Implementation of Vision Transformers for Lung Abnormality

Detection Using Low Dose CT Images

Alfred D., M. N. Deephak and D. Lakshmi

Department of Electronics and Communication Engineering, St Joseph’s College of Engineering, Chennai, Tamil Nadu,

India

Keywords: Vision Transformers, Lung Abnormalities, Low‑Dose CT, Self‑Attention, CNN Comparison, Diagnostic

Accuracy, Medical Imaging.

Abstract: Detection of lung abnormalities originating from a variety of infections, inflammation, and environmental

exposures in the patient needs high accuracy to help improve diagnostic efficiency by means of medical

imaging. Regular CNNs have the weakest long-range dependencies, being very weak with almost zero

receptive fields, which, therefore, induce many false positives and negatives. The future with ViTs is bright

because the self-attention mechanism can extract local as well as global features from images and perform far

better than regular CNNs. This work proposes a ViT-based model for the detection of lung abnormality in

low-dose CT images. Most of the existing systems are prone to high classification error rates because of their

poor quality towards understanding the context present in the images. The proposed ViT model bridges that

gap by using a pre-trained architecture, and patch-based processing is used to focus more on the essential

features of the image. We show how ViT's performance gets superseded by CNN for metrics through

comparison. The overall goal behind the project would be facilitating the early detection of lung abnormalities,

avoiding false results, and pushing the potential clinical uses with a dense, efficient solution for abnormality

detection of the lung.

1 INTRODUCTION

L. Devan., et al, 2013 Lung abnormalities range from

infections or inflammatory responses to any

environmental, genetic, or exposure kind of insult that

causes physical structures in the lungs to be damaged.

Any of them can fall into the groupings of structural,

obstructive, restrictive, or infectious and may call for

identification to enable proper strategy into treatment.

Early and appropriate detection of lung abnormalities,

especially malignant growths, in clinical settings is

pretty important for outcomes since it enables

interventions to occur in time.

X. Zhang., et al, 2023 In the last few years, some

of the techniques which have taken a lead in the field

of abnormality detection in lungs include medical

imaging techniques such as Computed Tomography

(CT). With CT imaging, cross-sectionals in lung

structures are resolved to great detail, and hence small

lesions or structural irregularities can be identified.

However, this is challenging to accomplish with high

diagnostic accuracy due to factors like image noise,

high anatomical variation, and overlapping features in

complex cases. Traditional approaches to deep

learning-deep, especially CNNs-have indeed achieved

promising performance for processing various

medical images. However, with their small receptive

fields, they are not able to capture long-range

dependencies that might arise in the medical

diagnostic context as false positives or false negatives.

R. Mahum and A. S. Al-Salman, et al, 2023

Recently, Vision Transformers (ViTs) have emerged

as a very powerful alternative to CNNs in computer

vision. Unlike CNNs that mainly depend on localized

filters, ViTs are based on a self-attention mechanism

that will allow capturing global contextual

information across an entire image. This characteristic

places ViTs as one that is uniquely advantageous for

medical image analysis, where localized and long-

range information is of critical importance in arriving

at accurate diagnoses. Spatial relations and context are

encoded regardless of the size or structure of the

images due to the division of an image into smaller

patches by ViTs.

266

D., A., Deephak, M. N. and Lakshmi, D.

Implementation of Vision Transformers for Lung Abnormality Detection Using Low Dose CT Images.

DOI: 10.5220/0013926400004919

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 5, pages

266-272

ISBN: 978-989-758-777-1

S. R. Vinta, et al, 2024 This work aims at the

application of ViTs in the detection of lung

abnormalities using low-dose CT images for patient

safety and proper diagnoses. Specifically, it aims to

analyze the performance of ViTs in identifying

regions of abnormality in CT scans and

histopathological images compared with traditional

CNN architectures. Through this comparison, we aim

to note how these benefits can potentially reduce false

positives and false negatives, thus improving

diagnostic accuracy and enhancing early detection.

This would revolutionize diagnostic practices because

clinicians would get better accuracy and reliability in

the identification of lung abnormalities. M. Irtaza,et al,

2024 The advanced architectures of ViTs are

exploited to demonstrate their potential contribution

toward faster, more accurate diagnostic decisions so

that improved patient care is attained.

The objective of this study is to develop a Vision

Transformer (ViT)-based model for detecting lung

abnormalities in low-dose CT images with high

accuracy. The model aims to improve diagnostic

efficiency by leveraging the self-attention mechanism

of ViTs to capture both local and global features,

reducing false positives and negatives. By utilizing a

pre-trained architecture and patch-based processing,

the model enhances feature extraction and contextual

understanding. The study also compares ViT with

CNN-based approaches to demonstrate its advantages.

Ultimately, this research seeks to facilitate early

detection, minimize diagnostic errors, and promote

clinical adoption of AI-driven lung abnormality

detection.

This work is organized with review of the

literature survey as Section II. Methodology described

in Section III, highlighting its functionality. Section

IV discusses the results and discussions. Lastly,

Section V concludes with the main suggestions and

findings.

2 LITERATURE SURVEY

The evolution of medical imaging and computational

methods, diagnosis in lung disease and cancer is even

more accurate. Several works have been undertaken

in the realm of enhancing segmentation of lung

parenchyma, tissue differentiation using impedance

spectroscopy, biomedical antenna design, and

lncRNA-disease association prediction. Progress in

disease classification that is rooted in deep learning

as well as multi-omics data also emphasizes early

detection alongside precision in treatment

approaches. This survey examines these new methods

and their addition to the diagnosis and prognosis of

lung disease.

C. Wu, et al., 2024 This work categorized lung

sounds by an enhanced model of Bi-ResNet with the

aim of diagnostic accuracy, incorporating both skip

and direct connections in feature combination. STFT

and wavelet transform information is used to improve

the training of the model. Based on this, the

introduced model performed greatly better than

standard Bi-ResNet on the ICBHI database,

particularly for noise and composite heart sound

patterns. T. Nguyen and F. Pernkopf, 2022 This work

talks about the classification of lung sounds in terms

of applying methods like co-tuning, stochastic

normalization with data augmentation on unbalanced

datasets and explored performance on both ICBHI

and a personal dataset, where such methods will turn

out to be remarkable improvements in classification

accuracy, particularly when adventitious lung sounds

and respiratory diseases are of interest.

T. Wanasinghe, et al, 2024 A CNN model is

engineered for enhancing lung sound classification

using techniques in feature extraction including Mel

spectrograms, MFCCs, and Chromatograms. Tested

on public datasets the model works exceptionally well

to classify 10 classes for the purpose of attaining

automated auscultation assistance in early lung

disease detection. N. Babu, et al, 2024 This work

addresses a data augmentation method that is

eigenvectors-based for enhancing the accuracy of

lung sound signal classification using an automated

diagnostic system. The authors employ features that

are spectrogram-derived and integrate them with

machine learning classifiers for effective use in low-

resource health environments and improved accuracy

with less noise.

K. Liu, et al, 2022 The method in this work is

founded on a better YOLO-based model towards lung

nodule detection. It introduces an enhanced YOLO-

v5 architecture through stochastic pooling with multi-

scale feature fusion and optimized loss function,

which achieves state-of-the-art performance in

efficiency and accuracy metrics. Thus, it assists the

radiologists to detect the nodules better, and assists in

overcoming the difficulties of the misdiagnosis

primarily because of its delicate appearance. H.

Alqahtani, et al, 2024 In this work, the proposed are

the developed convolutional autoencoders using the

enhanced Water Strider Algorithm to categorize lung

and colon cancers. In this, the process involved

includes noise elimination and MobileNetv2, which

assists in feature extraction to discriminate cancer

Implementation of Vision Transformers for Lung Abnormality Detection Using Low Dose CT Images

267

from histopathological images correctly. It enhances

the diagnostic precision in the direction of appropriate

treatment and prognosis of cancer patients. The

electronic nose equipment is suggested to detect

respiratory disease from sweat. Malikhah et al., 2022

A stacked model of DNN, combined with recently

developed techniques for feature extraction, offers a

cost-effective, rapid method to screen for infections

without processing the respiratory samples so as to

negatively affect risk decrease and enhance the

separability between classes of the data.

A. Tripathi, et al, 2024 It proposes two models as

a type of MobileNetV2. According to the fine-tuning

and L2 regularization, the models were fine-tuned to

improve the accuracy of early detection of lung

disease. The mobile models perform better than the

conventional architectures in the classification

performance on various datasets for the

pulmonologists in the process of offering early and

accurate diagnostic assistance and preventive care

services. M. Fontanellaz et al., 2024 The work

explored a CAD for lung fibrosis diagnosis; this CAD

was optimized with focus on segmentation precision

and radiomic feature examination. It carried out

comparison of 2D vs 3D representations for data and

treatment of segmentation tasks using MLP-Mixers in

addition to the baseline UNets and made a

comparison of diagnostic accuracy when competing

with skilled radiologists, and inferred the possibility

of CAD in medical imaging.

M. Obayya, et al 2023 Tuna Swarm Algorithm

with GhostNet is designed here for colon and lung-

cancer detection tasks. Proposed system uses Gabor

filtering for image preprocessing that maximizes

feature extraction pertaining to the required features,

enhancing high classification accuracy. High speed

computation and effective handling of massive

databases enables fast cost-effective diagnosis in

cancer. L. Zhu, et al, 2024 The proposed architecture

enhances the state-of-the-art U-Net with a shape

stream branch and multi-scale convolutional blocks

to better segment lung parenchyma, particularly in

small and blurry areas. The experiment results in the

present work indicated massive superiority scores on

Dice Similarity Coefficient over state-of-the-art

networks, thereby providing evidence of the

efficiency and effectiveness of the proposed method

towards challenging lung parenchyma areas

segmentation.

G. Company-Se et al., 2022 Electromagnetic

impedance spectroscopy, applied in a bronchoscopic

environment, is an alternate, non-invasive, non-

ionizing method for lung tissue discrimination over

CT and PET. The article analyzed both the 3- and the

4-electrode techniques and presented results in

support of the 3-electrode technique, thus confirming

its capability for discriminating bronchial and healthy

tissues and proposing complementary use in lung

pathology diagnosis. A. R. Chishti et al., 2023 The

work has listed the biomedical problems that have

size, efficiency, and biocompatibility as the ones to

be solved for the uses in disease detection such as in

cancer. Discussion of the function of antennas in

diagnostic imaging, biotelemetry, and biosensing is

presented along with the advancements that enable

the design of efficient biomedical devices for various

severe diseases.

J. Ha, 2024 A new matrix factorization

methodology in the prediction of the lncRNA-disease

association with regard to some of the limitations in

the existing computational models is introduced. In

this, the developed method integrates heterogeneous

biological information for enhanced accuracy over

performance in the identification of disease-related

biomarkers. M. Magdy Amin, et al, 2024 A deep

learning method for the classification of non-small

cell lung cancer is created based on multi-omics data

in the form of RNA and miRNA sequencing, in terms

of CNNs. This multimodal method obtained accuracy

for most classes much greater than some of the earlier

single-modality work and suggests that early

detection and proper classification of cancer subtypes

can be dramatically enhanced.

Limitations: Even with major breakthroughs,

various limitations remain for medical imaging and

computational diagnosis of lung disease and cancer.

Most deep learning models are plagued by limited

data, especially for rare conditions, causing

overfitting and decreased generalizability. High

computational expense and model training

complexity prevent real-world applicability,

particularly in resource-poor environments. Noisy

and artifact-ridden lung sound signals and imaging

data are affecting classification performance. Further,

single-modality data restricts in-depth analysis, while

clinical uptake is still hindered by interpretability.

Ethical implications of patient data privacy and bias

in AI-assisted diagnosis add to the complexities of

mass use in clinical environments.

3 METHODOLOGY

The methodology elucidates the systematic approach

adopted in this research work for lung abnormality

detection in low-dose computed tomography images

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

268

with the help of Vision Transformers. The approach

starts with exhaustive data acquisition, followed by

important preprocessing steps to enhance the quality

and variance of the image. Segmentation techniques

cut out the regions of interest from images thus

offering scope for further concentrated analysis. For

effective classification, the ViT architecture was

employed. Feature extraction and performance

analysis ensure that the final predictions are robust

and provide insights into the efficiency of the model

compared to traditional methods. Figure 1 shows the

ViT Architecture Diagram.

Figure 1: ViT Architecture Diagram.

3.1 Data Collection

The work begins by gathering an extensive dataset

composed of low-dose CT images and

histopathological slides. The images are acquired

from the med databases and institutions. They ensure

there is a variation in the lung abnormalities

particularly in the normal and cancerous regions.

Each image is labeled to allow for the supervised

learning as this model can distinguish between

healthy tissues and abnormal tissues. Figure 2 shows

the Lung CT Scans.

Figure 2: Lung CT Scans.

3.2 Preprocessing

Once they have the dataset, they preprocess it in

various steps to improve the quality and consistency

of images. Normalization is then done to standardize

the pixel intensity value, and data augmentation

techniques include rotation, flip, and scaling in order

to increase the diversity. Then they split the data into

training, validation, and test sets to ensure that the

model generalizes well towards unseen data. Figure 3

shows the Preprocessing Stage.

Figure 3: Preprocessing Stage.

3.3 Segmentation

To isolate the region of interest in the CT images,

segmentation is an important step. Advanced

techniques like thresholding and contour detection

facilitate the identification and definition of abnormal

regions. This step improves the model's ability to

spotlight the most critical features while washing

away noise from surrounding healthy tissues. Figure

4 shows the Segmentation Stage.

Figure 4: Segmentation Stage.

3.4 Classification

Classification is done through the architecture of

Vision Transformer (ViT) while processing images

that were segmented. These images were broken into

smaller patches, which were linearly embedded in

transformer layers. The self-attention in the ViT

would help the model capture both local and global

features, which would be beneficial for the model.

The output was classified into healthy, cancerous, or

non-cancerous categories with a multi-layer

perceptron head.

3.5 Feature Extraction

ViT utilizes feature extraction within its processing as

it extracts fitting features from the acquired images.

Implementation of Vision Transformers for Lung Abnormality Detection Using Low Dose CT Images

269

The self-attention layers focus on significant patterns

and structures regarding lung abnormalities. Such a

capability will be particularly beneficial in medical

imaging as it can identify minute details that would

be perceivable only when a disease has been

developed.

3.6 Analysis and Prediction

It evaluates the accuracy, precision, recall, and F1-

score of the model. It finds the mistakes of prediction

by ViT and points out areas to improve. Interesting

visualizations like confusion matrices and ROC

curves are presented to interpret results. This is an

elaborative analysis of whether the superiority of the

proposed model in comparison with the traditional

CNN methods actually exists for the task of lung

abnormality detection, which benefits the

applications in clinical practices.

Figure 5: Proposed Flow Diagram.

A Vision Transformer architecture comprises a

number of key components designed to properly

process images. It begins with an input image,

converted into fixed-size patches, then flattened, and

finally, linearly embedded. As a result, each patch is

augmented with the use of positional encoding to

preserve spatial information. The embeddings are

then fed to a stack of transformer encoder layers.

Here, the model captures dependencies across all

patches through self-attention mechanisms. This

enables the model to learn both local and global

features. The output from these layers is pooled and

then passed through a classification head for final

predictions. In other words, ViT excels at complex

image contexts and improves the diagnostic accuracy.

Figure 5 shows the Proposed Flow Diagram.

4 RESULT AND DISCUSSION

Early results of ViT for lung abnormality detection

task are highly promising since the performance of

this model has been successfully proven in analysis

of low-dose CT images. Training the suggested ViT

architecture achieves higher performance compared

to conventional CNN architectures dependent on

spatial interactions in particular, particularly when

accuracy and F1-score are at stake. A controlled

assessment was carried out on a test data set of

cancerous and non-cancerous areas, and compared to

the top-performing CNNs, it achieved an accuracy of

over 99%, whereas the latter obtained remarkable up

to 99%. This kind of superb performance would

validate that ViT indeed does offer superiority since

it is capable of leveraging all kinds of global

contextual information required to detect very small

abnormalities in medical images.

Figure 6: Output with accuracy.

Precision and recall values are equivalent

advantages for ViTs. One such model achieved a

precision of 99.4% and recall of 99.3% over the CNN,

which achieved a precision of 99.85% and recall of

99.80%. These high-level abilities to actually identify

real positives with minimum false negatives can

further enhance its performance in medical

environments where such an error could have quite

serious implications. The under the curve value of the

ViT model was also much greater than for the

aforementioned two models, which suggests a

potential for even more accuracy with good overall

performance in diagnosis based on differentiation

between healthy and unhealthy areas, as illustrated in

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

270

Table 1: Evaluation Metrics of Model Performance.

NORMAL MILD MODERATE SEVERE TUBERCULOSIS

ACCURACY 99.0891 99.0925 99.095 99.1014 99.0847

SENSITIVITY 99.03 99.03 99.03 99.03 99.03

SPECIFICITY 99.0891 99.0925 99.095 99.1014 99.0847

figure 6. Table 1 shows the Evaluation Metrics of

Model Performance.

Within the comprehensive explanation of the

model misclassifications, it was realized that a vast

majority of CNN false positives arose due to an

inadequate small receptive field unable to capture

very vital contextual cues. Contrary to this, self-

attention in the case of the ViT model made the model

better equipped to comprehend features anywhere in

the image and make a superior and more intelligent

classification. The above visualizations, i.e., the

confusion matrices and ROC curves, suggest that the

ViT is more robust compared to its counterpart where

differentiation between normal and abnormal

classifications is very evident

The third research was on the data augmentation

methods that were utilized in preprocessing to

improve the model's generalization capability from a

training set. The methods utilized, such as rotation,

scaling, and flipping, would also contribute to

robustness against overfitting and enhance the

performance of the model on new data, critical

characteristics particularly where there is limited

diversity, as appears to be the case with medical

imaging tasks.

The comparative study also confirmed the fact

that ViT's inference and training time was less than

that of conventional CNNs, indicating that ViT is

significantly more efficient than these CNNs for real-

world clinical use. This efficiency is critical in real-

time diagnostic procedures in healthcare, where

prompt results can be pivotal in deciding patient

outcomes. In short, the use of ViTs in detecting lung

abnormalities not only resulted in improved

performance metrics but also presents a strong

argument in using

ViTs for medical imaging applications. The

research thus lends merit to the assumption that ViTs

will transform diagnostic procedures, which will

enhance early diagnosis and improve patient care for

patients suffering from lung health issues. Future

research may be aimed at model integration into

clinical workflows and validation on a broad

spectrum of patient populations and imaging

modalities.

5 CONCLUSIONS

This work effectively utilized Vision Transformers

(ViTs) for lung abnormality detection in low-dose CT

scans, proving their capacity to learn both global and

local image features using self-attention mechanisms.

Our findings show that ViTs perform better than

conventional CNNs on various performance

measures such as accuracy, precision, recall, and F1-

score. The enhanced contextual perception of ViTs

resulted in a significant decrease in false positives and

false negatives, which points towards their capability

to improve diagnostic accuracy in clinical settings.

Major contributions of this work are strict data

preprocessing, data augmentation methods, and using

pre-trained ViT models to improve generalization to

intricate medical imaging tasks. By successfully

overcoming CNNs' weakness in capturing long-range

dependencies, ViTs offer a promising alternative for

more accurate and trustworthy lung abnormality

detection.

Even with these developments, some

misclassifications indicate avenues for enhancement.

Future studies will need to aim to optimize hybrid

models that integrate CNNs and ViTs to maximize

the strengths of both architectures. Further, increasing

the dataset to encompass heterogeneous imaging

modalities and patient populations will enhance

model robustness and clinical utility. As medical

imaging becomes increasingly AI-driven,

incorporating ViTs into diagnostic pipelines could

dramatically improve early detection capabilities,

ultimately leading to better patient outcomes and

more efficient clinical decision-making.

Implementation of Vision Transformers for Lung Abnormality Detection Using Low Dose CT Images

271

REFERENCES

A. R. Chishti et al., "Advances in Antenna-Based

Techniques for Detection and Monitoring of Critical

Chronic Diseases: A Comprehensive Review," in IEEE

Access, vol. 11, pp. 104463-104484, 2023, doi:

10.1109/ACCESS.2023.3316149.

A. Tripathi, T. Singh, R. R. Nair and P. Duraisamy,

"Improving Early Detection and Classification of Lung

Diseases with Innovative MobileNetV2 Framework,"

in IEEE Access, vol. 12, pp. 116202-116217, 2024, doi:

10.1109/ACCESS.2024.3440577.

C. Wu, N. Ye and J. Jiang, "Classification and Recognition

of Lung Sounds Based on Improved Bi-ResNet Model,"

in IEEE Access, vol. 12, pp. 73079-73094, 2024, doi:

10.1109/ACCESS.2024.3404657.

G. Company-Se et al., "Minimally Invasive Lung Tissue

Differentiation Using Electrical Impedance

Spectroscopy: A Comparison of the 3- and 4-Electrode

Methods," in IEEE Access, vol. 10, pp.

7354- 7367, 2022, doi: 10.1109/ACCESS.2021.31392

23.

H. Alqahtani, E. Alabdulkreem, F. A. Alotaibi, M. M.

Alnfiai, C. Singla and A. S. Salama, "Improved Water

Strider Algorithm with Convolutional Autoencoder for

Lung and Colon Cancer Detection on Histopathological

Images," in IEEE Access, vol. 12,

pp. 949- 956, 2024, doi: 10.1109/ACCESS.2023.3346

894.

J. Ha, "LncRNA Expression Profile-Based Matrix

Factorization for Predicting lncRNA- Disease

Association," in IEEE Access, vol. 12, pp. 70297-

70304, 2024, doi: 10.1109/ACCESS.2024.3401005.

K. Liu, "STBi-YOLO: A Real-Time Object Detection

Method for Lung Nodule Recognition," in IEEE

Access, vol. 10, pp. 75385-75394, 2022, doi:

10.1109/ACCESS.2022.3192034.

L. Devan, R. Santhosham and R. Hariharan, "Non-invasive

Method of Characterization of Fibrosis and Carcinoma

Using Low-Dose Lung CT Images," 2013 IEEE

International Conference on Systems, Man, and

Cybernetics, Manchester, UK, 2013, pp. 2168-2172.

L. Zhu, Y. Cai, J. Liao and F. Wu, "Lung Parenchyma

Segmentation Based on U-Net Fused with Shape

Stream," in IEEE Access, vol. 12, pp. 29238-29251,

2024, doi: 10.1109/ACCESS.2024.3365577.

M. Obayya, M. A. Arasi, N. Alruwais, R. Alsini, A.

Mohamed and Yaseen, "Biomedical Image Analysis for

Colon and Lung Cancer Detection Using Tuna Swarm

Algorithm with Deep Learning Model," in IEEE

Access, vol. 11, pp. 94705-94712, 2023, doi:

10.1109/ACCESS.2023.3309711.

M. Irtaza, A. Ali, M. Gulzar and A. Wali, "Multi-Label

Classification of Lung Diseases Using Deep Learning,"

in IEEE Access, vol. 12, pp. 124062-124080, 2024, doi:

10.1109/ACCESS.2024.3454537.

M. Fontanellaz et al., "Computer-Aided Diagnosis System

for Lung Fibrosis: From the Effect of Radiomic

Features and Multi-Layer-Perceptron Mixers to Pre-

Clinical Evaluation," in IEEE Access,

vol. 12, pp. 25642- 25656, 2024, doi:10.1109/ACCES

S.2024.3350430.

M. Magdy Amin, A. S. Ismail and M. E. Shaheen,

"Multimodal Non-Small Cell Lung Cancer

Classification Using Convolutional Neural Networks,"

in IEEE Access, vol. 12, pp. 134770-134778, 2024, doi:

10.1109/ACCESS.2024.3461878.

Malikhah et al., "Detection of Infectious Respiratory

Disease Through Sweat from Axillary Using an E-Nose

with Stacked Deep Neural Network," in IEEE Access,

vol. 10, pp. 51285-51298, 2022, doi:

10.1109/ACCESS.2022.3173736.

N. Babu, D. Pruthviraja and J. Mathew, "Enhancing Lung

Acoustic Signals Classification with Eigenvectors-

Based and Traditional Augmentation Methods," in

IEEE Access, vol. 12, pp. 87691-87700, 2024, doi:

10.1109/ACCESS.2024.3417183.

R. Mahum and A. S. Al-Salman, "Lung-RetinaNet: Lung

Cancer Detection Using a RetinaNet with Multi-Scale

Feature Fusion and Context Module," in IEEE Access,

vol. 11, pp. 53850-53861, 2023, doi:

10.1109/ACCESS.2023.3281259.

S. R. Vinta, B. Lakshmi, M. A. Safali and G. S. C. Kumar,

"Segmentation and Classification of Interstitial Lung

Diseases Based on Hybrid Deep Learning Network

Model," in IEEE Access, vol.12, pp.50444- 50458, 20

24, doi: 10.1109/ACCESS.2024.3383144.

T. Nguyen and F. Pernkopf, "Lung Sound Classification

Using Co-Tuning and Stochastic Normalization," in

IEEE Transactions on Biomedical Engineering, vol. 69,

no. 9, pp. 2872-2882, Sept. 2022, doi:

10.1109/TBME.2022.3156293.

T. Wanasinghe, S. Bandara, S. Madusanka, D. Meedeniya,

M. Bandara and I. D. L. T. Díez, "Lung Sound

Classification with Multi-Feature Integration Utilizing

Lightweight CNN Model," in IEEE Access, vol. 12, pp.

21262-21276, 2024, doi:

10.1109/ACCESS.2024.3361943.

X. Zhang, D. Maddipatla, B. B. Narakathu, B. J. Bazuin and

M. Z. Atashbar, "Intelligent Detection of Adventitious

Sounds Critical in Diagnosing Cardiovascular and

Cardiopulmonary Diseases," in IEEE Access, vol. 11,

pp. 100029-100041, 2023, doi:

10.1109/ACCESS.2023.3313605.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

272