A COMPARATIVE ANALYSIS USING DEEP

LEARNING TECHNIQUES FOR DETECTING

LUNG CANCER
Vishali Joyce P Nancy Arokia Rani S
Postgraduate Student Assistant Professor
Department of Computer Science Department of Computer Science
Stella Maris College(Autonomous) Stella Maris College(Autonomous)
Chennai, India Chennai, India
[email protected] [email protected]
Abstract— Convolutional neural networks are a
specialized kind of deep learning model that is adapted
primarily for the handling of grid-like data, especially
images. Its applications currently involve image
classification, object detection, and segmentation because
CNNs can automatically detect patterns in an image through
learning spatial hierarchies of features. Lately, CNNs have
been significantly developed in the healthcare sector
especially when dealing with medical images and visual data.
One important application is lung cancer detection, which
arises from the uncontrolled multiplication of abnormal cells
in the lungs. Lung cancer detection based on CNNs at an
early stage is considered an advanced stage of detecting
cancerous patterns in medical images. This paper compares
three of the most widely used CNN architectures such as
VGG-16 ,DenseNet-121 and ResNet-50 with the goal of
identifying an effective model for real-time lung cancer Fig. 1.1: Lung Cancer Diagram
detection. The analysis is to further explore the steps of lung
Among various deep learning models developed, CNN
cancer detection that distinguish between cancerous and
non-cancerous data from CT scans, thereby providing very seems to stand out as a great instrument of image pattern
valuable insights into the efficacy of each method for early recognition and analysis. Designed to take grid-like data
disease detection. structures such as images, CNNs are particularly effective
at capturing spatial hierarchies and identifying key features
Keywords— Deep Learning, Convolutional Neural Networks, across scales of the input data, establishing itself as one of
VGG-16, DenseNet-121, ResNet-50 the most suitable tools for utilizing in medical imaging
analysis where critical patterns in intricate datasets can be
I. INTRODUCTION automatically learned and subsequently detected.
Lung cancer is one of the leading causes of deaths due to
The cases of the discovery of lung cancer is subjected to
cancer around the world, largely because it is diagnosed
processing within CT scan images by CNNs for
way too late. Survival chances would be much better if the
identification and differentiation of malignant and non-
patients were diagnosed at an early stage of cancer. The
malignant tissues. It has made tremendous progress in
traditional diagnostic methods- like biopsies and X-rays-
early disease diagnosis through the processing of large
take a much longer time, and the diagnosis hardly becomes
dimensional medical images along with the extraction of
possible on an early-stage basis. Advances in the medical
subtle features. Specifically, CNNs have been used in
imaging technology, in particular, with regard to
order to obtain the precision and efficiency of diagnosing
Computed Tomography (CT) scans, have dramatically
lung cancer and therefore assist in proper on-time
improved lung cancer early detection. The poor
treatment for the patient. This improves the detection of
interpretation of such medical images, however, remains
abnormal tissue growth in the lung and, therefore reduces
an arduous challenge to radiologists that has only now
overall mortality from lung cancer using CNNs.
been well addressed through deep learning technologies.
The last few years have seen dramatic changes in the field This paper presents the evaluation of three of the most
of medicine and are perhaps going to change with what has widely discussed CNN architectures: VGG-16, ResNet-50,
appeared as a gift-artificial intelligence and deep learning and DenseNet-121 with a lung cancer detection problem
methods. on CT scan images. The models have achieved great
success in benchmarked image classification tasks, but the
real-time performance of the model with respect to lung
cancer detection is still an exploratory field of research.
VGG-16 is simple and deep, ResNet-50 uses residual
connections to combat the problem of vanishing gradients
and DenseNet-121 emphasizes feature reuse through dense
connections.
 Fig 2.1.1 VGG-16 Architecture

Simonyan and Zisserman introduced the first VGG-16

model in 2014. It is one of the first deep learning models
that particuraly performance well on image classification
Fig. 1.2: Lung cancer CT - Scan tasks. This architecture adopts a very straightforward and
simple 16-layer design comprising 13 convolutional layers,
This study tries to compare the performance of such which then feed into max-pooling layers. Such feature
architectures in the hope that it will validate which model extraction and reducing the dimensionality in this design
is more effective in drawing the difference between will always make the design significantly simple, as it only
cancerous and non-cancerous lung tissues, a guide towards has 3x3 convolutional filters well-suited to medical
their potential application in early lung cancer diagnosis. imaging applications, such as the detection of lung cancer.
Fully connected layers are the last layers of VGG-16; these
enable the classification function to allocate images to
II. LITERATURE SURVEY predefined categories, such as benign and malignant
tumors. This models introduces non linear and to improve
efficiency after every convolutional layer uses the ReLU
A. “Improved VGG-16 Convolutional Neural function.
Network Based Lung Cancer Classification and
Identification on Computed Tomography” Although VGG-16 is enormously successful, it has several
shortcomings in both computational complexity and time.
The field of medical image processing has experienced Deeper layers and higher parameters also lead to a higher
significant growth in interest for the recent development of computational cost for the architecture. Ammendments that
deep learning, and particularly CNNs. A significant researchers suggested to deal with these constraints are that
number of studies focused on CNN architectures that may adding additional convolutional layers were proposed,
help in the early detection of various diseases, like lung along with increasing kernel sizes, similar to the Improved
cancer, by extracting meaningful features from high- VGG-16 model. These changes have been proven to
dimensional image data. enhance accuracy and reduce false positives in the lung
cancer detection tasks, and enhanced accuracy to 97% with
Traditionally, the methods employed for the detection of the new model at a rate of 86% compared with the VGG-
lung cancer include biopsies, X-rays, and CT scans, which 16 model (vgg16).
depend on the interpretation of images by radiologists.
Machine learning and especially CNNs rectify some of the More broadly, this shows that the application of VGG-16
flaws in these traditional methods because they enable the and all of its variants in lung cancer detection reflects a
automation of image analysis for very accurate early major milestone in medical image processing, which is
detection. Among several CNNs applied to this problem, efficiency in accuracy to diagnose diseases or conditions at
VGG-16, ResNet, and DenseNet are prominently used. an early stage.
There are several studies that have been done in literature
on CNN-based approaches with regard to the classification
of lung cancer nodules from CT images. For instance, B. “ResNet based Lung Nodules Detection from
Sharma and Jindal, in 2011, developed an image Computed Tomography Images”
processing technique for lung cancer detection with
excellent results but unscalable due to expert-extracted This paper proposes an efficient lung nodule detection
features. Farahani et al., in 2015, addressed ensemble system by a ResNet-inspired 2D CNN architecture. This
learning techniques on lung CT images, obtaining proposed system works on two stages: the pre-processing
increased detection rates but with model complexity and segmentation, followed by the detection of a
limited in its application in real-time. Jin et al. (2016) and candidate nodule. This first stage of preprocessing adjusts
Golan et al. (2017) were CNN-based approaches, which the slices from the CT scan in such a way that it removes,
presented even superior results of deep models in in effect, all lung regions from other irrelevant structures,
comparison with traditional machine learning classifiers, which include bones, tissues, and gaps, based on the aim
such as SVM. of nodule detection. Scan data is converted into
Hounsfield units (HU). Voxels outside the bore of the
scanner are assigned 0 values. This way only important
parts are input in scan. Then a thresholding process is
executed, in which a set of HU value, such as 600, is used
to separate lung regions from non-lung regions, and it
produces a binary image.
Morphological operations fine-tune the binary image. number of key advantages. The method used for
Erosion removes small, inconsequential objects, such as preprocessing stage is a relatively simple morphological
noise, and dilation brings back in the eroded areas so they operation, one which optimizes computational time and
are the same as before the erosion is applied. If both cost while highly accurately producing segmentations of
erosion and dilation are applied together, the boundaries the lung region. A significant reduction in the number of
are then smoothed out so that the segmented lung region is layers constitutes the residual-based architecture for fewer
complete and accurate. Another closing operation is residual blocks than in more complex models but still
applied, using a small structuring element only to fill up gives optimal performance in the aspect of demonstrating
minute holes, for making the segmented lung region high accuracy even with a more compact network. The
appear more continuous and noise reduced. Output is the general reduction in the number of layers will lead to a
lung-only segmented image. This is then divided into decrease in computational complexity, implying a more
smaller 32x32 pixel windows, which represents different efficient system without affecting the detection accuracy.
areas of the lung and would be the input to the next stage,
nodule detection.
C. “Lung Cancer Detection Using Deep Learning”
Using a deep convolutional neural network based on the
ResNet architecture, it enters into the next phase for Deep learning approaches for medical imaging analysis in
potential lung nodule detection. This specific architecture the field of lung cancer detection attain much attention
of ResNet is used for training networks even deeper due to their property of extracting complex features of
without degradation in performance, because of residual medical images automatically. Traditional methods of
connections. These shortcut connections bypass one or diagnosis rely on the interpretation of the imaging data; be
more layers, passing along the more efficient gradient it a CT scan or an X-ray, where human error and time
flow at the training phase, preventing the vanishing consume the valuable resource. To counter this,
gradient problem, and hence ensuring that the model researchers often favor automated approaches which rely
actually learns from deep layers. The network architecture on AI, especially deep learning methods, including
comprises six modified residual blocks. The first four techniques by CNN and its advanced variants, such as
blocks comprise two layers of 2D convolutional layers ResNet which is a Residual Network. There are several
sandwiched between batch normalization layers. reasons for applying CNNs in lung nodule detection. In
Convolutional layers bring out most of the spatial this regard, these include possibly processing structured
features, including edges, shapes, and textures, to grid data, which are higher-order features of medical
differentiate nodules from other structures in the CT images based on the principle of hierarchical feature
scans. Batch normalization normalizes the output learning.CNNs apply various layers of convolutional
generated by each convolution layer; this stabilizes filters to detect the spatial hierarchies of features from
training and improves overall generalization performance. low-level patterns such as edges toward high-level
The fifth residual block is a bit special as it possesses a abstract features that are required to differentiate benign
single 2D convolutional layer with the aim to conserve and malignant nodules. Importantly, the features are learnt
computational complexity and to capture pertinent by CNNs automatically during the training phase without
features. The sixth and final residual block is larger and manually feature engineering, as typically done in the
contains four 2D convolutional layers for obtaining deeper earlier approaches.
features or deeper understanding of the features from
input windows. Following all these, max-pooling is Recent Works. Different recent works have demonstrated
applied and reduces the spatial dimensions of feature important improvements brought by CNNs in medical
maps by simply picking up the most important features imaging tasks, such as lung nodule classification. For
discarding less useful information. This simply means the instance, in the recent past, Zhang et al. (2019) reported
network was focused more on the areas that are prominent that hybrid models of CNNs can achieve good accuracy if
or most relevant within each window. The output of the features are handcrafted and automatically learned from
final block is passed through a dense layer which the CT scans for lung nodule classification. Like other
generates 256 probability values corresponding to each above-mentioned studies, Nageswaran et al. (2022)
32x32 window. It indicates the likelihood of a particular elaborated about works and discussions involving the
window containing a lung nodule. techniques on machine learning and image processing for
lung cancer prediction. This study presented the fact that
This system was tested on the LIDC-IDRI dataset, which CNN-based models are better than the traditional
is composed of CT images with annotated lung nodules. classification methods like SVM due to its capabilities to
The whole dataset was divided into the training group, identify finer nuances of difference in the data
validation group, and testing group, where 70% was used encountered in imaging.
for training purposes and the rest used for testing. The
Adam optimizer with a learning rate of 0.001 was used to
train the model since it adjusts the learning rate during
training for faster convergence. With 40 epochs of
training, the system achieved a top-5 accuracy rate of
95.24% on the test dataset. This implies that for 95.24%
of cases for each window, the correct candidate nodule Fig 2.3.1 ResNet-50 architecture
appeared within the first five predictions in the model.
The architecture of the proposed system presents a
One of the other significant changes of CNN architecture from raw images, thereby eliminating the need to
is the Residual Network, or ResNets, particularly ResNet- handcraft features.
50, as it overcomes the vanishing gradient problem that
tends to occur while training deep neural networks. With CNN Architectures Adapted in Lung Nodule Detection
the residual connections that skip layers and make easier a Different CNN architectures have been adapted in the
better gradient flow and feature propagation throughout medical image analysis domain. Used here are VGG-16
the network, now it is simply possible to train very deep and VGG-19, ResNet-50, DenseNet-121, MobileNet,
networks using ResNet-50. For these reasons, due to its Xception, NASNetMobile, and NASNetLarge. The VGG
capability to recognize very subtle and complex patterns model brought in deep architecture using 3×3
on scans that are also viewed as early stages of lung convolutional filters that permitted more abstract feature
cancer, this architecture performed fairly well in medical extraction at deeper layers while reducing parameters. The
image analysis. Kalavagunta et al. (2024) further advance ResNet architecture introduces residual blocks to avoid
their work with the introduction of CNNs and ResNet-50 the vanishing gradient problem, thus allowing much
architectures for lung cancer detection. The above deeper networks than previously used.
architectures were trained on annotated lung CT scan
datasets, and it was proven that the ResNet-50 version has
an increased accuracy of 92.54% compared to a regular
CNN at 85.67%, mainly caused by its deeper architecture
and ability to learn residual features. This work evidences
the ability of these architectures in elevating the accuracy
of the automatic diagnostic system, hence enabling faster,
less hassling manual diagnosis.
Fig 2.4.1 DenseNet-121 architecture
These are complemented by even more evidence, for
example that of Bruntha et al. (2022), who used transfer DenseNet is another advanced CNN architecture, which
learning to fine-tune deep learning models including improves on the feature reuse by providing dense
CNNs for the purposes of lung nodule classification. connections between every layer and every other layer of
Those indicated that CNNs can substantially improve every type, thus enhancing information flow and bringing
performance for very small sizes of medical datasets. in a fewer number of parameters. MobileNet and Xception
Further, Huang et al. (2019) provided insights into how both use depthwise separable convolutions, thereby
CNNs can be combined with other features to make more making the model less complex and also more suitable for
accurate malignancy predictions on lung nodules. mobile as well as embedded applications.

Based on the literature presented above, CNNs along with Dataset and Methodology
advanced variants such as ResNet-50 are gaining traction The Lung Image Database Consortium and Image
in the detection and classification of lung cancer from Database Resource Initiative (LIDC-IDRI) dataset, which
medical images. Since it increases both the precision and is used for containing the training and evaluation data set
the reliability of lung nodule detection and reduces the in this work, is a set of annotated CT scans of lung
time to reach a diagnosis, their use advances early nodules by four radiologists. For the nodules, the final
detection and treatment of lung cancer. Future research diagnosis was determined by a median rating assigned to
could further improve these models, perhaps by making the nodules. Pre-processing includes cropping raw CT
use of synthetic datasets, GANs, or other advanced scans of 3D lung nodule images and normalizing pixel
techniques that help alleviate some of the limitations due values. The deep learning models to be trained were fed
to data scarcity and generalization across various patient 3D patches from the CT images; and in order to treat the
populations. slices on the Z-axis as different channels, they were
compatible with 3D data without significantly increasing
the computation requirements. They used Adam optimizer
D. “Lung Nodule Diagnosis on 3D Computed along with a binary cross-entropy loss function over 500
Tomography Images Using Deep Convolutional epochs for training.
Neural Networks “
Performance Evaluation:
In the field of lung cancer diagnosis, CAD systems have The performances of the CNN models are measured in
increasingly become valuable tools for early diagnosis and terms of accuracy, sensitivity, specificity, and precision as
design of treatments in lung cancer diagnosis. Lung well as Area Under the Curve (AUC). DenseNet-121 and
cancer is an illness whose mortality rate ranking is quite Xception proved to be the best models, and DenseNet-121
high globally and whose late stage diagnosis accounts for yields the highest accuracy, specificity, and AUC values
most mortality; therefore CAD systems can try to assist at 87.77%, 92.38%, and 93.79% respectively, while the
the radiologist in detection of lung nodules as one of the highest sensitivity is achieved by Xception at 82.73%.
early symptoms of lung cancer. Traditionally, CAD Meanwhile, models such as VGG-16, VGG-19, and
systems rely on feature extraction and nodule ResNet-50 showed poor sensitivity; hence they failed to
segmentation processes for a machine learning model to classify between malignant nodules from benign nodules.
classify the nodule as either benign or malignant. Although Xception was less accurate than DenseNet-121
However, with deep learning, particularly CNNs, the in diagnostics, it was faster during training.
possibility has been established of end-to-end learning
 . IV. CONCLUSIONS
III. RESULTS / COMPARITIVE STUDY In conclusion, there is a comparison analysis of
architectures involving CNNs, namely VGG-16, ResNet-
50, and DenseNet-121a that has made strong differences
This section provides the comparison of five leading deep in detecting early-stage lung cancer from CT scans. VGG-
learning models: DenseNet121, ResNet, VGG16, 16 is very deep and is well for feature extraction but it
Improved VGG16, and ResNet-50 in terms of does not work well with complicated patterns because of
performance metrics like accuracy, advantages, and the absence of residual connections. ResNet-50 improves
disadvantages in the application to lung nodule detection this by using residual learning which assists in
using CT scans of the LIDC-IDRI dataset. It covers all overcoming the problem of a vanishing gradient and
aspects regarding feature extraction capabilities, enables the network to learn complex patterns better.
complexity in computations, and suitability for medical DenseNet-121 is an improvement in performance by
image classification tasks and the trade-offs between adding dense connections between layers that foster
achieving superior performance and resource utilization. feature reuse and improved gradient flow are very
TABLE 1: EVALUATION OF METHODS BASED ON THE effective for classification. DenseNet-121 showed promise
DATASET AND ARCHITETURE
in sustaining high information flow and feature diversity,
an important characteristic of change detection in medical
Dataset- Dis- images. Further future research would include
Algorithm Accuracy(%) Advantages
Used Advantages optimization of CNN architectures for early lung cancer
Excellent detection and testing them in more clinical settings.
feature reuse; Requires large
LIDC-
DenseNet- IDRI
handles computational REFERENCES
87.77% overfitting, resources due
121 (CT [1] Javed, Rabia, et al. "Deep learning for lungs cancer detection: a
deep network to deep
scans) review." Artificial Intelligence Review 57.8 (2024): 197.
with dense architecture
connections [2] Journal of Network Communications and Emerging Technologies
Residual (JNCET) Volume 11, Issue 2, February (2021) Improved VGG-16
learning Convolutional Neural Network Based Lung Cancer Classification
Complex and Identification on Computed Tomography
avoids
LIDC- architecture;
vanishing [3] Zhang, Qianqian, et al. "Lung nodule diagnosis on 3D computed
IDRI may have tomography images using deep convolutional neural networks."
ResNet 95.24% gradients;
(CT longer training Procedia Manufacturing 39 (2019): 363-370.
highly
scans) times with
effective at [4] Kumar, Ch VMS N. Pavan, et al. "LUNG CANCER DETECTION
deep layers
feature USING DEEP LEARNING." Journal of Nonlinear Analysis and
extraction Optimization 15.1 (2024)..
Requires more [5] Elnakib, Ahmed, Hanan M. Amer, and Fatma EZ Abou-Chadi.
Simple
layers for "Early lung cancer detection using deep learning optimization."
architecture;
LIDC- good (2020): 82-94.
good for
IDRI accuracy;
VGG-16 86% straightforwar [6] Asuntha, A., and Andy Srinivasan. "Deep learning for lung Cancer
(CT slower
d tasks and detection and classification." Multimedia Tools and Applications
scans) convergence
easy to 79.11 (2020): 7731-7762.
compared to
modify [7] Kumar, Vinod, et al. "Unified deep learning models for enhanced
newer models
More accurate lung cancer prediction with ResNet-50–101 and EfficientNet-B3
Higher using DICOM images." BMC Medical Imaging 24.1 (2024): 63.
than VGG16
LIDC- computational [8] Malik, Hassaan, and Tayyaba Anees. "BDCNet: Multi-
due to
Improved IDRI cost compared classification convolutional neural network model for classification
97% additional
VGG-16 (CT to original of COVID-19, pneumonia, and lung cancer from chest
layers and
scans) VGG16 due to radiographs." Multimedia Systems 28.3 (2022): 815-829.
better feature
more layers
extraction [9] Subramanian, R. Raja, et al. "Lung cancer prediction using deep
Handles very learning framework." International Journal of Control and
deep Automation 13.3 (2020): 154-160.
Requires high
LIDC- networks [10] Kalaivani, N., et al. "Deep learning based lung cancer detection and
computational
IDRI effectively; classification." IOP conference series: materials science and
ResNet-50 92.54% resources due
(CT good feature engineering. Vol. 994. No. 1. IOP Publishing, 2020.
to deeper
scans) extraction due
layers [11] Ashhar, Sarah Mohd, et al. "Comparison of deep learning
to residual
convolutional neural network (CNN) architectures for CT lung
connections
cancer classification." International Journal of Advanced
Technology and Engineering Exploration 8.74 (2021): 126.
[12] Gupta, Jaya, Sunil Pathak, and Gireesh Kumar. "Deep learning
(CNN) and transfer learning: a review." Journal of Physics:
Conference Series. Vol. 2273. No. 1. IOP Publishing, 2022.
[13] Hasanah, Syifa Auliyah, et al. "A deep learning review of resnet
architecture for lung disease Identification in CXR Image." Applied
Sciences 13.24 (2023): 13111.
[14] Pathan, Sameena, et al. "An optimized convolutional neural
network architecture for lung cancer detection." APL
bioengineering 8.2 (2024).
[15] Pandian, R., et al. "Detection and classification of lung cancer
using CNN and Google net." Measurement: Sensors 24 (2022):
100588.