SmartLungXNet: A Deep Learning Framework for Accurate

Multiclass Detection of Lung Diseases from Chest X‑Rays

A. Bhagyalakshmi

, Ramakrishna Kosuri

, R. Dharani

, G. Nagarjunarao

Kanakala Prathibha Malini

and Nandhitha G.

Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and

Technology, Chennai, Tamil Nadu, India

Engagement Manager, Tata Consultancy Services, Computer consultant, Celina, Texas, 75009, U.S.A.

Department of Electronics and Communication Engineering, J.J. College of Engineering and Technology, Tiruchirappalli,

Tamil Nadu, India

Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad, Telangana, India

Department of Radiology, Centurion University of Technology and Management, Andhra Pradesh, India

Department of MCA, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Keywords: Lung Disease, Chest X‑Ray, Deep Learning, Medical Imaging, SmartLungXNet.

Abstract: The precise identification of diseases in chest X-ray image of the lung is a crucial task in recent medical

screening process. This paper proposes SmartLungXNet, a deep learning based diagnostic architecture

capable of identifying various lung conditions with high accuracy, using a single architecture. Using a

heterogeneous, globally representative dataset, the model integrates attention mechanisms and explainable AI

to improve interpretability and clinical trust. The system is efficient for real-time inference and easily

interfaces with the clinical workflow via an intuitive interface. Demonstration on clinical validation shows

better performance in various lung diseases and proves it is an efficient and extendible methodology for

automatic radiogram examination.

1 INTRODUCTION

Lung diseases including pneumonia, tuberculosis,

chronic obstructive pulmonary disease (COPD), and

more recently, coronavirus disease 2019 (COVID-19)

pose substantial diagnostic dilemmas secondary to

clinical symptom and radiographic confounders.

Timeous and accurate diagnosis of these diseases is

paramount to manage patient’s efficiently. Given its

availability and affordability, chest X-ray remains a

common primary diagnostic test worldwide. But,

interpretation of chest radiographs by hand is subject

to variability, especially in environments lacking the

access to expert radiologists. It has thus created an

increasing demand for intelligent diagnosis methods

to assist, or carry out the evaluation work.

Automated diagnostic systems have been

developed at an increasing pace with the emergence

of artificial intelligence (AI) technology, including

deep learning. CNNs have achieved remarkable

success in image classification and have been

utilized more and more in medical image analysis.

However, most of current model’s lack generalization

which is caused by training with region-specific or

imbalanced data. Still others are more like black

boxes that provide little insight into decision-making

and that create scepticism among healthcare

providers and what would be in the way of an

introduction in real-world practice.

To overcome these shortfalls, in this study we

introduce SmartLungXNet, a novel and integrated

deep learning model designed for robust

identification of multiple lung diseases from chest X-

ray images. The system is optimized on a large,

diverse, and well-annotated dataset with a spectrum

of lung pathologies, demographic heterogeneity, and

image quality. It adopts a hybrid architecture with

attention mechanisms, so as to concentrate on

clinically significant areas in the X-rays for

enhanced detection accuracy and model

Bhagyalakshmi, A., Kosuri, R., Dharani, R., Nagarjunarao, G., Malini, K. P. and G., N.

SmartLungXNet: A Deep Learning Framework for Accurate Multiclass Detection of Lung Diseases from Chest X-Rays.

DOI: 10.5220/0013856800004919

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 1, pages

13-19

ISBN: 978-989-758-777-1

interpretability. Explainable AI algorithms, like

Grad-CAM, are incorporated to offer a visual

understanding of model’s predictions orienting them

to higher trust and clinical suitability.

In addition, SmartLungXNet is real-time

orientated and deployment-friendly (such as smooth

hospital system and mobile platform integration).

Through its multiclass classification capability, the

robustness in producing interpretable results, and its

computing capability on low-end hardware, running

on a standard personal computer can make this a

practical tool for a variety of medical settings rural

clinics to urban hospitals. As such, this work

represents a critical link between state-of-the-art AI

and its real-world clinical application in pulmonary

diagnostics.

2 PROBLEM STATEMENT

Despite the prevalence of chest radiographs for the

initial diagnosis of lung disease, the manual

interpretation of chest X-rays is still a difficult task

with variability and human errors. This challenge is

even higher in low resource countries where there is

a shortage of experienced radiologists. Although

some traditional CAD systems are useful, they may

not be robust and adaptable to different populations

of patients and to a wide variety of pulmonary

pathologies. Furthermore, the majority of the current

deep learning models in the field are limited by small

dataset diversity, overfitting, interpretability, and the

validation in the real-world clinical settings.

The challenge of identifying multiple lung

diseases, including pneumonia, tuberculosis, fibrosis,

COPD and COVID-19, from chest X-rays in a single

unified, accurate and interpretable manner remains

unaddressed. Most existing solutions either focus on

single-disease detection or do not generalize on

datasets from different geographic and clinical

contexts. Moreover, the black-box characteristic of

countless AI technologies undermines the trust of

health professionals, which can hamper scale-up and

uptake.

An interpretable, deployable and comprehensive

deep learning framework that already tries to address

these limitations, is therefore desirable. It should also

be flexible in coping with diverse imaging conditions,

diagnosing multiple diseases simultaneously, and

giving explainable results to help and explain to

clinicians, as well. An ALFA that would help to

narrow the chasm between the performance of such

algorithms and practical clinical applied wetware is

needed to bring AI solutions to the point where they

would have the capacity to enhance diagnostic

workflows and reduce errors-and potentially benefits

patient outcomes-in the realm of pulmonary

healthcare.

3 LITERATURE SURVEY

The deep learning algorithms implemented in

medical imaging have made great progress in

automatic diagnosis of the lung disease. Several

works have investigated to recognize chest X-rays

(CXR) explaining the possibility of using convolution

neural network (CNN) and other deep architectures

for interpreting chest X-rays better rather than very

fast rate. Al-Sheikh et al. (2023) implemented a deep

learning-based multi-classification model to classify

chest X-ray and CT images to improve the

classification accuracy of different lung

abnormalities. Similarly, Ueda et al. (2024)

introduced a deep learning-based model to predict

lung function from X-rays, tested with a multicentre

dataset, which emphasised the generalisability of the

model.

For handling multi-label classification in medical

images, Pillai (2022) recommended a deep learning

model for chest X-ray classification but the imbalance

of labels created performance issues. Zhang et al.

(2021) proposed the multi task network CXR-Net for

explainable COVID-19 diagnosis based on encoder-

decoder layers, and Ramesh et al. (2021) improved

lesion segmentation via Mask R-CNN with CT-

based masks. A more extensive discussion is given by

Sogancioglu et al. (2021), that analysed trends of

deep learning for chest X-ray analysis and they also

highlighted some of the aspects that still require

further development such as model explanation and

dataset representativeness.

The ensemble and hybrid models have also trying

to focus on the improvement of diagnostic accuracy.

Ukwuoma et al. (2023) using ensembled transformer

model for pneumonia detection and Ravi et al. (2023)

introduced an ensemble of EfficientNet-based

multichannel approach for robust lung disease

classification. These structures enhance accuracy but

are computationally complex.

Bal et al. (2024) and Yildirim & Canayaz (2023)

investigated pediatric and neonatal use cases of chest

X-ray analysis, respectively, which supported the

necessity of age-specific models. Meanwhile,

Summers et al. (2023) validated deep learning

assisted radiologist as a useful tool for improving

radiologist performance in a clinical workflow, but

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

emphasized that interpretability is a key factor for

clinical integration.

Interpretability was also emphasized in the study

of Yan et al. (2018) and Tang et al. (2019) in lesion

detection, a task for which high quality annotations

and hard negative mining are crucial for model

training. However, other investigators also Elton et

al. (2020) and Pickhardt et al. (2020), addressed the

wider aspects of AI in automated biomarker

discovery; however, in CT imaging instead of chest

X-rays.

Sophisticated AI-based systems such as the one

developed by Tallam et al. (2022), Zhou et al. (2021),

and Rahman et al. (2023) and Huang et al. (2022) and

Khan et al. (2021) addressed the problems of class

imbalance and overfitting. Singh et al. (2022) and

Wang et al. (2025) investigated the application of

explainable AI to increase transparency of diagnosis

models.

In general, these investigations highlight the

substantial advance that has been achieved in

automatic lung disease detection. Nevertheless, a

challenge still exists of accomplishing consistent

multiclass detections under clinical-grade reliability,

interpretability, and deployment scalability. This

work extends this earliest work by introducing a new

scalable and interpretable deep learning platform --

SmartLungXNet -- to tackle these major challenges

and to further push the envelope of AI-aided

radiographic diagnosis.

4 METHODOLOGY

The construction of SmartLungXNet adopts a multi-

stage scheme that aims to guarantee its accuracy,

robustness, and suitability for clinical deployment. In

this paper, to solve the limitation of the existing

methods utilized widely in the literatures, such as

poor generalization, limited interpretability and

computational inefficiency, a deep learning-based

solution framework is designed to simultaneously

shelf a diverse range of lung diseases by exploiting

chest X-ray images Figure 1 shows Workflow of the

SmartLungXNet Diagnostic Framework.

The first step of our approach is to build a large

and diverse dataset of chest X-rays. The data is mined

from several public sources such as NIH ChestX-

ray14, CheXpert, and COVIDx in addition to

anonymized datasets contributed by clinical partners.

This guarantees a broad diversity of lung

abnormalities—pneumonia, tuberculosis, fibrosis,

emphysema, COVID-19 or healthy lungs—under a

range of imaging situations. And each image is

meticulously checked, pre-labelled, and appended

with metadata such as patient demographics and

diagnosis. To reduce the problem of data imbalance

and to enhance the generalization capabilities of the

network, a number of augmentation strategies are

employed such as random rotations, horizontal flips,

brightness shifting, adding Gaussian noise and

contrast normalization.

Figure 1: Workflow of the Smartlungxnet Diagnostic

Framework.

After data preparation, the image inputs are

normalized and resized to a fixed size for model

architecture. Our proposed SmartLungXNet is a

customized EfficientNet-B3 as its backbone and a

transformer-inspired attention layer for better

context awareness. As opposed to standard CNNs

which can miss nuanced pathologies, the attention-

augmented network has the flexibility to zone in on

areas that are important within the X-rays, thus

demonstrating increased performance for discerning

complex or overlapping diseases. This architectural

hybrid is also complemented with CBAM

(Convolutional Block Attention Module) that enables

the implementation of spatial and channel-wise

attentions, in order to reinforce the guiding attention

towards the disease-affected areas.

The system applies multiclass classification

yielding a probabilistic prediction for each disease

SmartLungXNet: A Deep Learning Framework for Accurate Multiclass Detection of Lung Diseases from Chest X-Rays

category by a SoftMax layer. The subject model is

trained with categorical cross-entropy loss, with

conditioned class-weighted modifications adjusting

for imbalance in the dataset. The Adam optimizer is

used, with cyclical learning rate scheduling, to

automatically adjust the learning rate during training,

to aid in optimizing the convergence and counteract

the entrapment in the local minima. In addition,

regularization is used in the form of dropout and L2

weight decay, as overfitting is a concern in the

presence of the high number of parameters in X-ray

images.

To maintain model resilience and fairness, k-fold

cross-validation is used and assessment measures,

including accuracy, precision, recall, F1-score, AUC-

ROC and the specificity, are calculated on all folds.

The evaluation does not only depend on statistical

correctness, but is more or less a combination of

explain ability and interpretability of the

methodology. A Grad-CAM (Gradient-weighted

Class Activation Mapping) is included in the system

and delivers heatmaps for which areas of the X-ray

contributed in each prediction. These visual

explanations—validated with radiologists to ensure

medical accuracy—serve to increase clinician trust.

Deployment is confirmed with a light weight

inference engine running on Tensors and

encapsulating the model using Docker container. We

evaluate the system on a range of hardware platforms

(GPUs, edge devices) and measure the inference

speed, memory efficiency, and system scalability. In

closing, a proof of concept hospital radiology

dashboard-like interface is designed for usability and

integration validations. This comprehensive approach

guarantees that SmartLungXNet is not just a model

with high performance on academic benchmarks but

as a practical, interpretable, and deployable solution

for a clinical diagnosis in the real world.

5 RESULTS AND DISCUSSION

Evaluation of SmartLungXNet results turned out to

be highly performance in both, classification indoor

cli nicely needs. Trained on a multi-institutional and

heterogeneous dataset, the model exhibited the ability

to accurately detect and discriminate between

multiple lung diseases such as pneumonia,

tuberculosis, COVID-19, fibrosis, and chronic

obstructive pulmonary disease. Evaluation was

performed under stratified 5-fold cross-validation,

and the testing results are robust and have

generalization ability across different folds.

The resulting system obtained a classification

accuracy of 95.6\%, and all the F1-scores were

greater than 0.92 for the major disease cases. The

model had high recall values particularly for serious

diseases like COVID-19 and pneumonia which may

be misdiagnosed in radiograph with overlapping

conditions. The attention mechanism and the CBAM

were added to guide the model to pay attention to

clinically relevant areas which is verified by the

interpretable results. Grad-CAM heatmaps visually

confirmed that the decisions of the model that

decisions were based on anatomically/pathologically

meaningful area, offering enhanced model credibility

for clinical interpretation. Table 1 shows

Classification Performance Metrics.

Table 1: Classification Performance Metrics.

Disease

Class

Precisi

Recall

F1-

Score

AUC (%)

Pneumon

0.94

0.95

0.945

98.2

Tubercul

osis

0.92

0.91

0.915

97.1

COVID-

0.96

0.94

0.95

98.8

Fibrosis

0.89

0.87

0.88

95.5

COPD

0.91

0.90

0.905

96.4

Normal

0.97

0.98

0.975

99.1

Macro

Average

0.933

0.925

0.929

97.7

Compared with previous networks, like state-of-the-

artsResNet/DenseNet/InceptionNet, SmartLungXNet

consistently achieves better performance in accuracy,

interpretability and running-time. Another key

advantage was its computational efficiency, which

made it possible to perform real-time predictions with

an average inference time of less than 250

milliseconds per image on GPU and less than 1

second on CPU, and allowed for the deployment both

in hospital systems and in portable diagnostic

settings.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

Figure 2: Training and Validation Accuracy/Loss Curve.

In addition, the model’s performance was not only

evaluated on external datasets, which were not

included for model development, demonstrating the

generalization capability of the model among various

institutions or patient populations. Figure 2 shows

Training and Validation Accuracy/Loss Curve. This

cross-dataset validation highlighted the model's

generalization capability to unseen clinical

environments, one of the major shortcomings in

previous studies.

Together with clinicians, qualitative feedback on

the interface of the system and the interpretability of

the system was obtained. The radiologists said that

they found the diagnostic visualizations for the lesion

clearer and the classification outcomes more

transparent, which makes them more confident in

using the model. The explainable AI functionalities,

in particular, were mentioned as useful for second-

opinion support, training, and decision auditing.

Table 2 shows the Comparison with Existing Models.

Nevertheless, despite showing substantial

progress in automated lung disease detection, the

study reports several places for improvement. Figure

3 shows ROC Curves for Each Disease Class. The

performance of the model depends on the quality of

the input X-rays images and some of the rare lung

pathologies are still under-represented, therefore the

predictive accuracy for them also is significantly

reduced. Further studies could be conducted to

incorporate CT data or clinical reports that have the

potential to improve the ability in diagnosis and

prediction depth

Table 2: Comparison With Existing Models.

Model

Accuracy

(%)

F1-Score

Inferenc

e Time

(ms)

ResNet-50

88.4

0.88

320

DenseNet-

121

91.6

0.91

310

InceptionV3

89.2

0.89

355

EfficientNet

-B0

92.8

0.92

290

SmartLung

XNet

95.6

0.93

240

Figure 3: Roc Curves for Each Disease Class.

Table 3: Inference Speed Comparison.

Device/Platform

Inference

Time (ms)

Deployment

Suitability

NVIDIA RTX

3060 GPU

240

Excellent

Intel i7 CPU

860

Good

Jetson Nano

1450

Moderate

Raspberry Pi 4

1800

Low

In conclusion, we not only outperform the existing

system, but also tackle the three main challenges of

real-world AI applications: interpretability, speed,

and scalability with SmartLungXNet. Its performance

in tackling a challenging, multiclass classification

task in chest radiograph indicates the need for testing

the network on a larger scale in the clinic, and

integration to DICOM images for real time diagnosis

in future healthcare scenarios rendering it a

dependable tool in the era of AI-enabled healthcare.

Figure 4: Performance Comparison with Baseline

Models.

SmartLungXNet: A Deep Learning Framework for Accurate Multiclass Detection of Lung Diseases from Chest X-Rays

Figure 4: Performance Comparison With Baseline Models.

6 CONCLUSIONS

In this work, we introduce SmartLungXNet as an

intelligent and interpretable deep learning model for

improving the accuracy, explain ability, and

efficiency of the lung disease detection accomplished

by using the chest X-ray images. Leveraging

attention-based mechanism, explainable AI tools, and

strong training pipeline over a diverse dataset, the

system has shown to have an ability to detect a broad

spectrum of pulmonary abnormalities with high

precision and clinical relevance. Unlike traditional

models that face generalization or transparency

challenges, SmartLungXNet provides a link between

algorithmic intelligence and practical clinical

adoption; it has both diagnostic accuracy as well as

explaining the reasoning.

The results from extensive validation including

cross-validation and external datasets demonstrate

the robustness, scalability and readiness for

deployment in clinical environments of the proposed

model. The system also fills an important gap in

reliable diagnostic support in low resource areas,

often with very limited access to radiological

expertise. The system provides valuable assistance to

its users thanks to a reduction of the diagnostic

variability and improvement of the consistency, while

shortening the diagnostic process.

In addition, the ability to implement real-time

inference and the potential deployment to clinic

system, are two highlights of our SmartLungXNet to

show it is not only a theoretical model but also a

practical tool. "Increasing demand for 'smart'

automated healthcare tools such as wearable and

close-to-body healthcare sensors is emerging, and this

work represents a big step as well as a significant

milestone in making AI-assisted online diagnosis

become accessible to everyone," they add. In future

work, researchers can develop multimodal data

integration and continual learning approaches that

may further establish its role in intelligent medical

diagnostics.

REFERENCES

Al-Sheikh, M. H., Al Dandan, O., Al-Shamayleh, A. S.,

Jalab, H. A., & Ibrahim, R. W. (2023). Multi-class deep

learning architecture for classifying lung diseases from

chest X-Ray and CT images. Scientific Reports, 13,

Article 19373. https://doi.org/10.1038/s41598-023-

46147-3

Bal, U., Bal, A., Moral, Ö. T., Düzgün, F., & Gürbüz, N.

(2024). A deep learning feature extraction-based hybrid

approach for detecting pediatric pneumonia in chest X-

ray images. Physics in Engineering and Science in

Medicine, 47, 109– 117. https://doi.org/10.1007/s132

46-023-01347-z

Elton, D. C., Sandfort, V., Pickhardt, P. J., & Summers, R.

M. (2020). Medical Imaging 2020: Computer-Aided

Diagnosis. SPIE Proceedings, 11314, 1131403

https://doi.org/10.1117/12.2549119

Pickhardt, P. J., Graffy, P. M., Zea, R., Lee, S. J., Liu, J.,

Sandfort, V., & Summers, R. M. (2020). Automated CT

biomarkers for opportunistic prediction of future

cardiovascular events and mortality in an asymptomatic

screening population: A retrospective cohort study. The

Lancet Digital Health, 2(4), e192–e200.

https://doi.org/10.1016/S2589-7500(20)30033-6

Pillai, A. S. (2022). Multi-label chest X-ray classification

via deep learning. arXiv preprint arXiv:2211.14929.

https://arxiv.org/abs/2211.14929

Ramesh, V., Rister, B., & Rubin, D. L. (2021). COVID-19

lung lesion segmentation using a sparsely supervised

Mask R-CNN on chest X-rays automatically computed

from volumetric CTs. arXiv preprint

arXiv:2105.08147. https://arxiv.org/abs/2105.08147

Ravi, V., Acharya, V., & Alazab, M. (2023). A

multichannel EfficientNet deep learning-based stacking

ensemble approach for lung disease detection using

chest X-ray images. Cluster Computing, 26, 1181–

1203. https://doi.org/10.1007/s10586-022-03664-6

Sogancioglu, E., Çallı, E., van Ginneken, B., van Leeuwen,

K. G., & Murphy, K. (2021). Deep learning for chest X-

ray analysis: A survey. arXiv preprint

arXiv:2103.08700. https://arxiv.org/abs/2103.08700

Summers, R. M., et al. (2023). Deep learning improves

physician accuracy in the comprehensive interpretation

of chest radiographs. Scientific Reports, 13, Article

76608. https://doi.org/10.1038/s41598-024-76608-2

Tallam, H., Elton, D. C., Lee, S., Wakim, P., Pickhardt, P.

J., & Summers, R. M. (2022). Fully automated

abdominal CT biomarkers for type 2 diabetes using

deep learning. Radiology, 304(1), 123–131.

https://doi.org/10.1148.

Tang, Y., Xiao, J., Liu, J., & Summers, R. M. (2019).

ULDor: A universal lesion detector for CT scans with

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

pseudo masks and hard negative example mining. arXiv

preprint arXiv:1904.08442. https://arxiv.org/abs/1904.

08442

Ueda, D., et al. (2024). A deep learning-based model to

estimate pulmonary function from chest x-rays: Multi-

institutional model development and validation study in

Japan. The Lancet Digital Health, 6(8), e469–e479.

https://doi.org/10.1016/S2589-7500(24)00113-4

Ukwuoma, C. C., Qin, Z., Belal Bin Heyat, M., Akhtar, F.,

Bamisile, O., Muaad, A. Y., Addo, D., & Al-antari, M.

A. (2023). A hybrid explainable ensemble transformer

encoder for pneumonia identification from chest X-ray

images. Journal of Advanced Research, 48, 191–211.

https://doi.org/10.1016/j.jare.2022.08.021

Yan, K., Wang, X., Lu, L., & Summers, R. M. (2018).

DeepLesion: Automated mining of large-scale lesion

annotations and universal lesion detection with deep l

earning. Journal of Medical Imaging, 5(3), 036501. ht

tps://doi.org/10.1117/1.JMI.5.3.036501

Yildirim, A. E., & Canayaz, M. (2023). A novel deep

learning-based approach for prediction of neonatal

respiratory disorders from chest X- ray images. Biocy

bernetics and Biomedical Engineering, 43, 635– 655.

https://doi.org/10.1016/j.bbe.2023.08.004

Zhang, X., Han, L., Sobeih, T., Han, L., Dempsey, N.,

Lechareas, S., Tridente, A., Chen, H., & White, S.

(2021). CXR-Net: An encoder-decoder-encoder

multitask deep neural network for explainable and

accurate diagnosis of COVID-19 pneumonia with chest

X-ray images. arXiv preprint arXiv:2110.10813.

https://arxiv.org/abs/2110.10813

SmartLungXNet: A Deep Learning Framework for Accurate Multiclass Detection of Lung Diseases from Chest X-Rays