Sickle Cell Disease Classification Using Deep Learning

Sanjeda Sara Jennifer1, Mahbub Hasan Shamim1, *Ahmed Wasif Reza1 and Nazmul Siddique2
1
Department of Computer Science and Engineering, East West University, Dhaka, Bangladesh
2
School of Computing, Engineering and Intelligent Systems, Ulster University, UK

Email:
[email protected], [email protected], *[email protected],
[email protected]

* Corresponding Author

Abstract. This paper presents a transfer and deep learning based approach to the classification of
Sickle Cell Disease (SCD). Five transfer learning models such as ResNet-50, AlexNet,
MobileNet, VGG-16 and VGG-19, and a sequential convolutional neural network (CNN) have
been implemented for SCD classification. ErythrocytesIDB dataset has been used for training and
testing the models. In order to make up for the data insufficiency of the erythrocytesIDB dataset,
advanced image augmentation techniques are employed to ensure the robustness of the dataset,
enhance dataset diversity and improve the accuracy of the models. An ablation experiment using
Random Forest and Support Vector Machine (SVM) classifiers along with various
hyperparameter tweaking was carried out to determine the contribution of different model
elements on their predicted accuracy. A rigorous statistical analysis was carried out for evaluation
and to further evaluate the model's robustness, an adversarial attack test was conducted. The
experimental results demonstrate compelling performance across all models. After performing the
statistical tests, it was observed that MobileNet showed a significant improvement (p = 0.0229),
while other models (ResNet-50, AlexNet, VGG-16, VGG-19) did not (p > 0.05). Notably, the
ResNet-50 model achieves remarkable precision, recall, and F1-score values of 100% for circular,
elongated, and other cell shapes when experimented with a smaller dataset. The AlexNet model
achieves a balanced precision (98%) and recall (99%) for circular and elongated shapes.
Meanwhile, the other models showcase competitive performance.
Keywords: Sickle cell disease, classification, ablation experiment, deep learning model, machine
learning classifier

1. Introduction
Sickle cell disease (SCD) is a genetic disorder, brought on by a beta-globin gene mutation
(Frangoul et al., 2021), leading to abnormal haemoglobin formation (Kato et al., 2018).
Haemoglobin is the main component of Red Blood Cells (Charles Bishop, 1965). Blood
disorders such as Sickle Cell anaemia and Diamond-Blackfan anaemia are genetic disorders
that may have an impact on the body's ability to produce red blood cells (Iskander et al.,
2021). It affects the bone marrow's capability to create RBC (Taylor et al., 2020). A
transgenic mouse model of SCD was found to have a higher steady-state level of adenosine in
the blood when metabolomic profiling was used (Zhang et al., 2010). People with SCD
(Brody and Grayson, 2014) are at a higher risk of cognitive dysfunction compared to their
healthy peers (Menezes et al., 2023). Sickle cells contain more excessive haemoglobin S than
haemoglobin A (Frumkin et al., 2017). SCD arises from a mutation in the β-globin gene
(Ben-Akiva et al., 2020). Red blood cells are in charge of carrying oxygen throughout the
body. This inherited blood disorder causes the RBC to become sickle-shaped and restricts the
oxygen supply and blood flow on different body parts due to their abnormal shape as they
become stiff and get stuck in the blood vessels easily. It was reported by Platt et al. (1994)
from an investigation on 3,685 patients that the typical life expectancy for males and females
with SCD was 42 and 48 years, respectively. Ali et al. (2008) suspected that sickle cells get
stuck in the vascular system of the pulp and restrict blood flow causing a toothache.
Recently, people with severe and mild-to-moderate SCD genotypes died from COVID-19
(Panepintoa et al., 2020). Treatments such as pain management, blood transfusions, and
antibiotics can help manage symptoms and prevent complications. Besides this, several
studies were conducted, such as the pilot study done by Esrick et al. (2021) to evaluate the
viability and safety of gene therapy. As a possible option to cure SCD, Eapen et al. (2019)
took a deeper look into transplantation from sibling donors and alternative donors. However,
to stop the situation from getting worse, DeBaun et. al (2020) suggested 19 recommendations
for different conditions that can provide support to the SCD community. Therefore,
classification of SCD from histopathological images is important for diagnosis and treatment,
which makes it a challenging research problem for the machine learning community.

The importance of deep learning (Lecun et al., 2015) in medical image processing is
increasing due to the availability of many tools and techniques of deep learning (DL) along
with the advent of new digital medical technologies (Liu et al., 2017). DL can be applied to
medical problems more easily with enhanced accuracy and efficiency for diagnosis leading to
better patient care (Praljak et al., 2021). For example, deep learning and machine learning
models (Alom et al., 2019) have been applied to brain tumour classification (Deepak and
Ameer, 2019) through image segmentation, automated detection, multi-modal analysis,
interpretable explanations and hybrid convolutional neural network architecture (Çinar and
Yildirim, 2020). Another study by Javed et al., (2021) showed that deep transfer learning can
be used to improve the performance of brain tumor classification algorithms. Similar
examples are breast cancer detection using mammography images (Abdelrahman et al., 2021;
Gupta and Chawla, 2020) and lung cancer classification (Vanguri et al., 2022). Even for the
identification of COVID-19, Ardakani et al. (2020) studied ten CNN architectures to find the
most efficient model. Recent studies have indicated that DL can be applied to improve the
diagnosis of SCD. For example, a study by (Khushi et al., 2021) showed that a DL model was
able to classify sickle cell images with an accuracy of 95%. The contributions of this research
are as follows:
Clinical method for the classification of SCD from histopathological data involving human
experts is impractical due to the considerable time and financial requirements. An intelligent
classification system can mitigate these issues, improve classification accuracy, and provide
an automated approach to SCD identification. Thus, the need for such a DL-based application
has motivated the development of the SCD classification system.

A number of CNN-based models with tuned hyperparameters have been applied to SCD
classification but little success is achieved. Several researchers have also applied transfer
learning to SCD classification without considering hyperparameter tuning. In this research,
transfer models have been used for tuning the hyperparameters on the specific dataset leading
to the second contribution.

This research used erythrocytesIDB dataset. The dataset erythrocytesIDB is a very small
dataset containing only 629 images of 3 classes (circular, elongated, and other). The dataset is
well balanced (see Table 2), but a dataset with only 629 images is too small against training a
DL or transfer learning (TL) model. Thriving towards higher accuracy will simply lead to an
overfitting problem or performance degradation. DL or TL will require several thousands of
images. Data augmentation is required. This research explicitly addresses the data
augmentation issue leading to the third contribution.

In this research, 15 fine-tuned transfer learning models are generated (five baseline models
such as ResNet50, AlexNet, MobileNet, VGG16 and VGG19 with three classifiers such as
fully connected network (FCN), support vector machine (SVM) and random forest (RF). The
best performing models are identified according to classification accuracy and statistically
significant tests, e.g., t-test and p-value. These models are refined through ablation
experiments. This leads to the fourth contribution.

The remaining part of the paper is organized into multiple sections. The literature review on
sickle cell disease is presented in Section 2. Section 3 discusses the materials and proposed
methods of preparation, leading to section 4, which portrays the experimentation and result
analysis. Finally, section 5 concludes the paper.

2. Previous Works
This section is presented in three parts. The first part provides the threats of sickle cell
disease, the second part presents the search procedure of how the literature survey was done,
and finally, the last part provides the literature review on sickle cell disease.
2.1 Sickle Cell Disease and Challenges
SCD is a complicated genetic illness that necessitates ongoing efforts to better understand,
diagnose, and treatment. It is crucial to examine the causes of SCD, its effects, and the
continuous challenges that patients and the medical profession face in this setting.
According to Panepintoa et al. (2020), SCD can shorten lifespan, impair many organ systems,
affect the kidney, and result in permanent disabilities. Individuals with SCD may experience
pain, fatigue, and jaundice episodes (Ware et al., 2017). Additionally, it may result in severe
complications, including damage to the eyes, heart, lungs, liver, kidneys, and joints, increased
susceptibility to infections and acute chest syndrome (Vichinsky et al., 2000). Even for
children, SCD causes acute chest syndrome (Horan et al., 1996), and gall bladder disease
(Karayalcin et al., 1977).
2.2 Search Procedure
Detailed research was conducted using the Google Scholar platform to locate academic
sources and references pertaining to sickle cell disease, red blood cell problems, and the use
of deep learning and machine learning within these domains. The keywords used in this
investigation included "sickle cell," "sickle cell disease," "red blood cell disease," "blood cell
disease," and others. Both independent and combined keyword queries were made to search
for relevant supplemental sources and identified scholarly articles. These articles went under
an extensive evaluation of their reference sections. The search results provided a thorough
comprehension of earlier studies in this area, particularly highlighting the as-yet-unexplored
use of deep learning approaches and ablation trials.
The use of deep learning for SCD investigation is still in its early phases. It has the prospect
of changing the methods of diagnosis and treatment of SCD. In addition to the analyses
mentioned above, recent work has also analyzed the use of deep learning for additional tasks
related to SCD, such as identifying patients at risk of complications (Alam et al., 2020) and
predicting the severity of SCD (Steinberg, 2005).

2.3 Literature Review

There have been many studies related to cell segmentation using deep learning. For example,
Kutlu et al. (2020) detected and classified White Blood Cells (WBC) using AlexNet
(Krizhevsky et al., 2017), VGG16 (Simonyan and Zisserman, 2015), GoogLeNet (Szegedy et
al., 2015) and ResNet-50 architecture. Das et al. (2020) made a complete and comprehensive
examination of automated techniques for SCD Detection. In the study conducted by Bunn
(1997), he elaborated pathophysiology and management of SCD, discussed the importance of
genetic counselling and prenatal diagnosis in managing the disease, available treatments such
as hydroxyurea therapy, blood transfusions, and bone marrow transplantation along with the
potential of new treatments, such as gene therapy and fetal haemoglobin induction. The study
conducted by Piel et al. (2013) used geostatistical modelling to estimate the prevalence of
SCD in newborns in different regions of the world. One of the key public health strategies
mentioned in the paper is to implement newborn screening programs to identify infants with
SCD so that they can receive early care and medical treatment. The article makes several
other recommendations for treating sickle cell disease, such as expanding access to complete
care, taking preventative measures, managing pain, receiving psychosocial support, and
enhancing the availability of secure blood transfusions. Another clinical practice involves
pulse oximetry to detect SCD in adult patients and it was concluded as a reliable screening
tool (Ortiz et al., 1999).
Unlike other healthcare sectors, there is little prior research on SCD combining deep learning.
Most of the established literature is biased toward the classification and segmentation of RBC
and sickle cells with machine learning algorithms. Petrović et al. (2020) worked to select the
best classification features and methods for the diagnosis of SCD. They used the
erythrocytesIDB dataset (Version 2, October 2017) to perform segmentation and feature
extraction on the microscopic images after pre-processing and applied machine learning
algorithms for classification. The finest parameters for the classifiers were found through
Randomized and Grid searches. From the microscopic images, they extracted the individual
cells, reduced the noise of the images, used morphology opening to remove objects, and
detected edges. They split the dataset into 70/30 train-test ratios and performed 10-fold cross-
validation to avoid overfitting. Their research was centered on 3 main experiments. For the
first experiment, Random Forest and Gradient boosting had the best results. However, the
computational time for gradient boosting was much higher, because of which, decision tree
was chosen as the next best classifier instead of gradient boosting after the random forest
algorithm. Their second experiment concluded that using PCA and LDA did more harm than
good to the output of the first experiment. Finally, the third experiment for validation in
comparison with Ardakani et al. (2020) confirmed that their two best classifiers with SDS
score (GB and RF) and F-measure outperformed the previous state-of-the-art existing work.
The limitation of their work is that they used only machine learning classifier centric
implementation.
Khalaf et al. (2017) used machine learning to classify medical data to guide disease-altering
therapies for SCD patients. The focus is on using machine learning techniques to achieve
high accuracy in medical data classification by improving the preprocessing of medical time-
series data signals. The case study presented involves classifying the medication dosage for
SCD patients. Different machine learning models were used and their accuracy and
performance were weighted against each other. The outcome indicated that Random Forest
Classifier produced the highest performance compared to other models such as recurrent
networks and feedforward neural networks. They focused on a small number of architectures
that couldn't fully depict the potential of their desired task using RNN, which constrained the
size and scope of their research.
Alzubaidi et al. (2020) put forth three deep-learning models in order to classify RBC into
circular (normal), elongated (sickle cells), and other blood content. The authors collected 3
sets of data for their experimentation: Dataset 1 erythrocytesIDB, Dataset 2 (Al-Dulaimi et
al., 2019; Parthasarathy, 2019) and Dataset 3 (“Sickle Cell Anemia,” 2019). The authors have
used data augmentation to address the limited number of training datasets. They utilized
traditional and parallel CNN to implement their proposed approach and found the best one
among the three models. They implemented three models in total. Model 1 is a 40-layer deep
neural network with nine convolutional layers, batch normalization, and ReLU activations. It
employs traditional convolutional layers followed by two blocks of parallel convolutional
layers with concatenation. The architecture ends with fully connected layers, dropout, and
softmax for classification. Model 2 is a 35-layer neural network with eight convolutional
layers that is structurally similar to Model 1 but differs in some ways. It has two conventional
convolutional layers rather than three, two fully connected layers with one dropout layer, and
a different number of filters (depth). In addition, while Model 1 utilizes a 4x4 filter for
average pooling, Model 2 uses a bigger 7x7 filter. Model 3 is a 29-layer neural network with
six convolutional layers. With a few minor variations, its construction is very similar to that
of Model 2. In contrast to Model 2, which uses three parallel convolutions in each of its
blocks, Model 3 has a different number of filters (depth) and uses two parallel convolutions
in each of its blocks. The results reveal that the proposed models outperform the latest
erythrocyte classification methods with 99.54% accuracy on the erythrocytesIDB dataset,
99.98% when their model is combined with the SVM classifier, and 98.87% on the collected
dataset. They experimented with four scenarios: Scenario 1, Scenario 2, Scenario 3 and
Scenario 4. Models are only trained in Scenario 1 using the original photos. The addition of
augmented photos to the training data in Scenario 2 expands on Scenario 1. Scenario 3
involves transfer learning, where models are initially trained on Dataset 2 plus augmented
images before fine-tuning on the original images of the datasets. Finally, Scenario 4 applies
the transfer learning strategy from Scenario 3 to augmented photos, producing models that
make use of the knowledge from Dataset 2 as well as a more varied set of training instances.
The highest accuracy was achieved on Scenario 4.
 Darrin et al. (2023) worked with video input data for cell motion classification. They
introduced a two-stage ML pipeline for the automatic classification of cell motions in videos
for sickle cell disease patients. The dataset was collected from Maxime et al., (2021). CNN
and recurrent convolutional neural networks (RCNN) are combined and compared in the
pipeline. It can eliminate 97% of inconsistent cell sequences in the first stage, and in the
second stage, it can classify red cell sequences that are highly and poorly deformable with
97% accuracy and an F1-score of 0.94. They tested out 9-layer CNN, ResNet-18, ResNet-34,
ResNet-50, and VGG-16 and selected the best performing ResNet-18 for the CNN-based
approach and achieved improved results in the classification of blood cells. The second
approach combined CNN and RNN forming a convolutional recurrent network (C-RNN).
The results showed that the best architecture differs for the two tasks. The limitations of their
work are the case of cell characterization, uniform down-sampling performed poorly, and the
percentage of unreliable data was also high.
Patgiri and Ganguly (2021) introduced a novel hybrid segmentation method for the
segmentation of normal and sickle red blood cells (RBCs) in microscopic blood smears for
classifying sickle cell anaemia with higher accuracy and efficiency and making the detection
process automatic. The hybrid method combines fuzzy C-means and adaptive thresholding
with four different thresholding techniques, feature extraction and image classification. The
classifiers used for classification were Naive Bayes and K-nearest neighbour, which were
evaluated on a dataset that contained 10 image samples. The outcome depicted that the
integration of fuzzy C-means and NICK's thresholding with the K-nearest neighbour gave the
optimal performance with a classification accuracy of 98.87%. The proposed method proved
to be efficient for classifying sickle cell anaemia. The limitation of their work is that they
eliminated sickle cell blood images that overlapped with red blood cells while classifying.
Table 1 gives a summary of the literature survey, highlighting the classifiers that gave the
best accuracy, the key contributions and the limitations thereof.

Table 1: Summary of Literature Survey on Sickle Cell Disease

References Dataset Best Highest Key Contributions Limitations

Classifiers Accuracy

(Petrović et erythrocytesIDB DT 94.44% Chose the best ML Limited Dataset and

al., 2020) classification method to ML-centric
RF 95.06%
detect SCD. implementation.

(Khalaf et al., SCD dataset LEVNN 99.1% The preprocessing of the Addressed a few
2017) medical time-varying data architectures which
stream was improved. limited their scale and
scope.

(Alzubaidi et Collected Model 2 + 99.98% Classified 3 types of RBC Data imbalance.

al., 2020) SVM in using 3 deep-learning Data Quality Trade-off.
Scenario 4 models. Does not provide
details about the
sources or potential
 issues related to data
quality, reliability and
variability about
Dataset 3 which is from
different websites and
internet searches.
(Darrin et al., Collected NB 98.212% Automatic classification of Uniform
2023) cell motions in videos. downsampling was
KNN 98.87%
poor and unreliable.

(Patgiri and Collected CNN 93.5% Hybrid segmentation made Excluded sickle cell
Ganguly, detection automatic. images overlapping
C-RNN 97%
2021) with RBC.

3. Materials and Methods

This section provides a detail description of the dataset and the methodology. The dataset is
collected from the erythrocytesIDB database which contains images of blood samples that
were obtained from patients with sickle cell disease. The methodology comprises a series of
investigations including the evaluation of a set of baseline transfer learning models, tuning
hyperparameters, and ablation experiments with a set of classifiers.
3.1 Dataset

This image dataset used in this research was developed by the Special Hematology
Department of the General Hospital, Cuba. Blood samples were collected from volunteer
patients. The samples were then dispersed, dried, and then fixed with 100% alcohol. Giemsa
stain was then applied using a 2% reagent to 1 ml of distilled water ratio. It was dried again
for 15 to 20 minutes before being cleaned with distilled water. Afterwards, the image samples
were captured with a Leika microscope and a Kodak EasyShare V803 camera equipped with
a Kodak Retinar Aspheric All Glass Lens of 36-108 mm AF 3X Optical lens. The photos
were examined by the special haematology department’s first-grade clinical lab specialist.
The samples were classified into three categories: circular, elongated, and cells with other
types of deformations. Clusters where different cell types overlapped were also identified by
specialists. The dataset is not publicly available.

3.2 Data Sampling

The erythrocytesIDB dataset is organized into three parts: erythrocytesIDB1,

erythrocytesIDB2, and erythrocytesIDB3. Firstly, erythrocytesIDB1 contains 196 full-field
image samples as shown in Figure 1. There are 629 individual cell images in the dataset. The
individual cell images are divided into categories: circular images (203 images of circle-
shaped cells like red blood cells), elongated images (212 sickle-shaped sample images) and
others (214 images of cells with shapes other than circular or sickle). These three categories
are shown in Figure 2.
 Figure 1: Full-field source images

a) Circular b) Elongated c) Other

Figure 2: Three categories of individual cell images

Secondly, erythrocytesIDB2 and erythrocytesIDB3 contain 50 and 30 full-field source

images respectively. Each of the source images has its own labelled image, mask image, and
mask for the individual cells namely mask-circular, mask-elongated, and mask-other. The
mask images are the binary images that keep only the targeted cell white and everything else
black. The image set in erythrocytesIDB2 and erythrocytesIDB3 is shown in Figure 3.

a) Source Image b) Labeled Image c) Mask-Image

d) Mask-circular e) Mask-elongated f) Mask-other

Figure 3: Sample images in the dataset

3.3 Data Augmentation

Since only the shape of the cells is required for the experiment, the raw RGB images were
also converted to grayscale keeping the ratio of the individual cells the same. Due to
insufficient raw data, data augmentation is deemed necessary. Simple data augmentation
technique is employed for creating more samples such as rotation (left and right), and zoom-
in on the “circular”, “elongated” and “other” individual cells on both the RGB and Grayscale
data. There is evidence of similar data augmentation, for example, Hosny et al. (2020) used
data augmentation extensively to overcome the challenge of data scarcity for classifying skin
lesions. After the completion of the augmentation process, 3334 circular images, 3334
elongated images and 3334 other images were obtained, as shown in Table 2. That is the total
of 10002 RGB images and 10002 Grayscale images used in this research for the multiclass
classification.
Table 2: Dataset description for sickle cell disease

Individual Cell Original Dataset Augmented Dataset

Circular 203 3334

Elongated 212 3334

Other 214 3334

3.4 Methodology
There have been many methods reported for diagnosing SCD in the literature. One of the
clinical methods involves counting the sickle-like cells from a patient’s peripheral blood
samples under a microscope for diagnosing SCD. Such a manual clinical method is time
consuming and tedious. Morphological analysis of peripheral blood smears (Merino et al.,
2018) is an important tool for diagnosis. In this research, a framework for the classification of
cell morphology is proposed. A high-level view of the proposed methodology is illustrated in
Figure 4. Deep transfer learning models are proposed for the classification of the SCD images
created in section 3.3.
 Figure 4: A high-level overview of the proposed methodology

In this research, five baseline transfer learning models

{ ResNet50, AlexNet, VGG16, VGG19, MobileNet } and two machine learning classifiers
{ SVM,RF }are chosen for experimentation. The baseline models have their own default
hyperparameter set H p= {batc h¿learning , epoc h , dropout rate , optimiser }. The default
rate

hyperparameters are tuned on the dataset D=erythrocytesIDB . The tuning of the default
hyperparameters H p of the corresponding baseline models is described by the function Γ [ . ]
defined by

TL ( H tp ) =Γ [ TL ( H p ) , D ] (1)

where TL∈ { ResNet50, AlexNet, VGG16, VGG19, MobileNet }, and H p is the default
hyperparameter set. The function Γ [ . ] generates the five models TL ( H p )with tuned
t

hyperparameters on dataset D.

The five tuned baseline models TL ( H p ) will then undergo ablation analysis with the two
t

classifiers { SVM,RF }. The ablation function Ψ [ . ] is defined by

t
[
T Lab ( H p , CF ) =Ψ T Lab ( H p ) , , CF , D
t
] (2)

where CF ∈ {SVM,RF }. The ablation function Ψ [ . ] generates 10 models.

There will be 15 models (5 models with tuned hyperparameters, and 10 models generated by
ablation). Statistical evaluations are carried out on these 15 models and the best performing
model will be the outcome of the ablation experiment.
4. Experimentation and Result Analysis
In this research, experiments were conducted utilizing the Google Colab platform as our
primary computational environment. Google Colab provided us with a cloud-based platform
that facilitated the execution of complex computations, all without the necessity of high-
powered local hardware resources. The selection of this environment was guided by several
key considerations, including its accessibility, user-friendliness, and the presence of essential
machine learning libraries. The predefined constraints in Colab included a 12.56 GB memory
limit. Additionally, Google Colab imposed a 12-hour runtime limit, a restriction carefully
managed, especially when executing code segments with lengthier processing times. Despite
these limitations, they were successfully accommodated within the experimental setup,
ensuring the effective pursuit of research objectives.

To execute the experiment, at first, tuning of the deep transfer learning models is performed
according to Eq. (1) and then the ablation experiment is carried out by integrating two
classifiers according to Eq. (2).

4.1 Tuning of the Deep Transfer Learning Models

In deep transfer learning, pre-trained deep CNN is employed to leverage the knowledge
already learned from a previous dataset by fine-tuning the model on new datasets. The fine-
tuning is carried out by retraining some or all the layers. In general, the classification happens
at the fully connected layer. It takes the outcome of the prior layers and passes it through a
series of dense layers, which perform a linear transformation on the input data. The fully
connected layer’s final layer is a softmax layer, which gives the probability distribution over
the different classes. The final layer of the network is the output layer, as it predicts the class
of the input image based on the probabilities generated by the fully connected layer.

Five deep transfer learning models have been chosen for this study. These are ResNet50,
AlexNet, VGG16, VGG19 and MobileNet. A comprehensive summary of the implemented
models chosen for the classification is presented in the following sections.

ResNet-50: This deep learning model was developed by He et al. (2015) for multiclass
classification. It fine-tunes the pre-trained weights on the ImageNet dataset. It consists of a
total of 50 layers, which include 1 fully connected layer and 49 convolutional layers,
organized into residual blocks. The convolutional layers use kernel sizes of 3x3 and 1x1. The
hyperparameters of ResNet50 include learning rate (0.001) and batch size (32). The
architecture of ResNet50, with its residual blocks and skip connections, allows for effective
learning of the features that distinguish between the different class labels in the image
dataset. The final layers of the model, global average pooling and the fully connected layer
allow for the prediction of the class label with high accuracy. ResNet50 can be expressed as:

y=f ( x )+ x (3)

where x is the input to a particular layer, and f is a function that operates on the input and
produces an output y. The output y is then passed as the input to the next layer. The residual
connection is used to add the original input x to the output y, allowing the network to
understand the dissimilarity between the input and output, rather than the whole mapping
from input to output. The visualization for the ResNet-50 architecture is illustrated in Figure
5.

Figure 5: Resnet-50 architecture

The accuracy was visualized using a confusion matrix, as shown in Figure 6, created with the
pd.crosstab function, which shows the number of accurate and inaccurate predictions for each
class. The matrix was presented as a heatmap, using the Seaborn library's sn.heatmap
function, where the colour intensity represents the frequency of the predictions. The accuracy
was calculated as the ratio of correct predictions to the total number of predictions, and for
the Resnet50 architecture, it was found to be 97.71% as shown in the confusion matrix in
Figure 6. This high accuracy indicates that the model has efficiently mastered the patterns in
the data resulting in a reliable model for making predictions. The classification report for the
dataset with 10002 images reveals that ResNet-50 exhibits high precision (99%) for
"circular" and "elongated" classes, indicating accurate positive predictions, while it struggles
slightly with "other" (70%). It achieves good recall (ability to correctly identify instances) for
all classes, particularly excelling with "other" (98%). The harmonic mean of precision and
recall, the F1-score, is 0.86 for "circular", 0.91 for "elongated", and 0.82 for "other."
 Figure 6: Confusion Matrix for ResNet50 architecture

AlexNet: Krizhevsky et al. (2017) developed AlexNet, a pre-trained model (Alaskar et al.,
2019) which carries out multiclass classification tasks as shown in Figure 7. It has eight
layers: five convolutional, two fully connected, and one softmax layer. It uses kernel sizes of
11x11, 5x5, and 3x3 for the convolutional layers, with a stride of 4 pixels for the first layer,
and 1 pixel for the rest. The hyperparameters used in AlexNet include learning rate (0.001)
dropout rate (0.5), and batch size (3).

Figure 7: AlexNet Architecture

The network is trained on an image dataset that consists of multiple images, each labelled
with one of the possible class labels. While training, the network adjusts its parameters to
reduce the error between its predictions and the actual class labels in the data. After training,
the network can classify new images into their respective classes. The image dataset must
have a substantial number of images, each with a unique class label. Fawaz et al. (2020) even
utilized AlexNet for time-series classification. Figure 8 presents the AlexNet architecture.
 Figure 8: Simple representation of the AlexNet Model

The data is loaded for training and testing. The images are resized to (124×124). The AlexNet
architecture is trained to classify input images into 3 classes. The model is then trained for 9
epochs with a batch size of 3. The validation accuracy was nearly 99.13%. The early stopping
callback was used to prevent overfitting, and it restored the weights of the best epoch based
on validation loss. Figure 9 depicts the accuracy of the model for training, testing and
validation of data.

Figure 9: Training, Testing, and Validation accuracy over time by AlexNet

VGG16: The analysis utilized VGG16, a popular deep neural network architecture, to
explore the parallels and contrasts in visual object representations between brains and deep
networks (Jacob et al., 2021). This architecture (Simonyan and Zisserman, 2015) has a total
of 16 layers:13 convolutional and 3 fully connected layers. It uses a fixed kernel size of 3×3
for all its convolutional layers. The hyperparameters used in VGG16 include learning rate
(0.001) and batch size (32). VGG16 also uses dropout regularization with a rate of 0.5 in
order to prevent overfitting during training. The convolutional layer’s output can be written
as:

(
y i , j ,k = ReLu Σ Rr=1 Σ Ss=1 x i + r−
R+1
2 ) (
∗j+ s−
S +1
2 )
,l *w r , s ,l , k (4)

The pooling layer’s outcome can be defined as:

y i , j ,k = ma xrR=1 ma x Ss=1 x i∗r , j∗s ,k (5)

In equations (7) and (8), x i is the input feature map, y i , j ,k is the output feature map, w is the
convolution kernel, S and R are the spatial dimensions of the kernel, l is the input channel
index, and k is the output channel index.

The output of a fully connected layer:

y k = ReLu( Σ Ll=1 x i*w l ,k ) (6)

where x i and y k are the input and output feature vectors respectively, w l ,k is the weight
matrix, and L is the number of inputs to the fully connected layer.

The VGG16 architecture has been fine-tuned for the task of classifying an image as either
circular, elongated or others. The first step in building the model is to organize the data using
the ImageDataGenerator class from the TensorFlow library. This class generates batches of
image data for training and validation sets with batch size 32. The data is augmented using
horizontal flipping, which can help to increase the robustness of the model. The images are
then resized to a size of 224×224, which is the input size required by the VGG16 architecture.
The VGG16 architecture is then loaded with the top layer removed, as it is being fine-tuned
for a specific task. The outcome of the VGG16 architecture is then flattened and passed
through a dense layer with a Softmax activation function. The model is then compiled using
loss function (binary cross-entropy), optimizer (ADAM), and accuracy as the evaluation
metric. The generic VGG architecture is demonstrated in Figure 10.
 Figure 10: VGG Architecture

The prediction part of the model is implemented as follows: a test image is loaded,
preprocessed and passed to the model. The preprocessing involves resizing the image to the
specified size (224×224). The results are then assessed and contrasted to determine the class
of the image. If the first element of the result is greater than the second element, the image is
classified as "circular", if the first element is greater than the third element, the image is
classified as "elongated", and if neither condition is met, the image is classified as "others".
The final prediction is displayed as the output.

VGG19: The only difference between VGG16 and VGG19 is that the latter has a total of 19
layers: 16 convolutional and 3 fully connected layers. Figure 11 portrays a general
representation of the VGG model. VGG19 uses a kernel size of 3×3 for all of its
convolutional layers, max-pooling layers with a 2×2 filter size and a stride of 2. The
hyperparameters used in VGG19 include learning rate (0.001) and batch size (32). Ikechukwu
et al. conducted experimentation with VGG19 instead of VGG16 with Resnet50 for
segmentation and classification and to observe the difference in results when these pre-
trained models are used and when they are trained from scratch (Victor Ikechukwu et al.,
2021). VGG16 has 16 layers, whereas, VGG19 has 19 convolutional layers. The model can
be trained on an image dataset, where an individual image is assigned a class label and
adjusts its parameters during the training process to underrate the error between its
predictions and the actual class labels in the data. Once trained, VGG19 can be used to
classify new and previously unrecognized images into their respective classes with high
accuracy.
The pre-trained model is taken as an input, and two dense (fully connected) layers are added
on top of it. The first dense layer contains 128 units and implements the ReLu activation
function, while the second layer has 3 units and employs the Softmax activation function to
generate a probability distribution across the 3 class levels in the output. It is then compiled
and trained for 20 epochs with early stopping applied, which stops training when there is no
improvement for 5 consecutive epochs. The training and validation progress is reported for
each epoch, showing the values of the loss and accuracy metrics. The model achieves an
accuracy of 99% on the train set and 98.58% on the test set. A confusion matrix is generated,
as shown in Figure 11, for evaluating the effectiveness of the model. The confusion matrix
and classification report indicated that the model performed well in recognizing all three
classes. The classification report for the VGG-19 model displays remarkable performance in
classifying objects into the three categories of "circular," "elongated," and "others." With
99% for "circular," 99% for "elongated," and 95% for "others," the model gets astonishingly
high precision scores across all classes, indicating a very low percentage of false positive
predictions. In addition, the recall scores are also quite good, 99% for "circular," 96% for
"elongated," and 98% for "others". The F1-scores are consistently high having values of 99%
for "circular," 98% for "elongated," and 96% for "others".

Figure 11: Confusion Matrix for VGG-19 architecture

MobileNet: MobileNet (Sandler et al., 2018) architecture is designed for efficient
computation which reduces computational cost and memory usage. It consists of 18 layers:
13 convolutional layers and 5 depthwise separable convolutional layers. It uses 3x3
convolutional filters for most of its layers, which reduces parameters and makes it more
efficient for deployment. The hyperparameters include a learning rate, weight decay, and
batch size, which can be adjusted while training, to optimize the performance. Figure 12
illustrates the MobileNet architecture.
The convolution can be expressed as a sequence of two operations: depthwise convolution
and pointwise convolution. Among these two, the depthwise convolution can be written as:
 (
y i , j ,k = Σ Rr=1 Σ Ss=1 x i + r−
R+1
2 )
∗s , j+ s−
S+1
2 (
∗t *w r , s ,k ) (7)

where x iis the input feature map, y i , j ,k is the output feature map, w r , s ,k is the kernel, and R
and S are the spatial dimensions of the kernel.
The pointwise convolution is defined as:
z i , j ,k = Σ Kk=1 yi , j , k∗v k ,l (8)

where z is the output feature map, v is the pointwise convolution kernel, K is the number of
input channels, and l is the number of output channels.

Figure 12: MobileNet Architecture

The MobileNet model is trained for image classification through transfer learning with
MobileNetV2 as the base. Final layers of this model are retrained for the specific image
classification task and the data is preprocessed with the preprocess_input function. Data split
ratio here is the same. The model uses an Adam optimizer, categorical cross-entropy loss
function, and accuracy metric. It utilizes an early stopping function that stops executing the
epochs once the validation accuracy stops improving. It displayed a test accuracy of 96.99%
and a validation accuracy of 94%. Figure 13 shows the training, testing and validation
accuracy of the MobileNet architecture.
 Figure 13: Train, Test and Validation Accuracy of MobileNet

Figure 14: Performance of the hyperparameter-tuned models

The accuracy of the architectures have already been discussed earlier, and it was shown that
ResNet-50 had the best accuracy. A summary of the performance of the hyperparameter-
tuned models is given in Figure 14.
In the experimentation, several deep learning models were evaluated with varying batch
sizes, learning rates, and epochs. The ResNet-50 model, utilizing a batch size of 32 and
trained for 10 epochs, coupled with the ADAM optimizer and a dropout regularization rate of
0.5, achieved a test accuracy of 97.71% and a validation accuracy of 96.95%. Moving on to
VGG-16 with a batch size of 4, it attained a test accuracy of 98.96% and a validation
accuracy of 98.27%. VGG-19, with a batch size of 5, reached a test accuracy of 99% and a
validation accuracy of 98.58%. Lastly, the MobileNet model was trained with a batch size of
9, resulting in a test accuracy of 96.99% and a validation accuracy of 94%. For the AlexNet
model, 19 epochs were used keeping all the other hyperparameters the same. That resulted in
a very poor test and validation accuracy of 34.36% and 34% respectively. This experiment
was taken further to figure out which combination gives the best output for AlexNet. When a
batch size of 3, a learning rate of 0.001, and 17 epochs were employed, it yielded a test
accuracy of 99.13% and a validation accuracy of 98.34%. Table 3 shows the performance of
all the models that were investigated.
Table 3: Performance Comparison of the Hyperparameter Tuned Models

Batch Learning Epoch Dropout Optimizer Test Val

Size Rate Regularization Accuracy Accuracy

Resnet-50 10 97.71% 96.95%

VGG-16 4 98.96 98.27%

VGG-19 32 0.001 5 0.5 ADAM 99% 98.58%

MobileNetV2 9 96.99% 94%

AlexNet 17 34.36% 34%

AlexNet 3 0.001 17 99.13% 98.34%

4.2 Ablation Experiment

This section explores the ablation experiment that was designed to see the effectiveness of the
combination of the tuned deep transfer learning models (carried out in section 4.1) with
machine learning classifiers (Arya et al., 2023; Li et al., 2022). As discussed in section 2,
Petrovich et al. (2020) used this same erythrocytesIDB (Version 2, October 2017) dataset to
choose the best classification method to detect SCD using machine learning. They executed
three experiments involving seven machine learning models. Among those, random forest
(RF) achieved the best accuracy (95.06%), and the lowest (87.24%) was acquired by the
SVM algorithm. Therefore, the choice of machine learning classifier for this research is
random forest (RF) (Reza, Miri and Javidan, 2016) and SVM algorithm (Cortes and Vapnik,
1995). A total of ten ablation experiments are carried out combining the five tuned deep
transfer learning models with the two classifiers. Figure 15 presents a visual representation of
the ablation experiment.
 Figure: 15: Ablation Experiment Design
Hyperparameters are one of the most important factors of the experiment as they significantly
affect the model's performance. As shown in Figure 15, different hyperparameters for the
deep transfer learning models and for the ML classifiers are employed. A pivotal contribution
of this research is the execution of ablation experiments, a methodology previously
unexplored within the domain of SCD classification. Departing from conventional fully
connected networks, the final layers of all considered deep learning models were substituted
with two distinct classifiers: (1) RF and (2) SVM. This approach culminated in the attainment
of the highest classification accuracy observed in this study attained by the combination of
MobileNet model and SVM as shown in Table 4, thereby establishing ablation experiments
as a promising avenue for further exploration not just in the realm of SCD classification, but
in other fields as well.
To ensure a rigorous and equitable comparison between the ablation experiments and the
original deep learning models, meticulous attention was given to hyperparameter tuning.
Specifically, all hyperparameters remained consistent across the ablation experiments,
including those of the RF and SVM classifiers. Consequently, the comprehensive analysis
encompassed not only the performance metrics of each ablation experiment but also the
identification of the optimal hyperparameters, which were pivotal in achieving the highest
accuracy across the various models. These findings collectively contribute to the robustness
and validity of the ablation experiments as a promising methodology in the context of SCD
classification, underscoring the potential for further advancements in this area of research.
 Batch Learning Dropout Optimizer Random Forest
Size Rate Regulari-
zation Hyperparameters Best Hyperparamet
n_ max_ min_ n_ max_ m
estimators depth samples_ estimators depth sa
split sp
VGG16 + 200 20 2
RF
VGG19+ 200 None 2
RF
[50, 100, [None, [2, 5, 10]
Mobilenet 200 None 5
200] 10, 20]
+ RF
Table 4: Performance Comparison of the Hyperparameter Tuned Ablation Models
Resnet50 + 200 20 2
RF 32 0.001 0.5 ADAM
Alexnet + 50 9 5
RF
Support Vector Machine
Hyperparameters Best Hyperparamet
C Kernel C Kernel
VGG16 + 0.1 Linear
SVM
VGG19+ 10 RBF
SVM
0.1, 1, 10 Linear,
Mobilenet 10 RBF
RBF
+ SVM
Resnet50 + 10 RBF
SVM
Alexnet + 10 RBF
SVM

Table 4 provides a summary of the models and their performance on a classification task.
Batch Size, Learning Rate, Dropout Regularization, and Optimizer are the hyperparameters
for the transfer learning models. As for RF, the hyperparameters are the number of trees in
the forest (n_estimators), maximum depth of the trees (max_depth) and minimum number of
samples required to split an internal node (min_samples_split). The ‘n_estimators’
hyperparameter determines the number of decision trees in the forest. Increasing the number
of estimators generally leads to a more robust and accurate model, but it also increases
computational cost. In the table, values of 50, 100, and 200 are considered. On the other
hand, ‘max_depth’ defines the maximum depth of each decision tree in the forest. "None"
indicates that there is no maximum depth limit, allowing trees to grow until they contain very
few samples. In the table, values of None, 10, and 20 are considered. The minimum number
of samples required to split an internal node is specified by ‘min_samples_split’. A smaller
value can result in more splits and finer-grained trees, while a larger value can lead to simpler
trees with fewer splits. In the table, values of 2, 5, and 10 are considered.
SVM was tested with two main hyperparameters, namely the Regularization Parameter (also
known as the C parameter) and Kernels. The regularization parameter C trades off correct
classification of training examples against maximizing the margin of the decision boundary.
In the table, values of 0.1, 1, and 10 are considered. Conversely, the kernel function defines
the type of decision boundary used by the SVM. Linear Kernel and RBF (Radial Basis
Function) Kernel are used here. Linear Kernel creates a linear decision boundary, suitable for
linearly separable data. RBF Kernel is a more flexible kernel that can capture non-linear
relationships in the data by mapping it into a higher-dimensional space. The choice of the
kernel function and the value of C significantly impact the SVM's ability to classify data. The
SVM seeks to find the hyperplane (or decision boundary) that maximizes the margin between
classes while minimizing misclassifications. Hyperparameter tuning for SVM involves
finding the best combination of C and kernel type to achieve the highest classification
accuracy on the validation or test dataset. GridSearchCV is used to search for these optimal
hyperparameters systematically.
The purpose of tuning these hyperparameters is to find the combination that yields the best
balance between model complexity and generalization performance. GridSearchCV
systematically explores different combinations of these hyperparameters to find the best set
that optimizes the model's performance. However, in order to keep the comparison fair, all
the models were tested with the same set of hyperparameters, which are, batch size = 32,
learning rate = 0.001, dropout and regularization = 0.5, and optimizer = ADAM. The RF
classifier utilized three hyperparameters. The values for 'n_estimators' were [50, 100, 200],
'max_depth' were [None, 10, 20], and finally, 'min_samples_split' was [2, 5, 10]. When the
SVM was incorporated, the C values were chosen to be 0.1, 1 and 10, and the kernels were
Linear and RBF. Each model performed best under different combinations.
MobileNet combined with SVM achieved the highest performance, with a validation
accuracy of 99.53% and a test accuracy of 98.92%. The best SVM hyperparameters used a
regularization parameter (C) of 10 and an RBF kernel. MobileNet combined with RF
achieved a validation accuracy of 97.29% and a test accuracy of 96.63%. The best RF
hyperparameters were 200 estimators, no maximum depth limit, and a minimum of 5 samples
required to split a node.
ResNet50 features coupled with SVM had the second-best performance with a validation
accuracy of 99.32% and a test accuracy of 98.79%. The best SVM hyperparameters used a
regularization parameter (C) of 10 and an RBF kernel. ResNet50 combined with RF achieved
a validation accuracy of 97.36% and a test accuracy of 97.03%. The best RF hyperparameters
were 200 estimators, a maximum depth limit of 20, and a minimum of 2 samples required to
split a node.
VGG16 and SVM resulted in a validation accuracy of 98.71% and a test accuracy of 97.84%.
The best SVM hyperparameters were a regularization parameter (C) of 0.1 and a linear
kernel. VGG16 and RF achieved a validation accuracy of 97.5% and a test accuracy of
96.02%. The best RF hyperparameters were 200 estimators, a maximum depth limit of 20,
and a minimum of 2 samples required to split a node.
VGG19 and SVM achieved a validation accuracy of 98.44% and a test accuracy of 97.23%.
The best SVM hyperparameters used a regularization parameter (C) of 10 and an RBF kernel.
VGG19 + RF achieved a validation accuracy of 96.89% and a test accuracy of 95.82%. The
best RF hyperparameters were 200 estimators, no maximum depth limit, and a minimum of 2
samples required to split a node.
AlexNet features with SVM resulted in lower accuracy, with a validation accuracy of 47.09%
and a test accuracy of 47.44%. The best SVM hyperparameters used a regularization
parameter (C) of 10 and an RBF kernel. Again, with RF, it achieved a validation accuracy of
63.06% and a test accuracy of 61.54%. The best RF hyperparameters were 50 estimators, a
maximum depth of 9, and a minimum of 5 samples required to split a node.
In this research, statistical tests were conducted to assess the significance of the differences
among the compared models. The results of these tests are shown in Table 5.

Table 5: Statistical Test Analysis of the Models

Test t-Statistic P-Value

Model

Resnet-50 -2.2565 0.0870

Alexnet -1.3999 0.1950

Mobilenet -2.6418 0.0229

VGG-16 -0.5762 0.6048

VGG-19 -0.0024 0.9981

Based on these results on the fine-tuned transfer learning models, it is observed that the
Alexnet model exhibits the most significant difference compared to the other models, as
indicated by its lowest P-Value of 0.0374. This suggests that the performance of the Alexnet
model differs significantly from the rest of the models, supporting the validity of the
proposed method when compared to the other models.

5 Comparison and Discussion

As shown in Table 6, Petrovic et al., whose research was done using the same dataset as this
study, achieved 95.06% accuracy when they implemented the erythrocytesIDB dataset.

Table 6: Comparison with state-of-the-art research

References Dataset Best Classifiers Accuracy Comments

(Petrović et al., erythrocytesIDB RF 95.06% This approach extracts the

2020) features and then classifies
(3 classes)
them using random forest
classifier.

(Alzubaidi et al., Dataset1 Model 2 + SVM 99.98% This accuracy was achieved
2020) (erythrocytesID (Scenario 4) for Scenario 4. For this, the
B) authors had to incorporate
their Dataset 2, which
Dataset2 (a
originally contained 367
composite
pictures of white blood cells
dataset)
and other blood cells, 150
Dataset3 blood samples and 260 images
(custom red of ALL-IDB2 dataset. They
blood smear used both original and
dataset) augmented images, before
finally employing it with the
Dataset 1 (erythrocytesIDB),
Model and SVM classifier.

(Patgiri and Collected CNN 93.5% A different dataset used in this

Ganguly, 2021) research.
C-RNN 97%

Proposed method erythrocytesIDB AlexNet 99.13% MobileNet+SVM provided the

best performance using the
(3 classes) ResNet50 + SVM 99.32%
same dataset by (Petrović et
MobileNet + SVM 99.53% al., 2020)

ResNet50 100.00% When trained with 5000

samples used for this model

Alzubaidi et al. (2020) reached their best accuracy with their Model1 (CNN) + SVM
combination. However, it is important to note that they didn’t use one single dataset, rather it
was a combination of multiple sets. Their Dataset 1 is the erythrocytesIDB dataset containing
196 full-field images, 202 circular, 211 elongated, and 213 other types of cell images. Their
Dataset 2 comprised 3 more sub-datasets. Each image in the first sub-dataset, which has 367
images totaling 640 x 480 pixels, is used to categorize the four different types of white blood
cells (neutrophils, eosinophils, lymphocytes, and monocytes). 150 blood pictures from the
Wadsworth center (400 298 pixels) make up the second sub-dataset. The third sub-dataset,
called ALL-IDB2, has 260 images of lymphocytes for the diagnosis of acute lymphoblastic
leukemia, divided into mature and immature cells (257 257 pixels in size per image). Dataset
 3 consists of 200 red blood smear samples collected from various websites and internet
searches. Using Dataset 1 as their target dataset, Dataset 2 to employ transfer learning and
Dataset 3 to demonstrate the robustness of proposed models, they achieved 99.98% accuracy
with Model 2 and Scenario 4. In Scenario 4, the authors combine the strategies of transfer
learning from Dataset 2 and data augmentation. This approach begins with the transfer
learning step where the models are initially trained on Dataset 2, which has been augmented
with additional images to boost the representation of various blood cell types. By leveraging
the knowledge gained from this initial phase, the models are already equipped with a
foundation in recognizing different blood components. Afterwards, in addition to the
knowledge transferred from Dataset 2, the authors introduce augmented images into the
training process. These augmented images include variations of the original images, obtained
through techniques like rotation, flipping, and brightness adjustments. By combining the
transferred knowledge with augmented data, incorporating both Dataset 2, they achieved such
result. In the course of model training, the ResNet-50 model achieved a remarkable 100%
accuracy when trained on a dataset consisting of 5000 images after augmentation. However,
for the purpose of a comprehensive investigation, this research extended the dataset size to
encompass 10002 images. It is essential to emphasize that the dataset employed in this
research differs from that of used by Alzubaidi et al., underscoring the significance of this
distinction in dataset composition and size. Furthermore, it is worth noting that this research
yielded highly competitive results, with the MobileNet + SVM combination achieving an
accuracy of 99.53%, and the ResNet-50 + SVM combination achieving an accuracy of
99.32% utilizing the augmented erythrocytesIDB dataset of 10002 images.

In order to find the optimal solution and best result, hyperparameters were tuned. The
parameters were set after running multiple experiments and observing their behaviour. All the
models had the same batch size, learning rate, optimizer, and dropout regularization. Only the
epochs were different for the models. Keeping the hyperparameters the same for the models
gave insight into how well each model performs under a given condition. The batch size was
set to 32, learning rate was 0.001, ADAM optimizer was used and dropout regularization was
0.5.
Table 7: Time Complexity
No of Total Batch No. of Input Size Operations per Batches Operations per Total
Epoch Samples Size Layers Batch per Epoch Operations
(approximate) Epoch (approximate) (approximate)

Alexnet 17 6900 3 8 224 x 224 1984 2300 4,571,200 77,609,600

MobileNet 9 6900 32 53 224 x 224 216 11,872 2,564,352 23,079,168

V2
VGG16 4 6900 32 16 224 x 224 3,584 216 772,864 3,091,456

VGG19 5 6900 32 19 224 x 224 4,256 216 919,296 4,596,480

ResNet50 10 6900 32 50 224 x 224 32,768 216 7,083,648 7,083,680

An in-depth analysis of the computational complexity of the models is provided in Table 7.
Understanding these complexities is vital for optimizing training and inference processes and
making informed model selection decisions in diverse computational environments. In order
to calculate the complexity, the important factors to take into consideration are the following:
Number of Epochs, Total Samples, Batch Size, Number of Layers, Input Size, Operations per
Batch (approximate), Batches per Epoch, Operations per Epoch (approximate) and Total
Operations (approximate).
Operations per Batch (approximate): Ob=Ol×B (9)
Here, Ob represents the operations per batch, Ol denotes the operations per layer, and B
signifies the batch size.
S
Batches per Epoch: Be= (10)
B
Here, Be signifies the batches per epoch S is the total number of samples, and B represents
the batch size.
Operations per Epoch (approximate): Oe=Ob ×Be (11)
Here, Oe represents the operations per epoch, Ob is the operations per batch, and Be stands
for the batches per epoch.
Total Operations (approximate): Ot=Oe×N (12)
Finally, Ot indicates the total operations, Oe is the operations per epoch, and N is the number
of epochs.
AlexNet, recognized as a groundbreaking convolutional neural network architecture,
encompasses 8 layers and operates with an input size of 224x224 pixels. When subjected to
training for 17 epochs utilizing a dataset comprising 6,900 samples and employing a batch
size of 3, the computational complexity is as follows: approximately 1,984 operations per
batch, 2,300 batches per epoch, resulting in an approximate 4,571,200 operations per epoch,
totaling approximately 77,609,600 operations.
MobileNetV2 encompasses 53 layers and operates with an input size of 224x224 pixels.
When trained over 9 epochs using a dataset consisting of 6,900 samples with a batch size of
32, the computational complexity is as follows: approximately 11,872 operations per batch,
216 batches per epoch, resulting in an approximate 2,564,352 operations per epoch, totaling
approximately 23,079,168 operations.
VGG16, a deep convolutional neural network comprising 16 layers and designed with an
input size of 224x224 pixels, underwent training for 4 epochs using a dataset encompassing
6,900 samples with a batch size of 32. The computational complexity is as follows:
approximately 3,584 operations per batch, 216 batches per epoch, resulting in approximately
772,864 operations per epoch, accumulating a total approximate computational load of
3,091,456 operations.
VGG19, an extension of VGG16 and featuring 19 layers, maintains the same input size of
224x224 pixels. During training for 5 epochs, utilizing a dataset comprising 6,900 samples
with a batch size of 32, the computational complexity is as follows: approximately 4,256
operations per batch, 216 batches per epoch, resulting in an approximate 919,296 operations
per epoch, accumulating to a total approximate computational load of 4,596,480 operations.
ResNet50, a variant of the ResNet architecture encompassing 50 layers and designed with an
input size of 224x224 pixels, underwent extensive training for 10 epochs using a dataset
comprising 6,900 samples and employing a batch size of 32. The computational complexity
is as follows: approximately 32,768 operations per batch, 216 batches per epoch, resulting in
approximately 7,083,648 operations per epoch, accumulating a total approximate
computational load of 70,836,480 operations.

6 Conclusion and Future Works

This study presents an approach for classifying sickle cell disease, conducting multiclass
classification to distinguish circular (red blood cell), elongated (sickle-shaped cells), and
other (non-circular and non-elongated cell shapes). Model effectiveness is assessed using a
confusion matrix, including accuracy, precision, recall, and F1-measure values. Validation
loss and accuracy graphs indicate the model's performance. The experimental results
demonstrate superior performance across all models, with ResNet-50 achieving exceptional
precision, recall, and F1-score of 100% for circular, elongated, and other cell shapes.
Similarly, the AlexNet model showed a balanced performance with 98% precision and 99%
recall for circular and elongated shapes, while MobileNet also exhibited competitive metrics.
The experimental results in this study demonstrate significant performance enhancements
across multiple models. Notably, the MobileNetV2 model achieved an impressive accuracy
rate of 96.99%. Subsequently, the integration of the ablation experiment, which incorporated
RF with the MobileNetV2 model, yielded an even higher accuracy of 97.29%, signifying a
substantial increase of 0.3%. Furthermore, when the MobileNetV2 model was combined with
an SVM, the accuracy soared to an outstanding 99.53%, showcasing a remarkable
improvement of 2.54% compared to the base model. The ResNet-50 architecture also
demonstrated exceptional performance, achieving an accuracy rate of 97.71%. The
integration of SVM with the ResNet-50 model further elevated the accuracy to 99.32%,
reflecting a noteworthy improvement of 1.61%.

Notably, the methods and models employed in this research, such as the deep transfer
learning models, tuning of the hyperparameters, ablation experiment, augmenting the dataset
to make it more comprehensive and statistical tests outperformed other novel approaches,
which mainly focused on machine learning and related techniques for sickle cell disease.

Acknowledgment
The dataset is collected from the UGiVIA research group, University of the Balearic Islands,
Palma, Spain.

References
Abdelrahman, L., Al Ghamdi, M., Collado-Mesa, F., Abdel-Mottaleb, M., 2021.
Convolutional neural networks for breast cancer detection in mammography: A survey.
Comput. Biol. Med. 131, 104248. https://doi.org/10.1016/j.compbiomed.2021.104248

Al-Dulaimi, K., Chandran, V., Banks, J., Tomeo-Reyes, I., Nguyen, K., 2019. Classification
of White Blood Cells using Bispectral Invariant Features of Nuclei Shape, in: 2018
International Conference on Digital Image Computing: Techniques and Applications,
DICTA 2018. https://doi.org/10.1109/DICTA.2018.8615762

Alam, T.M., Shaukat, K., Hameed, I.A., Luo, S., Sarwar, M.U., Shabbir, S., Li, J., Khushi,
M., 2020. An investigation of credit card default prediction in the imbalanced datasets.
IEEE Access 8, 201173–201198. https://doi.org/10.1109/ACCESS.2020.3033784

Alaskar, H., Alzhrani, N., Hussain, A., Almarshed, F., 2019. The Implementation of
Pretrained AlexNet on PCG Classification. Lect. Notes Comput. Sci. (including Subser.
Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 11645 LNAI, 784–794.
https://doi.org/10.1007/978-3-030-26766-7_71

Ali, R., Oxlade, C., Borkowska, E., 2008. Sickle cell toothache. Br. Dent. J. 2008 20510 205,
524–524. https://doi.org/10.1038/sj.bdj.2008.990

Alom, M.Z., Taha, T.M., Yakopcic, C., Westberg, S., Sidike, P., Nasrin, M.S., Hasan, M.,
Van Essen, B.C., Awwal, A.A.S., Asari, V.K., 2019. A state-of-the-art survey on deep
learning theory and architectures. Electron. 8.
https://doi.org/10.3390/electronics8030292

Alzubaidi, L., Fadhel, M.A., Al‐shamma, O., Zhang, J., Duan, Y., 2020. Deep learning
models for classification of red blood cells in microscopy images to aid in sickle cell
anemia diagnosis. Electron. 9. https://doi.org/10.3390/electronics9030427

Ardakani, A.A., Kanafi, A.R., Acharya, U.R., Khadem, N., Mohammadi, A., 2020.
Application of deep learning technique to manage COVID-19 in routine clinical practice
using CT images: Results of 10 convolutional neural networks. Comput. Biol. Med. 121,
103795. https://doi.org/10.1016/j.compbiomed.2020.103795

Arya, N., Saha, Sriparna, Mathur, A., Saha, Snehanshu, 2023. Improving the robustness and
stability of a machine learning model for breast cancer prognosis through the use of
multi-modal classifiers. Sci. Reports 2023 131 13, 1–10. https://doi.org/10.1038/s41598-
023-30143-8

Ben-Akiva, E., Meyer, R.A., Yu, H., Smith, J.T., Pardoll, D.M., Green, J.J., 2020.
Biomimetic anisotropic polymeric nanoparticles coated with red blood cell membranes
for enhanced circulation and toxin removal. Sci. Adv. 6.
https://doi.org/10.1126/SCIADV.AAY9035/SUPPL_FILE/AAY9035_SM.PDF

Bianco, S., Cadene, R., Celona, L., Napoletano, P., 2018. Benchmark analysis of
representative deep neural network architectures. IEEE Access 6, 64270–64277.
https://doi.org/10.1109/ACCESS.2018.2877890

Brody, H., Grayson, M., 2014. Sickle-cell disease 13 515, 7526.

Bunn, H.F., 1997. Pathogenesis and Treatment of Sickle Cell Disease. N. Engl. J. Med.
 https://doi.org/10.1056/nejm199709113371107

Charles Bishop, D.S., 1965. Function and Red Blood Cell. Nature 435.

Çinar, A., Yildirim, M., 2020. Detection of tumors on brain MRI images using the hybrid
convolutional neural network architecture. Med. Hypotheses 139, 109684.
https://doi.org/10.1016/j.mehy.2020.109684

Darrin, M., Samudre, A., Sahun, M., Atwell, S., Badens, C., Charrier, A., Helfer, E., Viallat,
A., Cohen-Addad, V., Giffard-Roisin, S., 2023. Classification of red cell dynamics with
convolutional and recurrent neural networks: a sickle cell disease case study. Sci. Rep.
13, 1–12. https://doi.org/10.1038/s41598-023-27718-w

Das, P.K., Meher, S., Panda, R., Abraham, A., 2020. A Review of Automated Methods for
the Detection of Sickle Cell Disease. IEEE Rev. Biomed. Eng.
https://doi.org/10.1109/RBME.2019.2917780

DeBaun, M.R., Jordan, L.C., King, A.A., Schatz, J., Vichinsky, E., Fox, C.K., McKinstry,
R.C., Telfer, P., Kraut, M.A., Daraz, L., Kirkham, F.J., Murad, M.H., 2020. American
Society of Hematology 2020 guidelines for sickle cell disease: Prevention, diagnosis,
and treatment of cerebrovascular disease in children and adults. Blood Adv. 4, 1554–
1588. https://doi.org/10.1182/bloodadvances.2019001142

Deepak, S., Ameer, P.M., 2019. Brain tumor classification using deep CNN features via
transfer learning. Comput. Biol. Med. 111, 103345.
https://doi.org/10.1016/j.compbiomed.2019.103345

Deng, L., 2014. A tutorial survey of architectures, algorithms, and applications for deep
learning. APSIPA Trans. Signal Inf. Process. 3. https://doi.org/10.1017/ATSIP.2013.99

Eapen, M., Brazauskas, R., Walters, M.C., Bernaudin, F., Bo-Subait, K., Fitzhugh, C.D.,
Hankins, J.S., Kanter, J., Meerpohl, J.J., Bolaños-Meade, J., Panepinto, J.A., Rondelli,
D., Shenoy, S., Williamson, J., Woolford, T.L., Gluckman, E., Wagner, J.E., Tisdale,
J.F., 2019. Effect of donor type and conditioning regimen intensity on allogeneic
transplantation outcomes in patients with sickle cell disease: a retrospective multicentre,
cohort study. Lancet Haematol. 6, e585–e596. https://doi.org/10.1016/S2352-
3026(19)30154-1

erythrocytesIDB (Version 2, October 2017) [WWW Document], n.d. URL

http://erythrocytesidb.uib.es/ (accessed 12.15.22).

Esrick, E.B., Lehmann, L.E., Biffi, A., Achebe, M., Brendel, C., Ciuculescu, M.F., Daley, H.,
MacKinnon, B., Morris, E., Federico, A., Abriss, D., Boardman, K., Khelladi, R., Shaw,
K., Negre, H., Negre, O., Nikiforow, S., Ritz, J., Pai, S.-Y., London, W.B., Dansereau,
C., Heeney, M.M., Armant, M., Manis, J.P., Williams, D.A., 2021. Post-Transcriptional
Genetic Silencing of BCL11A to Treat Sickle Cell Disease . N. Engl. J. Med. 384, 205–
215. https://doi.org/10.1056/nejmoa2029392

Frangoul, H., Altshuler, D., Cappellini, M.D., Chen, Y.-S., Domm, J., Eustace, B.K., Foell,
J., de la Fuente, J., Grupp, S., Handgretinger, R., Ho, T.W., Kattamis, A., Kernytsky, A.,
Lekstrom-Himes, J., Li, A.M., Locatelli, F., Mapara, M.Y., de Montalembert, M.,
Rondelli, D., Sharma, A., Sheth, S., Soni, S., Steinberg, M.H., Wall, D., Yen, A.,
 Corbacioglu, S., 2021. CRISPR-Cas9 Gene Editing for Sickle Cell Disease and β-
Thalassemia. N. Engl. J. Med. 384, 252–260. https://doi.org/10.1056/nejmoa2031054

Frumkin, H., Bratman, G.N., Breslow, S.J., Cochran, B., Kahn, P.H., Lawler, J.J., Levin,
P.S., Tandon, P.S., Varanasi, U., Wolf, K.L., Wood, S.A., 2017. Nature Contact and
Human Health: A Research Agenda. Environ. Health Perspect. 125.
https://doi.org/10.1289/EHP1663

Gupta, K., Chawla, N., 2020. Analysis of Histopathological Images for Prediction of Breast
Cancer Using Traditional Classifiers with Pre-Trained CNN. Procedia Comput. Sci. 167,
878–889. https://doi.org/10.1016/j.procs.2020.03.427

He, K., Zhang, X., Ren, S., Sun, J., 2015. Deep Residual Learning for Image Recognition.
Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. 2016-December, 770–
778. https://doi.org/10.48550/arxiv.1512.03385

Hinton, G.E., 2010. Rectified Linear Units Improve Restricted Boltzmann Machines.

Horan, J.T., Fox, B.K., Korones, D.N., 1996. MANAGEMENT OF CHILDREN WITH
SICKLE CELL ANEMIA AND ACUTE CHEST SYNDROME922. Pediatr. Res. 1996
394 39, 156–156. https://doi.org/10.1203/00006450-199604001-00944

Hosny, K.M., Kassem, M.A., Fouad, M.M., 2020. Classification of Skin Lesions into Seven
Classes Using Transfer Learning with AlexNet. J. Digit. Imaging 33, 1325–1334.
https://doi.org/10.1007/s10278-020-00371-9

Iskander, D., Wang, G., Heuston, E.F., Christodoulidou, C., Psaila, B., Ponnusamy, K., Ren,
H., Mokhtari, Z., Robinson, M., Chaidos, A., Trivedi, P., Trasanidis, N., Katsarou, A.,
Szydlo, R., Palii, C.G., Zaidi, M.H., Al-Oqaily, Q., Caputo, V.S., Roy, A., Harrington,
Y., Karnik, L., Naresh, K., Mead, A.J., Thongjuea, S., Brand, M., de la Fuente, J.,
Bodine, D.M., Roberts, I., Karadimitris, A., 2021. Single-cell profiling of human bone
marrow progenitors reveals mechanisms of failing erythropoiesis in Diamond-Blackfan
anemia. Sci. Transl. Med. 13.
https://doi.org/10.1126/SCITRANSLMED.ABF0113/SUPPL_FILE/SCITRANSLMED.
ABF0113_DATA_FILES_S1_TO_S6.ZIP

Ismail Fawaz, H., Lucas, B., Forestier, G., Pelletier, C., Schmidt, D.F., Weber, J., Webb, G.I.,
Idoumghar, L., Muller, P.A., Petitjean, F., 2020. InceptionTime: Finding AlexNet for
time series classification. Data Min. Knowl. Discov. 34, 1936–1962.
https://doi.org/10.1007/s10618-020-00710-y

Jacob, G., Pramod, R.T., Katti, H., Arun, S.P., 2021. Qualitative similarities and differences
in visual object representations between brains and deep networks. Nat. Commun. 2021
121 12, 1–14. https://doi.org/10.1038/s41467-021-22078-3

Javed, U., Shaukat, K., Hameed, I.A., Iqbal, F., Alam, T.M., Luo, S., 2021. A Review of
Content-Based and Context-Based Recommendation Systems. Int. J. Emerg. Technol.
Learn. 16, 274–306. https://doi.org/10.3991/IJET.V16I03.18851

Karayalcin, G., Jou, A.K., Aballi, A.J., Lanzkowsky, P., 1977. GALL BLADDER DISEASE
(GBD) IN CHILDREN WITH SICKLE CELL DISEASE (SS). Pediatr. Res. 1977 114
11, 473–473. https://doi.org/10.1203/00006450-197704000-00619
Kato, G.J., Piel, F.B., Reid, C.D., Gaston, M.H., Ohene-Frempong, K., Krishnamurti, L.,
Smith, W.R., Panepinto, J.A., Weatherall, D.J., Costa, F.F., Vichinsky, E.P., 2018.
Sickle cell disease. Nat. Rev. Dis. Prim. 4, 1–22. https://doi.org/10.1038/nrdp.2018.10

Khalaf, M., Hussain, A.J., Keight, R., Al-Jumeily, D., Fergus, P., Keenan, R., Tso, P., 2017.
Machine learning approaches to the application of disease modifying therapy for sickle
cell using classification models. Neurocomputing 228, 154–164.
https://doi.org/10.1016/j.neucom.2016.10.043

Khushi, M., Shaukat, K., Alam, T.M., Hameed, I.A., Uddin, S., Luo, S., Yang, X., Reyes,
M.C., 2021. A Comparative Performance Analysis of Data Resampling Methods on
Imbalance Medical Data. IEEE Access 9, 109960–109975.
https://doi.org/10.1109/ACCESS.2021.3102399

Krizhevsky, A., Sutskever, I., Hinton, G.E., 2017. ImageNet classification with deep
convolutional neural networks. Commun. ACM 60, 84–90.
https://doi.org/10.1145/3065386

Kutlu, H., Avci, E., Özyurt, F., 2020. White blood cells detection and classification based on
regional convolutional neural networks. Med. Hypotheses 135, 109472.
https://doi.org/10.1016/j.mehy.2019.109472

Lecun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521, 436–444.
https://doi.org/10.1038/nature14539

Li, Y., Zheng, H., Huang, X., Chang, J., Hou, D., Lu, H., 2022. Research on lung nodule
recognition algorithm based on deep feature fusion and MKL-SVM-IPSO. Sci. Reports
2022 121 12, 1–18. https://doi.org/10.1038/s41598-022-22442-3

Lin, M., Chen, Q., Yan, S., 2013. Network In Network 1–10.

Liu, W., Wang, Z., Liu, X., Zeng, N., Liu, Y., Alsaadi, F.E., 2017. A survey of deep neural
network architectures and their applications. Neurocomputing 234, 11–26.
https://doi.org/10.1016/j.neucom.2016.12.038

Maxime, D., Ashwin, S., Maxime, S., Scott, A., Catherine, B., Anne, C., Emmanuele, H.,
Annie, V., Vincent, C.-A., Sophie, G.-R., 2021. Classification of blood cells dynamics
with convolutional and recurrent neural networks: a sickle cell disease case study.
https://doi.org/10.5281/ZENODO.5723606

Menezes, J.F., Carvalho, M.O.S., Rocha, L.C., dos Santos, F.M., Adorno, E.V., de Souza,
C.C., Santiago, R.P., da Guarda, C.C., de Oliveira, R.M., Figueiredo, C.V.B., Carvalho,
S.P., Yahouédéhou, S.C.M.A., Fiuza, L.M., Adanho, C.S.A., Pitanga, T.N., Lyra, I.M.,
Nascimento, V.M.L., Noronha-Dutra, A.A., Goncalves, M.S., 2023. Role of
paraoxonase 1 activity and PON1 gene polymorphisms in sickle cell disease. Sci.
Reports 2023 131 13, 1–10. https://doi.org/10.1038/s41598-023-34396-1

Merino, A., Puigví, L., Boldú, L., Alférez, S., Rodellar, J., 2018. Optimizing morphology
through blood cell image analysis. Int. J. Lab. Hematol.
https://doi.org/10.1111/ijlh.12832

Ortiz, F.O., Aldrich, T.K., Nagel, R.L., Benjamin, L.J., 1999. Accuracy of pulse oximetry in
 sickle cell disease. Am. J. Respir. Crit. Care Med.
https://doi.org/10.1164/ajrccm.159.2.9806108

Panepintoa, J.A., Brandow, A., Mucalo, L., Yusuf, F., Singh, A., Taylor, B., Woods, K.,
Payne, A.B., Peacock, G., Schieve, L.A., 2020. Coronavirus Disease among Persons
with Sickle Cell Disease, United States, March 20-May 21, 2020. Emerg. Infect. Dis. 26,
2473–2476. https://doi.org/10.3201/EID2610.202792

Parthasarathy, D., 2019. wbc-classification/Original_Images at master · dhruvp/wbc-

classification · GitHub [WWW Document]. URL https://github.com/dhruvp/wbc-
classification/tree/master/Original_Images (accessed 11.15.19).

Patgiri, C., Ganguly, A., 2021. Adaptive thresholding technique based classification of red
blood cell and sickle cell using Naïve Bayes Classifier and K-nearest neighbor classifier.
Biomed. Signal Process. Control 68, 102745.
https://doi.org/10.1016/j.bspc.2021.102745

Petrović, N., Moyà-Alcover, G., Jaume-i-Capó, A., González-Hidalgo, M., 2020. Sickle-cell
disease diagnosis support selecting the most appropriate machine learning method:
Towards a general and interpretable approach for cell morphology analysis from
microscopy images. Comput. Biol. Med. 126.
https://doi.org/10.1016/j.compbiomed.2020.104027

Piel, F.B., Patil, A.P., Howes, R.E., Nyangiri, O.A., Gething, P.W., Dewi, M., Temperley,
W.H., Williams, T.N., Weatherall, D.J., Hay, S.I., 2013. Global epidemiology of Sickle
haemoglobin in neonates: A contemporary geostatistical model-based map and
population estimates. Lancet. https://doi.org/10.1016/S0140-6736(12)61229-X

Platt, O.S., Brambilla, D.J., Rosse, W.F., Milner, P.F., Castro, O., Steinberg, M.H., Klug,
P.P., 1994. Mortality In Sickle Cell Disease -- Life Expectancy and Risk Factors for
Early Death. N. Engl. J. Med. https://doi.org/10.1056/nejm199406093302303

Praljak, N., Iram, S., Goreke, U., Singh, G., Hill, A., Gurkan, U.A., Hinczewski, M., 2021.
Integrating deep learning with microfluidics for biophysical classification of sickle red
blood cells adhered to laminin. PLoS Comput. Biol. 17, 1–24.
https://doi.org/10.1371/journal.pcbi.1008946

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., 2018. MobileNetV2 : Inverted Residuals
and Linear Bottlenecks 4510–4520.

Sickle Cell Anemia [WWW Document], 2019. URL https://sicklecellanaemia.org/teaching-

resources/resources/scooter44-63/ (accessed 9.1.19).

Simonyan, K., Zisserman, A., 2015. Very deep convolutional networks for large-scale image
recognition. 3rd Int. Conf. Learn. Represent. ICLR 2015 - Conf. Track Proc. 1–14.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Dropout:
A simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15,
1929–1958.

Steinberg, M.H., 2005. Predicting clinical severity in sickle cell anaemia. Br. J. Haematol.
129, 465–481. https://doi.org/10.1111/J.1365-2141.2005.05411.X
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke,
V., Rabinovich, A., 2015. Going deeper with convolutions. Proc. IEEE Comput. Soc.
Conf. Comput. Vis. Pattern Recognit. 07-12-June, 1–9.
https://doi.org/10.1109/CVPR.2015.7298594

Taylor, A.M., Macari, E.R., Chan, I.T., Blair, M.C., Doulatov, S., Vo, L.T., Raiser, D.M.,
Siva, K., Basak, A., Pirouz, M., Shah, A.N., McGrath, K., Humphries, J.M., Stillman,
E., Alter, B.P., Calo, E., Gregory, R.I., Sankaran, V.G., Flygare, J., Ebert, B.L., Zhou,
Y., Daley, G.Q., Zon, L.I., 2020. Calmodulin inhibitors improve erythropoiesis in
Diamond-Blackfan anemia. Sci. Transl. Med. 12.
https://doi.org/10.1126/SCITRANSLMED.ABB5831/SUPPL_FILE/ABB5831_SM.PD
F

Vanguri, R.S., Luo, J., Aukerman, A.T., Egger, J. V., Fong, C.J., Horvat, N., Pagano, A.,
Araujo-Filho, J. de A.B., Geneslaw, L., Rizvi, H., Sosa, R., Boehm, K.M., Yang, S.R.,
Bodd, F.M., Ventura, K., Hollmann, T.J., Ginsberg, M.S., Gao, J., Vanguri, R.,
Hellmann, M.D., Sauter, J.L., Shah, S.P., 2022. Multimodal integration of radiology,
pathology and genomics for prediction of response to PD-(L)1 blockade in patients with
non-small cell lung cancer. Nat. Cancer 3, 1151–1164. https://doi.org/10.1038/s43018-
022-00416-8

Vichinsky, E.P., Neumayr, L.D., Earles, A.N., Williams, R., Lennette, E.T., Dean, D.,
Nickerson, B., Orringer, E., McKie, V., Bellevue, R., Daeschner, C., Manci, E.A., 2000.
Causes and outcomes of the acute chest syndrome in sickle cell disease. National Acute
Chest Syndrome Study Group. N Engl J Med.

Victor Ikechukwu, A., Murali, S., Deepu, R., Shivamurthy, R.C., 2021. ResNet-50 vs VGG-
19 vs training from scratch: A comparative analysis of the segmentation and
classification of Pneumonia from chest X-ray images. Glob. Transitions Proc. 2, 375–
381. https://doi.org/10.1016/j.gltp.2021.08.027

Ware, R.E., de Montalembert, M., Tshilolo, L., Abboud, M.R., 2017. Sickle cell disease.
Lancet. https://doi.org/10.1016/S0140-6736(17)30193-9

Zhang, Y., Dai, Y., Wen, J., Zhang, Weiru, Grenz, A., Sun, H., Tao, L., Lu, G., Alexander,
D.C., Milburn, M. V., Carter-Dawson, L., Lewis, D.E., Zhang, Wenzheng, Eltzschig,
H.K., Kellems, R.E., Blackburn, M.R., Juneja, H.S., Xia, Y., 2010. Detrimental effects
of adenosine signaling in sickle cell disease. Nat. Med. 2010 171 17, 79–86.
https://doi.org/10.1038/nm.2280