 21, no. 1, pp. 1–11, Dec. 2021.
[23] T. Diwan, G. Anirudh, and J. V. Tembhurne, ‘‘Object detection using [42] K. Li and L. Cao, ‘‘A review of object detection techniques,’’ in Proc.
YOLO: Challenges, architectural successors, datasets and applications,’’ 5th Int. Conf. Electromechanical Control Technol. Transp. (ICECTT),
Multimedia Tools Appl., vol. 82, no. 6, pp. 9243–9275, Mar. 2023. May 2020, pp. 385–390.
[24] W. Fang, L. Wang, and P. Ren, ‘‘Tinier-YOLO: A real-time object [43] A. Kaur, Y. Singh, N. Neeru, L. Kaur, and A. Singh, ‘‘A survey on deep
detection method for constrained environments,’’ IEEE Access, vol. 8, learning approaches to medical images and a systematic look up into
pp. 1935–1944, 2020. real-time object detection,’’ Arch. Comput. Methods Eng., vol. 29, no. 4,
[25] P. Jiang, D. Ergu, F. Liu, Y. Cai, and B. Ma, ‘‘A review of YOLO algorithm pp. 2071–2111, Jun. 2022.
developments,’’ Proc. Comput. Sci., vol. 199, pp. 1066–1073, Jan. 2022. [44] N. Ganatra, ‘‘A comprehensive study of applying object detection
[26] M. J. Mortada, S. Tomassini, H. Anbar, M. Morettini, L. Burattini, and methods for medical image analysis,’’ in Proc. 8th Int. Conf. Comput.
A. Sbrollini, ‘‘Segmentation of anatomical structures of the left heart from Sustain. Global Develop. (INDIACom), Mar. 2021, pp. 821–826.
echocardiographic images using deep learning,’’ Diagnostics, vol. 13, [45] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, ‘‘You only look once:
no. 10, p. 1683, May 2023. Unified, real-time object detection,’’ in Proc. IEEE Conf. Comput. Vis.
Pattern Recognit. (CVPR), Jun. 2016, pp. 779–788.
[27] P. Zeng, S. Liu, S. He, Q. Zheng, J. Wu, Y. Liu, G. Lyu, and P. Liu,
‘‘TUSPM-NET: A multi-task model for thyroid ultrasound standard plane [46] S. Albawi, T. A. Mohammed, and S. Al-Zawi, ‘‘Understanding of a
recognition and detection of key anatomical structures of the thyroid,’’ convolutional neural network,’’ in Proc. Int. Conf. Eng. Technol. (ICET),
Comput. Biol. Med., vol. 163, Sep. 2023, Art. no. 107069. Aug. 2017, pp. 1–6.
[47] M. G. Ragab, S. J. Abdulkadir, and N. Aziz, ‘‘Random search one
[28] A. Baccouche, B. Garcia-Zapirain, C. C. Olea, and A. S. Elmaghraby,
dimensional CNN for human activity recognition,’’ Int. Conf. Comput.
‘‘Breast lesions detection and classification via YOLO-based fusion
Intell. (ICCI), pp. 86–91, Oct. 2020.
models,’’ Comput., Mater. Continua, vol. 69, no. 1, pp. 1407–1425, 2021.
[48] M. Hussain, ‘‘YOLO-v1 to YOLO-v8, the rise of YOLO and its
[29] C. Santos, M. Aguiar, D. Welfer, and B. Belloni, ‘‘A new approach for
complementary nature toward digital manufacturing and industrial defect
detecting fundus lesions using image processing and deep neural network
detection,’’ Machines, vol. 11, no. 7, p. 677, Jun. 2023.
architecture based on YOLO model,’’ Sensors, vol. 22, no. 17, p. 6441,
[49] C. Chen, Z. Zheng, T. Xu, S. Guo, S. Feng, W. Yao, and Y. Lan, ‘‘YOLO-
Aug. 2022.
based UAV technology: A review of the research and its applications,’’
[30] F. J. P. Montalbo, ‘‘A computer-aided diagnosis of brain tumors using
Drones, vol. 7, no. 3, p. 190, Mar. 2023.
a fine-tuned YOLO-based model with transfer learning,’’ KSII Trans.
[50] J. Terven and D. Cordova-Esparza, ‘‘A comprehensive review of YOLO
Internet Inf. Syst., vol. 14, no. 12, pp. 4816–4834, 2020.
architectures in computer vision: From YOLOv1 to YOLOv8 and YOLO-
[31] R. Rong, H. Sheng, K. W. Jin, F. Wu, D. Luo, Z. Wen, C. Tang, NAS,’’ 2023, arXiv:2304.00501.
D. M. Yang, L. Jia, M. Amgad, L. A. D. Cooper, Y. Xie, X. Zhan, [51] L. Aziz, M. S. B. Haji Salam, U. U. Sheikh, and S. Ayub, ‘‘Exploring deep
S. Wang, and G. Xiao, ‘‘A deep learning approach for histology-based learning-based architecture, strategies, applications and current trends in
nucleus segmentation and tumor microenvironment characterization,’’ generic object detection: A comprehensive review,’’ IEEE Access, vol. 8,
Modern Pathol., vol. 36, no. 8, Aug. 2023, Art. no. 100196, doi: pp. 170461–170495, 2020.
10.1016/j.modpat.2023.100196.
[52] P. Bharati and A. Pramanik, ‘‘Deep learning techniques—R-CNN to
[32] M. Safdar, S. Kobaisi, and F. Zahra, ‘‘A comparative analysis of data mask R-CNN: A survey,’’ in Computational Intelligence in Pattern
augmentation approaches for magnetic resonance imaging (MRI) scan Recognition. Singapore: Springer, 2020, pp. 657–668.
images of brain tumor,’’ Acta Inf. Medica, vol. 28, no. 1, p. 29, 2020, doi: [53] C. Yu, Z. Hu, R. Li, X. Xia, Y. Zhao, X. Fan, and Y. Bai, ‘‘Segmentation
10.5455/aim.2020.28.29-36. and density statistics of mariculture cages from remote sensing images
[33] J. Zhou, B. Zhang, X. Yuan, C. Lian, L. Ji, Q. Zhang, and J. Yue, ‘‘YOLO- using mask R-CNN,’’ Inf. Process. Agricult., vol. 9, no. 3, pp. 417–430,
CIR: The network based on YOLO and ConvNeXt for infrared object Sep. 2022.
detection,’’ Infr. Phys. Technol., vol. 131, Jun. 2023, Art. no. 104703. [54] Y. Cao, D. Pang, Y. Yan, Y. Jiang, and C. Tian, ‘‘A photovoltaic surface
[34] S. Bashir, R. Qureshi, A. Shah, X. Fan, and T. Alam, ‘‘YOLOv5-M: defect detection method for building based on deep learning,’’ J. Building
A deep neural network for medical object detection in real-time,’’ in Eng., vol. 70, Jul. 2023, Art. no. 106375.
Proc. IEEE Symp. Ind. Electron. Appl. (ISIEA), Jul. 2023, pp. 1–6, doi: [55] R. Kaur and S. Singh, ‘‘A comprehensive review of object detection
10.1109/ISIEA58478.2023.10212322. with deep learning,’’ Digit. Signal Process., vol. 132, Jan. 2023,
[35] H. M. Ali and N. K. El Abbadi, ‘‘Optic disc localization in retinal fundus Art. no. 103812.
images based on you only look once network (YOLO),’’ Int. J. Intell. Eng. [56] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S.
Syst., vol. 16, no. 2, pp. 1–11, 2023. Belongie, ‘‘Feature pyramid networks for object detection,’’ in Proc.
[36] J. Liang, Z. Wang, and X. Ye, ‘‘Application of deep learning in imaging IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017,
diagnosis of brain diseases,’’ in Proc. 3rd Int. Conf. Mach. Learn., Big pp. 936–944.
Data Bus. Intell. (MLBDBI), Dec. 2021, pp. 166–175. [57] F. Iandola, M. Moskewicz, S. Karayev, R. Girshick, T. Darrell, and
[37] H. Honda, S. Mori, A. Watanabe, N. Sasagasako, S. Sadashima, T. K. Keutzer, ‘‘DenseNet: Implementing efficient ConvNet descriptor
Dong, K. Satoh, N. Nishida, and T. Iwaki, ‘‘Abnormal prion protein pyramids,’’ 2014, arXiv:1404.1869.
deposits with high seeding activities in the skeletal muscle, femoral nerve, [58] M. G. Ragab, S. J. Abdulkadir, N. Aziz, Q. Al-Tashi, Y. Alyousifi,
and scalp of an autopsied case of sporadic Creutzfeldt–Jakob disease,’’ H. Alhussian, and A. Alqushaibi, ‘‘A novel one-dimensional CNN
Neuropathology, vol. 41, no. 2, pp. 152–158, Apr. 2021. with exponential adaptive gradients for air pollution index prediction,’’
[38] Z. Zhuang, G. Liu, W. Ding, A. N. J. Raj, S. Qiu, J. Guo, Sustainability, vol. 12, no. 23, p. 10090, Dec. 2020.
and Y. Yuan, ‘‘Cardiac VFM visualization and analysis based on [59] A. B. Amjoud and M. Amrouch, ‘‘Object detection using deep learning,
YOLO deep learning model and modified 2D continuity equation,’’ CNNs and vision transformers: A review,’’ IEEE Access, vol. 11,
Comput. Med. Imag. Graph., vol. 82, 2020, Art. no. 101732, doi: pp. 35479–35516, 2023.
10.1016/j.compmedimag.2020.101732. [60] W. Chen, Y. Li, Z. Tian, and F. Zhang, ‘‘2D and 3D object detection
[39] K. K. Wong, M. Ayoub, Z. Cao, C. Chen, W. Chen, D. N. Ghista, algorithms from images: A survey,’’ Array, vol. 19, Sep. 2023,
and C. W. J. Zhang, ‘‘The synergy of cybernetical intelligence with Art. no. 100305.
medical image analysis for deep medicine: A methodological per- [61] S. S. A. Zaidi, M. S. Ansari, A. Aslam, N. Kanwal, M. Asghar, and B. Lee,
spective,’’ Comput. Methods Programs Biomed., vol. 240, Oct. 2023, ‘‘A survey of modern deep learning based object detection models,’’ Digit.
Art. no. 107677. Signal Process., vol. 126, Jun. 2022, Art. no. 103514.

VOLUME 12, 2024 57833

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

[62] X. Cong, S. Li, F. Chen, C. Liu, and Y. Meng, ‘‘A review of YOLO object [84] A. A. T. Zade, M. J. Aziz, H. Majedi, A. Mirbagheri, and A. Ahmadian,
detection algorithms based on deep learning,’’ Frontiers Comput. Intell. ‘‘Spatiotemporal analysis of speckle dynamics to track invisible needle
Syst., vol. 4, no. 2, pp. 17–20, Jun. 2023. in ultrasound sequences using convolutional neural networks: A phantom
[63] A.-A. Tulbure, A.-A. Tulbure, and E.-H. Dulf, ‘‘A review on modern study,’’ Int. J. Comput. Assist. Radiol. Surgery, vol. 18, no. 8,
defect detection models using DCNNs—Deep convolutional neural pp. 1373–1382, Feb. 2023, doi: 10.1007/s11548-022-02812-y.
networks,’’ J. Adv. Res., vol. 35, pp. 33–48, Jan. 2022. [85] M. Mushtaq, M. U. Akram, N. S. Alghamdi, J. Fatima, and R. F. Masood,
[64] N.-N. Dao, T.-H. Do, S. Cho, and S. Dustdar, ‘‘Information revealed by ‘‘Localization and edge-based segmentation of lumbar spine vertebrae to
vision: A review on the next-generation OCC standard for AIoV,’’ IT identify the deformities using deep learning models,’’ Sensors, vol. 22,
Prof., vol. 24, no. 4, pp. 58–65, Jul. 2022. no. 4, p. 1547, Feb. 2022, doi:10.3390/s22041547.
[65] S. Gupta and S. Nair, ‘‘A review of the emerging role of UAVs in [86] Y. Ahmadyar, A. Kamali-Asl, H. Arabi, R. Samimi, and H. Zaidi,
construction site safety monitoring,’’ Mater. Today, Proc., 2023, doi: ‘‘Hierarchical approach for pulmonary-nodule identification from CT
10.1016/j.matpr.2023.03.135. images using YOLO model and a 3D neural network classifier,’’
[66] Y. Wang, H. Wang, and Z. Xin, ‘‘Efficient detection model of steel Radiological Phys. Technol., vol. 17, no. 1, pp. 124–134, Mar. 2024, doi:
strip surface defects based on YOLO-V7,’’ IEEE Access, vol. 10, 10.1007/s12194-023-00756-9.
pp. 133936–133944, 2022. [87] Y.-S. Huang, P.-R. Chou, H.-M. Chen, Y.-C. Chang, and R.-F.
[67] L. Cao, X. Zheng, and L. Fang, ‘‘The semantic segmentation of standing Chang, ‘‘One-stage pulmonary nodule detection using 3-D DCNN
tree images based on the YOLO v7 deep learning algorithm,’’ Electronics, with feature fusion and attention mechanism in CT image,’’ Comput.
vol. 12, no. 4, p. 929, Feb. 2023. Methods Programs Biomed., vol. 220, Jun. 2022, Art. no. 106786, doi:
[68] G. Jocher, A. Chaurasia, and J. Qiu. (2023). Ultralytics YOLOv8. 10.1016/j.cmpb.2022.106786.
[Online]. Available: https://github.com/ultralytics/ultralytics [88] C. Liu, S.-C. Hu, C. Wang, K. Lafata, and F.-F. Yin, ‘‘Automatic detection
of pulmonary nodules on CT images with YOLOv3: Development and
[69] S. Aharon, Louis-Dupont, O. Masad, K. Yurkova, L. Fridman,
evaluation using simulated and patient data,’’ Quant. Imag. Med. Surg.,
E. Khvedchenya, R. Rubin, N. Bagrov, B. Tymchenko, T. Keren, and
vol. 10, no. 10, pp. 1917–1929, Oct. 2020, doi: 10.21037/qims-19-883.
A. Zhilko, ‘‘Super-gradients,’’ Tech. Rep., 2021, doi: 10.5281/ZEN-
[89] G. H. Aly, M. Marey, S. A. El-Sayed, and M. F. Tolba, ‘‘YOLO based
ODO.7789328.
breast masses detection and classification in full-field digital mammo-
[70] D. Zhang, M. M. Islam, and G. Lu, ‘‘A review on automatic image
grams,’’ Comput. Methods Programs Biomed., vol. 200, Mar. 2021,
annotation techniques,’’ Pattern Recognit., vol. 45, no. 1, pp. 346–362,
Art. no. 105823, doi: 10.1016/j.cmpb.2020.105823.
Jan. 2012.
[90] Y. Su, Q. Liu, W. Xie, and P. Hu, ‘‘YOLO-LOGO: A transformer-based
[71] C. Sager, C. Janiesch, and P. Zschech, ‘‘A survey of image labelling for
YOLO segmentation model for breast mass detection and segmentation
computer vision applications,’’ J. Bus. Analytics, vol. 4, no. 2, pp. 91–110,
in digital mammograms,’’ Comput. Methods Programs Biomed., vol. 221,
Jul. 2021.
Jun. 2022, Art. no. 106903, doi: 10.1016/j.cmpb.2022.106903.
[72] A. Kumar, A. Kalia, K. Verma, A. Sharma, and M. Kaushal, ‘‘Scaling up
[91] Q. Fu and H. Dong, ‘‘Spiking neural network based on multi-scale
face masks detection with YOLO on a novel dataset,’’ Optik, vol. 239,
saliency fusion for breast cancer detection,’’ Entropy, vol. 24, no. 11,
Aug. 2021, Art. no. 166744.
p. 1543, Oct. 2022, doi: 10.3390/e24111543.
[73] S. Annadatha, M. Fridberg, S. Kold, O. Rahbek, and M. Shen, ‘‘A tool for [92] Y. Ku, H. Ding, and G. Wang, ‘‘Efficient synchronous real-time
thermal image annotation and automatic temperature extraction around CADe for multicategory lesions in gastroscopy by using multiclass
orthopedic pin sites,’’ in Proc. IEEE 5th Int. Conf. Image Process. Appl. detection model,’’ BioMed Res. Int., vol. 2022, pp. 1–9, Aug. 2022, doi:
Syst. (IPAS), vol. Five, Dec. 2022, pp. 1–5. 10.1155/2022/8504149.
[74] A. Dutta, A. Gupta, and A. Zissermann. (2316). Vgg Image Annotator [93] D.-C. Cheng, T.-C. Hsieh, K.-Y. Yen, and C.-H. Kao, ‘‘Lesion-based bone
(Via). [Online]. Available: http://www.robots.ox.ac.uk/~vgg/software/via metastasis detection in chest bone scintigraphy images of prostate cancer
[75] F. Ciaglia, F. Saverio Zuppichini, P. Guerrie, M. McQuade, and patients using pre-train, negative mining, and deep learning,’’ Diagnos-
J. Solawetz, ‘‘Roboflow 100: A rich, multi-domain object detection tics, vol. 11, no. 3, p. 518, Mar. 2021, doi: 10.3390/diagnostics11030518.
benchmark,’’ 2022, arXiv:2211.13523. [94] S. Li, Y. Li, J. Yao, B. Chen, J. Song, Q. Xue, and X. Yang, ‘‘Label-free
[76] Y. Hu, X. Wu, G. Zheng, and X. Liu, ‘‘Object detection of UAV for anti- classification of dead and live colonic adenocarcinoma cells based on 2D
UAV based on improved YOLO v3,’’ in Proc. Chin. Control Conf. (CCC), light scattering and deep learning analysis,’’ Cytometry A, vol. 99, no. 11,
Jul. 2019, pp. 8386–8390. pp. 1134–1142, Nov. 2021, doi: 10.1002/cyto.a.24475.
[77] Y. Jamtsho, P. Riyamongkol, and R. Waranusast, ‘‘Real-time bhutanese [95] N. Larpant, W. Niamsi, J. Noiphung, W. Chanakiat, T. Sakuldamrong-
license plate localization using YOLO,’’ ICT Exp., vol. 6, no. 2, panich, V. Kittichai, T. Tongloy, S. Chuwongin, S. Boonsang, and
pp. 121–124, Jun. 2020. W. Laiwattanapaisal, ‘‘Simultaneous phenotyping of five RH red blood
[78] M. J. Page, ‘‘The PRISMA 2020 statement: An updated guideline for cell antigens on a paper-based analytical device combined with deep
reporting systematic reviews,’’ Systematic Rev., vol. 10, no. 1, p. 89, learning for rapid and accurate interpretation,’’ Analytica Chim. Acta,
Dec. 2021, doi: 10.1186/s13643-021-01626-4. vol. 1207, May 2022, Art. no. 339807, doi: 10.1016/j.aca.2022.339807.
[79] M. A. Al-Antari, M. A. Al-Masni, and T. S. Kim, ‘‘Deep learning [96] Z. Han, H. Huang, D. Lu, Q. Fan, C. Ma, X. Chen, Q. Gu, and
computer-aided diagnosis for breast lesion in digital mammogram,’’ in Q. Chen, ‘‘One-stage and lightweight CNN detection approach with
Deep Learning in Medical Image Analysis, vol. 1213, 2020, pp. 59–72. attention: Application to WBC detection of microscopic images,’’
10.1007/978-3-030-33128-3_4. Comput. Biol. Med., vol. 154, Mar. 2023, Art. no. 106606, doi:
[80] L. Guo, Y. Yang, H. Ding, H. Zheng, H. Yang, J. Xie, Y. Li, T. Lin, and 10.1016/j.compbiomed.2023.106606.
Y. Ge, ‘‘A deep learning-based hybrid artificial intelligence model for the [97] M.-Y. Quan, Y.-X. Huang, C.-Y. Wang, Q. Zhang, C. Chang, and
detection and severity assessment of vitiligo lesions,’’ Ann. Transl. Med., S.-C. Zhou, ‘‘Deep learning radiomics model based on breast ultrasound
vol. 10, no. 10, p. 590, 2022, doi: 10.21037/atm-22-1738. video to predict HER2 expression status,’’ Frontiers Endocrinology,
[81] S. Afshari, A. BenTaieb, and G. Hamarneh, ‘‘Automatic local- vol. 14, Apr. 2023, Art. no. 1144812, doi: 10.3389/fendo.2023.1144812.
ization of normal active organs in 3D PET scans,’’ Computer- [98] R. Zhu, Y. Cui, J. Huang, E. Hou, J. Zhao, Z. Zhou, and H. Li,
ized Med. Imag. Graph., vol. 70, pp. 111–118, Dec. 2018, doi: ‘‘YOLOv5s-SA: Light-weighted and improved YOLOv5s for sperm
10.1016/j.compmedimag.2018.09.008. detection,’’ Diagnostics, vol. 13, no. 6, p. 1100, Mar. 2023, doi:
[82] Y. Huang, J. Li, X. Zhang, K. Xie, J. Li, Y. Liu, C. S. H. Ng, 10.3390/diagnostics13061100.
P. W. Y. Chiu, and Z. Li, ‘‘A surgeon preference-guided autonomous [99] G. Sun, C. Lyu, R. Cai, C. Yu, H. Sun, K. E. Schriver, L. Gao, and X. Li,
instrument tracking method with a robotic flexible endoscope based on ‘‘DeepBhvTracking: A novel behavior tracking method for laboratory
dVRK platform,’’ IEEE Robot. Autom. Lett., vol. 7, no. 2, pp. 2250–2257, animals based on deep learning,’’ Frontiers Behav. Neurosci., vol. 15,
Apr. 2022. Oct. 2021, Art. no. 750894, doi: 10.3389/fnbeh.2021.750894.
[83] B. Wang, J. Zheng, J. Yu, S. Lin, S. Yan, L. Zhang, S. Wang, S. Cai, [100] Z.-J. Huang, B. Patel, W.-H. Lu, T.-Y. Yang, W.-C. Tung, V. Bučinskas,
A. H. A. Ahmed, L. Lin, F. Chen, G. W. Randolph, and W. Zhao, M. Greitans, Y.-W. Wu, and P. T. Lin, ‘‘Yeast cell detection using
‘‘Development of artificial intelligence for parathyroid recognition fuzzy automatic contrast enhancement (FACE) and you only look
during endoscopic thyroid surgery,’’ Laryngoscope, vol. 132, no. 12, once (YOLO),’’ Sci. Rep., vol. 13, no. 1, p. 16222, Sep. 2023, doi:
pp. 2516–2523, Dec. 2022, doi: 10.1002/lary.30173. 10.1038/s41598-023-43452-9.

57834 VOLUME 12, 2024

M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

[101] Y.-F. Chen, Z.-J. Chen, Y.-Y. Lin, Z.-Q. Lin, C.-N. Chen, M.-L. Yang, [117] T. Ma, Z. Ma, X. Zhang, and F. Zhou, ‘‘Evaluation of effect of
J.-Y. Zhang, Y.-Z. Li, Y. Wang, and Y.-H. Huang, ‘‘Stroke risk study curcumin on psychological state of patients with pulmonary hyper-
based on deep learning-based magnetic resonance imaging carotid plaque tension by magnetic resonance image under deep learning,’’ Contrast
automatic segmentation algorithm,’’ Frontiers Cardiovascular Med., Media Mol. Imag., vol. 2021, pp. 1–10, Jul. 2021, doi: 10.1155/2021/
vol. 10, Feb. 2023, Art. no. 1101765, doi: 10.3389/fcvm.2023.1101765. 9935754.
[102] M. A. Al-masni, W.-R. Kim, E. Y. Kim, Y. Noh, and D.-H. Kim, [118] K.-C. Lee, Y. Cho, K.-S. Ahn, H.-J. Park, Y.-S. Kang, S. Lee, D. Kim, and
‘‘Automated detection of cerebral microbleeds in MR images: A two- C. H. Kang, ‘‘Deep-learning-based automated rotator cuff tear screening
stage deep learning approach,’’ NeuroImage, Clin., vol. 28, Apr. 2020, in three planes of shoulder MRI,’’ Diagnostics, vol. 13, no. 20, p. 3254,
Art. no. 10246410.1016/j.nicl.2020.102464. Oct. 2023, doi: 10.3390/diagnostics13203254.
[103] Y. Nambu, T. Mariya, S. Shinkai, M. Umemoto, H. Asanuma, I. Sato, [119] J. Fu, J.-W. Chai, P.-L. Chen, Y.-W. Ding, and H.-C. Chen, ‘‘Quan-
Y. Hirohashi, T. Torigoe, Y. Fujino, and T. Saito, ‘‘A screening assistance titative measurement of spinal cerebrospinal fluid by cascade arti-
system for cervical cytology of squamous cell atypia based on a two-step ficial intelligence models in patients with spontaneous intracranial
combined CNN algorithm with label smoothing,’’ Cancer Med., vol. 11, hypotension,’’ Biomedicines, vol. 10, no. 8, p. 2049, Aug. 2022, doi:
no. 2, pp. 520–529, Jan. 2022, doi: 10.1002/cam4.4460. 10.3390/biomedicines10082049.
[104] A. Boonrod, A. Boonrod, A. Meethawolgul, and P. Twinprai, ‘‘Diagnostic [120] B. Lv, L. Wu, T. Huangfu, J. He, W. Chen, and L. Tan, ‘‘Traditional
accuracy of deep learning for evaluation of C-spine injury from lateral Chinese medicine recognition based on target detection,’’ Evidence-
neck radiographs,’’ Heliyon, vol. 8, no. 8, Aug. 2022, Art. no. e10372, Based Complementary Alternative Med., vol. 2022, pp. 1–9, Jul. 2022,
doi: 10.1016/j.heliyon.2022.e10372. doi: 10.1155/2022/9220443.
[105] C.-P. Tang, C.-H. Hsieh, and T.-L. Lin, ‘‘Computer-aided image [121] T. Till, S. Tschauner, G. Singer, K. Lichtenegger, and H. Till, ‘‘Devel-
enhanced endoscopy automated system to boost polyp and adenoma opment and optimization of AI algorithms for wrist fracture detection in
detection accuracy,’’ Diagnostics, vol. 12, no. 4, p. 968, Apr. 2022, doi: children using a freely available dataset,’’ Frontiers Pediatrics, vol. 11,
10.3390/diagnostics12040968. Dec. 2023, Art. no. 1291804, doi: 10.3389/fped.2023.1291804.
[106] H. Matsui, S. Kamba, H. Horiuchi, S. Takahashi, M. Nishikawa, [122] F. Varçın, H. Erbay, E. Çetin, I. Çetin, and T. Kültü, ‘‘End-to-end
A. Fukuda, A. Tonouchi, N. Kutsuna, Y. Shimahara, N. Tamai, and computerized diagnosis of spondylolisthesis using only lumbar X-rays,’’
K. Sumiyama, ‘‘Detection accuracy and latency of colorectal lesions J. Digit. Imag., vol. 34, no. 1, pp. 85–95, Feb. 2021, doi: 10.1007/s10278-
with computer-aided detection system based on low-bias evaluation,’’ 020-00402-5.
Diagnostics, vol. 11, no. 10, p. 1922, Oct. 2021, doi: 10.3390/diagnos- [123] G. Oreski, ‘‘YOLO*C—Adding context improves YOLO performance,’’
tics11101922. Neurocomputing, vol. 555, Oct. 2023, Art. no. 126655.
[107] C.-P. Tang, H.-Y. Chang, W.-C. Wang, and W.-X. Hu, ‘‘A novel computer- [124] M. Baghbanbashi, M. Raji, and B. Ghavami, ‘‘Quantizing YOLOv7: A
aided detection/diagnosis system for detection and classification of comprehensive study,’’ in Proc. 28th Int. Comput. Conf., Comput. Soc.
polyps in colonoscopy,’’ Diagnostics, vol. 13, no. 2, p. 170, Jan. 2023, Iran (CSICC), Iran, Jan. 2023, pp. 1–5.
doi: 10.3390/diagnostics13020170. [125] M. Hu, Z. Li, J. Yu, X. Wan, H. Tan, and Z. Lin, ‘‘Efficient-lightweight
[108] T. Ozturk, M. Talo, E. A. Yildirim, U. B. Baloglu, O. Yildirim, YOLO: Improving small object detection in YOLO for aerial images,’’
and U. R. Acharya, ‘‘Automated detection of COVID-19 cases using Sensors, vol. 23, no. 14, p. 6423, Jul. 2023.
deep neural networks with X-ray images,’’ Comput. Biol. Med., [126] B. Aldughayfiq, F. Ashfaq, N. Z. Jhanjhi, and M. Humayun, ‘‘YOLO-
vol. 121, Jun. 2020, Art. no. 103792, doi: 10.1016/j.compbiomed.2020. based deep learning model for pressure ulcer detection and classifica-
103792. tion,’’ Healthcare, vol. 11, no. 9, p. 1222, Apr. 2023.
[109] Z. Kong, H. Ouyang, Y. Cao, T. Huang, E. Ahn, M. Zhang, and [127] H. Ghabri, W. Fathallah, M. Hamroun, S. B. Othman, H. Bellali, H. Sakli,
H. Liu, ‘‘Automated periodontitis bone loss diagnosis in panoramic and M. N. Abdelkrim, ‘‘AI-enhanced thyroid detection using YOLO to
radiographs using a bespoke two-stage detector,’’ Comput. Biol. Med., empower healthcare professionals,’’ in Proc. IEEE Int. Workshop Mech.
vol. 152, Jan. 2023, Art. no. 106374, doi: 10.1016/j.compbiomed.2022. Syst. Supervision (IW_MSS), Nov. 2023, pp. 1–6.
106374. [128] D. Chaudhary, A. Mathur, A. Chauhan, and A. Gupta, ‘‘Assistive object
[110] W. Panyarak, K. Wantanajittikul, A. Charuakkra, S. Prapayasatok, and recognition and obstacle detection system for the visually impaired
W. Suttapak, ‘‘Enhancing caries detection in bitewing radiographs using using YOLO,’’ in Proc. 13th Int. Conf. Cloud Comput., Data Sci. Eng.
YOLOv7,’’ J. Digit. Imag., vol. 36, no. 6, pp. 2635–2647, Dec. 2023, doi: (Confluence), Jan. 2023, pp. 353–358.
10.1007/s10278-023-00871-4. [129] B. Wu, C. Pang, X. Zeng, and X. Hu, ‘‘ME-YOLO: Improved YOLOv5
[111] Y. Tian, D. Zhao, and T. Wang, ‘‘An improved YOLO nano model for for detecting medical personal protective equipment,’’ Appl. Sci., vol. 12,
dorsal hand vein detection system,’’ Med. Biol. Eng. Comput., vol. 60, no. 23, p. 11978, Nov. 2022.
no. 5, pp. 1225–1237, May 2022, doi: 10.1007/s11517-022-02551-x. [130] N. Andrade, T. Ribeiro, J. Coelho, G. Lopes, and A. Ribeiro, ‘‘Combining
[112] S. Pang, T. Ding, S. Qiao, F. Meng, S. Wang, P. Li, and X. Wang, ‘‘A novel YOLO and deep reinforcement learning for autonomous driving in public
YOLOv3-arch model for identifying cholelithiasis and classifying roadworks scenarios,’’ in Proc. 14th Int. Conf. Agents Artif. Intell., 2022,
gallstones on CT images,’’ PLoS ONE, vol. 14, no. 6, Jun. 2019, pp. 793–800.
Art. no. e0217647, doi: 10.1371/journal.pone.0217647. [131] M. U. Hadi, R. Qureshi, A. Shah, M. Irfan, A. Zafar, M. B. Shaikh,
[113] M. A. Azam, C. Sampieri, A. Ioppi, S. Africano, A. Vallin, D. N. Akhtar, J. Wu, and S. Mirjalili, ‘‘Large language models: A com-
Mocellin, M. Fragale, L. Guastini, S. Moccia, C. Piazza, L. S. Mattos, prehensive survey of its applications, challenges, limitations, and future
and G. Peretti, ‘‘Deep learning applied to white light and narrow prospects,’’ Tech. Rep., 2023, doi: 10.36227/techrxiv.23589741.v4.
band imaging videolaryngoscopy: Toward real-time laryngeal cancer
detection,’’ Laryngoscope, vol. 132, no. 9, pp. 1798–1806, Sep. 2022, doi:
10.1002/lary.29960.
[114] J.-Y. Tsai, I. Y.-J. Hung, Y. L. Guo, Y.-K. Jan, C.-Y. Lin, T. T.-F. Shih,
B.-B. Chen, and C.-W. Lung, ‘‘Lumbar disc herniation automatic MOHAMMED GAMAL RAGAB received the
detection in magnetic resonance imaging based on deep learning,’’ Bachelor of Science degree in software engineer-
Frontiers Bioeng. Biotechnol., vol. 9, Aug. 2021, Art. no. 708137, doi: ing from Universiti Teknologi PETRONAS in
10.3389/fbioe.2021.708137. 2019, following the completion of his undergradu-
[115] B. Zhang, J. Li, Y. Bai, Q. Jiang, B. Yan, and Z. Wang, ‘‘An improved
ate degree, he continued his studies at Universiti
microaneurysm detection model based on SwinIR and YOLOv8,’’
Bioengineering, vol. 10, no. 12, p. 1405, Dec. 2023, doi: 10.3390/bio-
Teknologi PETRONAS, pursuing the master’s
engineering10121405. degree by research in machine learning. Currently,
[116] P. Rouzrokh, T. Ramazanian, C. C. Wyles, K. A. Philbrick, J. C. Cai, he is continuing his academic pursuits by pursuing
M. J. Taunton, H. M. Kremers, D. G. Lewallen, and B. J. Erickson, ‘‘Deep the Ph.D. degree in information technology with
learning artificial intelligence model for assessment of hip dislocation Universiti Teknologi PETRONAS. His ongoing
risk following primary total hip arthroplasty from postoperative radio- research builds on his previous work, focusing on the development of new
graphs,’’ J. Arthroplasty, vol. 36, no. 6, pp. 2197–2203, Jun. 2021, doi: and innovative techniques for optimizing the performance of deep learning
10.1016/j.arth.2021.02.028. models.

VOLUME 12, 2024 57835

 M. G. Ragab et al.: Comprehensive Systematic Review of YOLO for Medical Object Detection

SAID JADID ABDULKADIR (Senior Member, RIZWAN QURESHI (Senior Member, IEEE)
IEEE) received the B.Sc. degree in computer received the Ph.D. degree from the City University
science from Moi University, the M.Sc. degree of Hong Kong, Hong Kong, in 2021. His Ph.D.
in computer science from Universiti Teknologi thesis focused on lung cancer drug resistance
Malaysia (UTM), and the Ph.D. degree in infor- analysis using molecular dynamics simulation and
mation technology from Universiti Teknologi machine learning. He joined the Fast School
PETRONAS (UTP). He is currently an Associate of Computing, National University of Computer
Professor and a member of the Centre for Research and Emerging Sciences, Karachi, Pakistan, as an
in Data Science (CeRDaS), UTP. He is involved Assistant Professor. He is currently with the
in flagship consultancy projects for PETRONAS College of Science and Engineering, Hamad Bin
under pipeline integrity, materials corrosion, and inspection. His research Khalifa University, Doha, Qatar. He has published his findings and methods
interests include machine learning, deep learning architectures, optimiza- in IEEE TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, IEEE
tions, and applications in predictive analytics. He is serving as the Treasurer JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, Pattern Recognition, and
for the IEEE Computational Intelligence Society Malaysia Chapter and the IEEE BIBM Conference. His research interests include AI applications in
Editor-in-Chief for Platform journal. life sciences, cancer data sciences, computer vision, and machine learning.

AMGAD MUNEER received the B.Eng. degree

(Hons.) in mechatronic engineering from Asia
Pacific University of Technology and Innovation
(APU), in 2018, and the master’s degree in
information technology from Universiti Teknologi
PETRONAS, Malaysia, in 2022. He is currently
with the Department of Imaging Physics, The
University of Texas MD Anderson Cancer Center,
Houston, TX, USA, as a Research Assistant. SAFWAN MAHMOOD AL-SELWI received the
He has authored several high-impact articles in bachelor’s degree in software engineering from
well-reputed journals and conferences. His research interests include Taiz University, Yemen, in 2012, and the master’s
computer vision, AI applications for cancer data sciences, manufacturing degree in computer applications from Banga-
data analytics, medical imaging, and bioinformatics. He is a reviewer of lore University, India, in 2018. He is currently
many international impact-factor journals. a Research Assistant with the Department of
Computer and Information Sciences, Universiti
Teknologi PETRONAS (UTP), Malaysia. He has
ALAWI ALQUSHAIBI received the B.Sc. degree
a total experience of eight years both in academic
in computer networks and security from Universiti
institutions and in the industry. His industry
Teknologi Malaysia, in 2012, and the master’s
working experience is related to Android applications and website
(by Research) degree from Universiti Teknologi
development. His research interests include artificial intelligence, machine
PETRONAS (UTP), in 2021. He is currently
learning, predictive and time-series analysis, metaheuristic algorithms, and
an Academic Researcher. During his academic
optimization.
journey, he has acquired knowledge and skills
in conducting independent research, producing
academic writing, and teaching computer science
courses. His research interests include machine
learning, data science, optimization, feature selection, classification, data
analytics, and image processing, specifically in generative adversarial
networks (GANs).

EBRAHIM HAMID SUMIEA received the B.S.

degree in software engineering from Asia Pacific HITHAM ALHUSSIAN received the B.Sc. and
University of Technology and Innovation (APU), M.Sc. degrees in computer science from the
in 2014, and the master’s degree in management. School of Mathematical Sciences, University of
He is currently pursuing the Ph.D. degree with Khartoum, Sudan, and the Ph.D. degree from Uni-
Universiti Teknologi PETRONAS, delving deeper versiti Teknologi PETRONAS (UTP), Malaysia.
into the realm of artificial intelligence. He honed He was a Postdoctoral Researcher with the High-
his skills in various programming languages and Performance Cloud Computing Center, UTP. He is
software development methodologies at APU. currently a Senior Lecturer with the Department
He is eager to explore the fusion of technology and of Computer and Information Sciences, UTP. His
business. His work aims to leverage DDPG’s potential to address complex current research interests include real-time parallel
problems, demonstrating the powerful capacity of AI to transform various and distributed systems, big data and cloud computing, data mining, and real-
sectors. His research interests include reinforcement learning, specifically in time analytics.
the area of deep deterministic policy gradient (DDPG).

57836 VOLUME 12, 2024