Enhancing Lung Cancer Detection Using AI‑Based Deep Learning

Framework

E. S. Vinothkumar

, J. Vinoj

, M. Nivaashini

, A. Ravi Kumar

Ram Ganesh G. H.

and R. Senthilkumar

Department of CSE, Vel Tech Rangarajan Dr.Sagunthala R&D Institute of Science and Technology, Avadi, Chennai,

Tamil Nadu, India

Department of CSE, Vignan’s Foundation for Science, Technology and Research (VFSTR), Guntur, Andhra Pradesh, India

Department of AI&DS, Sri Eshwar College of Engineering, Coimbatore, Tamil Nadu, India

Department of CSE, Sridevi Women's Engineering College, Hyderabad, Telangana, India

Department of IT, Kamaraj College of Engineering and Technology, Viruthunagar, Tamil Nadu, India

Department of CSE, Hindusthan Institute of Technology, Coimbatore, Tamil Nadu, India

Keywords: Lung Cancer Detection, Artificial Intelligence (AI), Predictive Analytics, Deep Learning, Deep Convolution

Neural Network (DCNN).

Abstract: Cancer of the lungs is considered the global supreme deadly disease that is life-threatening. However,

premature diagnosis and appropriate cure are crucial for tumbling the transience rate concomitant with this

ailment. Computed tomography scans have emerged as among the most prevalent imagining methods for lung

cancer exposure, particularly when coupled with deep learning models. In this study, propose a deep learning

framework grounded on a Deep Convolutional Neural Network for the timely exposure of lung cancer

expending CT scan images. Additionally, we have analysed the enactment of supplementary models, such as

Inception V3, Xception, and ResNet-50, in comparison to our proposed model. Our comparative analysis

considered various metrics, including accurateness, Area beneath the Curve, recall, and loss. After appraising

the models' presentation, the outcomes show that the DCNN-based approach outperforms the other models

and demonstrates promising potential compared to traditional methods. Specifically, the proposed DCNN

model attained an precision of 98.27%, an Area Under Curve (AUC) of 97.12%, a recall of 98.70%, and a

loss of 0.328.

1 INTRODUCTION

This type of cancer is the deadliest and furthermost

miserable on the planet after all the others. It is

extremely complex in its nature and highly

stimulating to diagnose, as its symptoms are

frequently revealed only during the later and final

stages. However, mortality rates from lung cancer can

definitely be concentrated significantly done

premature recognition and timely therapy. This

disease mainly initiates in lungs but sometimes

completes the entire course with a few minor

noticeable symptoms before it has metastasized to the

other parts of the body. There has been much on-

going research and developments on different

methods, and more of them, in the recent past, have

produced really promising results toward an effective

identification and diagnosis in the case of lung cancer.

One of the best imaging modalities employed here to

assist in diagnosing early medical conditions would

definitely turn out to be CT scan images; however, the

interpretation and detection of such scans from cancer

is a very complicated and challenging practice for

most medical practitioners. Early detection helps the

timely intervention and thus can prove to be highly

crucial for the outcome of patients. Continued

research and innovation in lung cancer screening and

diagnostic methods are very necessary to reduce the

significant impact of this condition on individuals.

2 MATERIAL AND METHODS

Publicly available data set comprising computed

tomography scan images was used in the study, which

went through a whole processing pipeline beginning

Vinothkumar, E. S., Vinoj, J., Nivaashini, M., Kumar, A. R., H., R. G. G. and Senthilkumar, R.

Enhancing Lung Cancer Detection Using AI-Based Deep Learning Framework.

DOI: 10.5220/0013887700004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 2, pages

643-647

ISBN: 978-989-758-777-1

643

with image resizing, removing artifacts and noise, as

well as advanced image segmentation techniques to

cut out regions of interest. The resulting DCNN

model was used to train, test and validate pre-

processed CT sets with other widely recognized deep

learning architectures such as Inception V3, Xception

and ResNet-50, according to the normal hold-out-

validation method. This performance comparison of

these deep learning models was further evaluated and

analysed to find the best architecture that could

potentially detect different types of cancer. The

DCNN model was custom trained architecture and

ResNet-50, Inception V3, as well as Xception,

compared with pre-trained transmission learning

models that exploited their learned representations

within lung cancer detection capabilities.

2.1 Dataset Collection

The study makes use of a public dataset comprising

computed tomography images which had undergone

an entire processing pipeline from resizing the

pictures to artifact and noise elimination as well as

advanced image segmentation techniques to cut out

the regions of interest. Such that the resulting model

of convolutional neural networks deep was used as

training, testing, and validation in all the pre-

processed CT sets using the mainstream pour-deep

learning architectures such as Inception V3, Xception

and ResNet-50, following the normal threshold

method of comparison. This performance comparison

of such deep learning models was also further

evaluated and analyzed to find the probably best

architecture that might be able to detect cancer types.

Compared to the above models, the DCNN model

was custom trained architecture and pre-trained

transfer learning models Inception V3, Xception and

ResNet-50 were used in their detection capabilities

within lung cancer. As such, the study involved the

consideration of a publicly available computed

tomography scan image database that underwent a

very rigorous pre-processing pipeline that involved

the following strides: the resizing of images, removal

of noise and artifacts, and requiring advanced image

segmentation techniques to isolate areas of interest.

The projected DCCN model was tested, verified, as

well as trained on pre-processed CT scan images by

the regular hold-out-validation technique along with

other known DL architectures. With respect to these

three models of deep learning, thorough evaluation

and analysis were done to find the most suitable

architecture for identifying the three included types of

lung cancers.

2.2 Dataset Pre-Processing

Feature extraction pipelines a too important pre-

processing step before going for a model analysis

through deep learning. It has different components

which together perform certain important activities

on input data around the needed modelling tasks. Fist,

the raw image data is read, capturing all original pixel

level information. It is here that the rest of this

pipeline begins. The next important pre-processing

activity is resizing the images into a common format.

This is important in ensuring that the deep

learning models will process the inputs. This is

followed by the removal of noise and artefacts from

the images. The model can, otherwise, be affected by

those unwanted characteristics such that, eliminating

them is necessary. Advanced techniques for image

segmentation are applied so that the regions of

interest can be isolated in the images and allow

analysis focusing only on the relevant sections. Then,

further, operations such as dilation and deterioration

are conducted for morphological processes to make

the segmented regions better. It enhances the quality

and integrity of input data for the deep learning model

to be used for classification or pinpointing location

tasks. But by processing data through such an

exhaustive feature extraction pipeline, most

associated deep learning models tend to be more

capable and reliable themselves. Thus, superior

results can be achieved with real-world applications.

Figure 1: Deep Convolutional Neural Network

Architecture.

2.3 Validation Process

These days, with increased availability for datasets of

larger images, looking for an appropriate validation

approach becomes crucial. One of the most used and

effective ways employed for this study is a hold-out

validation scheme, where we have an allocation of a

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

644

70:15:15 percent split in data for training, testing, and

validating, respectively. This allows an objective

evaluation of the enactment and generalizability of

DL models. Furthermore, DL models were qualified

for about batch size for 50 epochs is 13. This

configuration was selected keeping convergence and

computation efficiency in a balance. Besides, all the

models were implemented with a random seed of

1000 during the execution so as to give reproducible

results. This step is essential, as it mitigates the

inherent variability that can arise from the random

initialization of model parameters, which could

otherwise lead to inconsistent outputs across different

iterations. By carefully designing and executing the

validation strategy, along with the appropriate

training configurations and seed setting, the study

was able to provide reliable and reproducible insights

into the efficiency of the DL models for initial lung

cancer recognition using CT scan images.

3 PROPOSED DEEP CNN

ARCHITECTURE

The proposed deep CNN has a first convolutional

layer that takes in the 64x64 input image. This layer

has 16 filters and is expected to represent the most

basic features, thus producing 62x62 feature maps.

The convolutional layer served as the primary

fundamental component of the DCNN. Subsequent to

this, the output was conceded over a max pooling

layer, which reduced the longitudinal data size by

half, yielding 31x31 feature maps. Max pooling

chooses the supreme features from the covered

correspond with characteristics region. For further

processing, the output was then fed into a second

convolutional layer with 32 filters and 29x29 feature

atlases. This was trailed by another max pooling

layer, which halved the spatial data size to 14x14

distinctive maps. An additional set of convolutional

and pooling layers was incorporated in the third stage.

The pooling layer in this case contained of 5x5

distinctive maps, while the convolutional layer

utilized 64 filters with 10x10 characteristics atlases.

Lastly, the end results was flattened and passed

through a 260-dimensional dense layer that is

completely interconnected. This was then routed to

the softmax activation function layer, which is

usually employed for multi-class grouping tasks.

Excluding for the end layer, all layers utilized the

ReLU triggering utility without failure. The described

DCNN architecture is represented in Figure 1. The

model was accomplished, authenticated, and verified

using a rate of learning is 0.02, 50 epochs, and a group

size of 13. The Adam optimizer was employed to

compile the model, and a Classification cross-entropy

loss function was utilized, along with other evaluation

indicators like accuracy, recall, and AUC.

Deep CNN Algorithm

Step 1: Convolution layer: The Initial layer serves

as where the input images include are collected.

Step 2: RELU Layer: The picture passes over the

RELU layer, which introduces non-linearity.

Step 3: Pooling Layer: The image is then sent to the

pooling layer, where, if it is too big, the number of

parameters is decreased.

Fully Connected Layer: This layer extracts the sorts

of the pictures with as much as extraordinary

accurateness. It is a vital layer of CNN Split data: Sort

your data with labels into sets for testing, validation,

and training. The model is taught by the training set,

and the validation set monitors its progress, and the

testing set assesses its final performance.

Loss function and back propagation: The predicted

output is associated to the correct label using a loss

function. After that, the error extends reluctant over

and done with the network, correcting the weights and

biases of each layer to minimize future errors.

Step 4: Optimization: Repeat the forward pass and

back propagation for all training images, iteratively

refining the model's parameters using an optimization

algorithm like Adam or SGD.

4 RESULTS AND DISCUSSIONS

The performance results show of four DL

classification models - DCNN, Inception V3,

Xception and ResNet-50 - applied to the Cancer of

the lungs is CT Scan Image collections need

comprehensively summarized in Tables 1, with

comparative insights presented in Figure 2,3,4,5.

These tables provide a detailed breakdown of the

training, validation, and testing show metrics for each

of the respective deep learning models. The inclusion

of these comprehensive performance results allows

for a thorough evaluation and comparison of the

capabilities of the different DL models in the task of

CT scan images to find lung cancer.

Enhancing Lung Cancer Detection Using AI-Based Deep Learning Framework

645

Table 1: Training outcomes for various DL models for lung cancer detection.

Models

Accuracy of

Training

AUC of Training Recall of Training Loss of Training

DCNN 98.27% 97.12 98.70 0.29

ResNet -50 95.20% 97.26 97.20 0.045

Ince

tion V3 93.36% 95.3 93.2 1.96

Xce

tion 95.24% 95.6 93.2 1.45

The analysis of the methods employed by

the deep convolutional neural network, and other

models reveals that the DCNN model surpasses the

other deep learning approaches, as evidenced by the

comprehensive performance results presented in

Tables 1. The DCNN model was selected as the most

suitable option for the proposed framework aimed at

detecting lung cancer by CT scan images due to its

exceptional performance metrics.

Figure 2: Comparisons for training accuracy levels.

Figure 3: Comparisons for training area under curve.

Specifically, the DCNN model achieved

impressive testing results, including an accurateness

of 98.27%, an area below the curve of 98.21%, a

recall of 98.70%, and a loss of 0.328. These

outstanding performance metrics demonstrate the

DCNN model's effectiveness in accurately

identifying and classifying different kinds of lung

cancer as well as normal cases, from the CT scan data.

The DCNN model's superior performance, compared

to the other DL architectures, for instance Inception

V3, Xception and ResNet-50, makes it a promising

choice for the proposed framework targeting CT

Scans for the preliminary credentials of lung cancer.

Figure 4: Comparisons for training recall.

Figure 5: Training loss function.

5 CONCLUSIONS

Cancer of the lungs is a foremost reason of cancer-

connected humanity global. While it cannot be fully

prevented, early diagnosis and treatment can

significantly improve patient outcomes and survival.

This is a critical priority for healthcare suppliers and

researchers, as lung cancer often goes undetected

until advanced stages. Our research explored a deep

learning framework founded on DCNN for primary

recognition of lung cancer using CT scans. This

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

646

DCNN model outperformed other approaches like

ResNet50, Inception V3, and Xception, achieving an

accurateness of 98.05%, AUC of 97.32%, recall of

98.70%, and training loss function of 0.29. To further

enhance early lung cancer diagnosis, we might

incorporate additional datasets and explore other ML

and DL frameworks in the upcoming, aiming to

improve the overall performance and reliability of our

detection methods.

REFERENCES

“Breast cancer statistics of american cancer socity.

[online]. availabe:https://www.cancer.org/cancer.html,

” [Accessed: 13- November-2022].

Abhir Bhandary , G. Ananth Prabhu , V. Rajinikanth , K.

Palani Thanaraj , Suresh Chandra Satapathy , David E.

Robbins , Charles Shasky , YuDong Zhang , Joao

Manuel R.S. Tavares , N. Sri Madhava Raja ,

DeepLearning Framework to Detect Lung ˜

Abnormality – A study with Chest X-Ray and Lung CT

Scan Images, Pattern Recognition Letters (2019), doi:

https://doi.org/10.1016/j.patrec.2019.11.013

Ausawalaithong, W., Thirach, A., Marukatat, S., &

Wilaiprasitporn, T. (2018). Automatic Lung Cancer

Prediction from Chest X-ray Images Using the Deep

Learning Approach. 2018 11th Biomedical Engineerin

g International Conference (BMEiCON).

Da Silva, G.L.F.; da Silva Neto, O.P.; Silva, A.C.; de Paiva,

A.C.; Gattass, M. Lung nodules diagnosis based on

evolutionary convolutional neural network. Multimed.

Tools Appl. 2017, 76, 19039–19055.

Deepa, R., Karthick, R., Velusamy, J., & Senthilkumar, R.

(2025). Performance analysis of multiple-input

multiple output orthogonal frequency division multiple

xing system using arithmetic optimization algorithm.

Computer Standards & Interfaces, 92, 103934.

Hany, M. (2020, August 20). Chest CT-scan images

dataset. Kaggle. Retrieved November 13, 2022, from

https://www.kaggle.com/datasets/mohamedhanyyy/ch

est-ctscan-images

Kaur, S., Hooda, R., Mittal, A., Akashdeep, & Sofat, S.

(2017). Deep CNN-Based Method for Segmenting

Lung Fields in Digital Chest Radiographs. Advanced

Informatics for Computing Research, 185– 194.

doi:10.1007/978-981-10-5780-9_17.

M. Mamun, A. Farjana, M. Al Mamun and M. S.

Ahammed, "Lung cancer prediction model using

ensemble learning techniques and a systematic review

analysis," 2022 IEEE World AI IoT Congress (AIIoT),

2022, pp. 187 193, doi:10.1109/AIIoT54504.2022.981

7326.

M. Mamun, M. I. Mahmud, M. I. Hossain, A. M. Islam, M.

S. Ahammed, M. M. Uddin, "Vocal Feature Guided

Detection of Parkinson's Disease Using Machine

Learning Algorithms", 2022 IEEE 13th Annual

Ubiquitous Computing, Electronics & Mobile

Communication Conference (UEMCON), 2022,

M. Mamun, S. B. Shawkat, M. S. Ahammed, M. M. Uddin,

M. I. Mahmud, A. M. Islam, "Deep Learning Based

Model for Alzheimer's Disease Detection Using Brain

MRI Images", 2022 IEEE 13th Annual Ubiquitous

Computing, Electronics & Mobile Communication

Conference (UEMCON), 2022.

Makaju, Suren, P. W. C. Prasad, Abeer Alsadoon, A. K.

Singh, and A. Elchouemi. "Lung cancer detection using

CT scan images." Procedia Computer Science 125

(2018): 107-114.

Naqi, S.M.; Sharif, M.; Jaffar, A. Lung nodule detection

and classification based on geometric fit in parametric

form and deep learning. Neural Comput. Appl. 2018,

3456789.

Nibali, A.; He, Z.; Wollersheim, D. Pulmonary nodule

classification with deep residual networks. Int. J.

Comput. Assist. Radiol. Surg. 2017, 12, 1799–1808.

Senthilkumar Ramachandraarjunan, Venkatakrishnan

Perumalsamy & Balaji Narayanan 2022, ‘IoT based

artificial intelligence indoor air quality monitoring

system using enabled RNN algorithm techniques’, in

Journal of Intelligent & Fuzzy Systems, vol. 43, no. 3,

pp. 2853-2868

Senthilkumar R, Dr.P.Venkatakrishnan,Dr.N.Balaji, Intelli

gent based novel embedded system based IoT Enabled

air pollution monitoring system, ELSEVIER

Microprocessors and Microsystems Vol.77, June 2020

Shaffie, A.; Soliman, A.; Fraiwan, L.; Ghazal, M.; Taher,

F.; Dunlap, N.; Wang, B.; van Berkel, V.; Keynton, R.;

Elmaghraby, A.; et al. A Generalized Deep

Learning Based Diagnostic System for Early Diagnosi

s of Various Types of Pulmonary Nodules. Technol.

Cancer Res. Treat. 2018, 17

Xie, Y.; Zhang, J.; Xia, Y.; Fulham, M.; Zhang, Y. Fusing

texture, shape and deep model-learned information at

decision level for automated classification of lung

nodules on chest CT. Inf. Fusion 2018, 42, 102–110.

Zhang, T.; Zhao, J.; Luo, J.; Qiang, Y. Deep belief network

for lung nodules diagnosed in CT imaging. Int. J.

Perform. Eng. 2017, 13, 1358– 1370

Enhancing Lung Cancer Detection Using AI-Based Deep Learning Framework

647