Lungs Disease Classification plays a crucial role in the detection, diagnosis, and treatment of various respiratory diseases. Given the complexity and variability of lung diseases, accurate classification methods are essential for distinguishing between different conditions based on medical imaging, clinical symptoms, and biomarkers.^1,2 Traditional diagnostic methods rely on radiologists interpreting chest X-rays or CT scans, which can be time-consuming and prone to human error. Machine learning and deep learning techniques, particularly convolutional neural networks (CNNs), have significantly improved the accuracy and efficiency of lung disease classification by automating image analysis and identifying patterns that may not be easily recognizable to the human eye.^1,2

The importance of lung disease classification extends beyond individual diagnosis, as it aids in public health monitoring and resource allocation. Timely and precise classification enables early intervention, reducing the severity of diseases and improving patient outcomes.^1,3 For instance, automated classification systems can assist in detecting infectious diseases like tuberculosis in high-risk populations, enabling quicker isolation and treatment to prevent further spread.^4,5

Classifying lung diseases into five categories— Tuberculosis, Coronavirus (COVID-19), Bacterial Pneumonia, Viral Pneumonia, and Normal Patients—is crucial for accurate diagnosis, effective treatment, and public health management. Each condition requires distinct medical interventions, and misclassification can lead to improper treatment, worsening patient outcomes.^3,6 Early identification of infectious diseases like COVID-19 and tuberculosis helps prevent transmission, while distinguishing between bacterial and viral pneumonia ensures appropriate antibiotic use, reducing the risk of antimicrobial resistance.^7,8 Automated classification systems powered by AI and deep learning can aid in rapid, cost-effective screening, especially in resource-limited areas.^9,10 This classification enhances clinical decision-making, outbreak control, and patient survival rates, ultimately improving healthcare efficiency and disease management.^11,12,13

This study performs an attention-based Convolutional Neural Network for the classification of lung diseases into 5 categories namely: Bacterial Pneumonia, Viral Pneumonia, Normal, Tuberculosis patients. Further, we also performed an Explainable local agnostic for explaining how these models made the prediction. Rest of the paper is as follows: section 2 has literature review, section 3 has methodology, section 4 has results and discussions, and section 5 has conclusions.

This section contains the information about the previous studies conducted on the detection of Lung Diseases Classification. We observed that many deep learning studies are used for the early detection of lung diseases such as Tuberculosis, COVID-19, Bacterial Pneumonia, Viral Pneumonia, and Normal Control patients. Researchers utilized Convolutional Neural Networks (CNN) to classify Tuberculosis, COVID-19, bacterial pneumonia, viral pneumonia, and normal patients. We classified the existing literature spanning from 2023 onwards into mainly 4 categories. They are:

Many of the latest studies performed on lung disease from Chest X ray images are focused on the Convolutional Neural Network without a greater focus on the interpretation and reasoning on why these models predicted a patient to be in a particular category. This is where our study is aiming for a region-focused neural network that can find out the lung disease category specifically into five categories namely: Bacterial Pneumonia, Viral Pneumonia, Corona Virus, and Normal patients. For this, we are using an attention-based neural network with interpretation layer using Local Interpretable Model Agnostic Explanations (LIME). Hence, the proposed model framework has the potential to explain a model explanation to the concerned stakeholder with the reason for such explanations.

We followed the methodology below in its sequential form for implementation. The first step involves image pre-processing, where the overall image is resized to a normalized size (620*620 pixel size). All the images used for the study is already in a Gray Scale format. Hence, we do not perform any further image conversions in the study. For example, the conversions from color to grayscale or grayscale or color conversions and all. In the second step, we performed the neural network architecture with a self attention mechanism. Then, in the third step the final results are evaluated for the execution purpose. Finally, we performed a LIME enabled visualization to understand it better which areas of the image region has made the significant contribution for the classification of lungs disease into the 5 categories.

We used three evaluation metrics for assessing the model. They are: 1. Sensitivity = TP/(TP+FN), 2. Specificity = TN/(TN+FP), 3. Balanced Classification Accuracy (BCA) = (Sensitivity + Specificity)/2, where TP is the True Positive ie Correctly identifying a class as such, FP is the False Positive ie, falsely identifying a positive class as negative, FN is the False Negative ie, falsely identifying a negative class as positive one. A better model tends to have higher value for all these metrics. AUC (Area Under the ROC Curve) measures how well a model ranks positives above negatives across all decision thresholds—it’s the probability a random positive gets a higher score than a random negative. Values range from 0.5 (random guessing) to 1.0 (perfect separation); higher is better.

We split the entire dataset into 3 categories namely: training, validation, and testing sets. After performing a 80%–20% split for the whole dataset, we created a training-validation set of 2780 and 695 testing data. We created a 80%, 10%, 10% split for training, validation, testing sets respectively (Table 1).

Table 2 illustrates the epoch wise training, validation, and testing set accuracy and error after the experimentation. It is observed that the result accuracy has significantly improved in each epoch resulting in a total overall training accuracy of 93% and testing accuracy of 88% after the epoch number 10. The self capture nature of SA-CNN made it possible to capture local patches in the image after each epoch. Hence, after increasing the number of epochs, the self attention mechanisms were able to increase the training accuracy to a greater extent. Further, we also observed that the training accuracy has reduced after epoch 10 as the self attention mechanism started to over adapt with the local patterns and starts to show the older patterns again and again.

Table 3 illustrates the classification of patients into Normal, Tuberculosis, Bacterial Pneumonia, Viral Pneumonia patients. As illustrated in Table 3, we can observe that the misclassifications are slightly larger while misclassifying Viral Pneumonia as Tuberculosis patients.

Table 4 has the class-wise Sensitivity, Specificity, F1 score, BCA for the developed model. An overall BCA of 84% is reported for the overall classification and the highest misclassification is reported for Bacterial Pneumonia which has to be rectified in the future studies.

Tables 5, 6, 7 indicates the ablation study results for the proposed model after the removal of attention, class weight, augmentation, data pre-processing layers of the model. Tables 5, 6, 7 has the results of BCA, Sensitivity, Specificity of the proposed architectures.

Ablation studies are critical in this work because they allow us to systematically evaluate the contribution of individual components to overall performance and interpretability, thereby strengthening the scientific rigor of the proposed model. Our experiments showed that removing the self-attention mechanism led to a marked reduction in Balanced Classification Accuracy, sensitivity, and specificity, confirming that the attention layer is not just an architectural enhancement but a vital element in capturing long-range spatial dependencies across chest radiographs. Unlike conventional convolutional filters that primarily focus on local features, the attention block enables the model to selectively emphasize clinically relevant lung regions and suppress irrelevant background patterns, which is particularly important for diseases like pneumonia and tuberculosis that present with subtle or diffuse manifestations. This demonstrates that the self-attention mechanism significantly improves both diagnostic accuracy and the clinical plausibility of the explanations, making it indispensable for robust and trustworthy lung disease classification.

Figures 5 and 6 presents the LIME interpretation for the lung disease classification after modeling with the SA-CNN. The results demonstrate that the crucial region responsible for the detection of Bacterial Pneumonia lies in the one lobe of the lungs and not with the respect to other. However. the SA-CNN models found both the lobes to be highly useful and found to be a very crucial factor for the detection of Tuberculosis patients. Thus, the proposed LIME framework can provide a detailed model agnostic local explanation with respect to every region of the lungs.

Table 8 has the illustration of results using various explainers. The insertion and deletion AUC results clearly highlight the importance of incorporating LIME-based explainability within the SA-CNN framework. Insertion AUC reflects how quickly classification confidence increases when pixels most emphasized by the explainer are revealed, while Deletion AUC captures how confidence drops when those pixels are removed. Lower Deletion AUC and higher Insertion AUC therefore indicate explanations that better align with the model’s true decision logic. The full SA-CNN combined with LIME achieves superior scores compared to its ablated counterparts (–Attention, –Augment, –ClassWeight), demonstrating that both the self-attention mechanism and robust training strategies are essential for producing faithful explanations. Moreover, when compared with alternative explainers such as Grad-CAM and Integrated Gradients, LIME maintains competitive performance, further validating its role as a reliable and interpretable method for understanding lung disease predictions.

The major limitation of the study lies in the improvement of the accuracy and reduction of misclassifications. Hence, we are planning to implement the model after a modified version of self attention CNN’s in the future. It is necessary to improve and self learn the regions of interest in distinguishing these regions of interest with an optimization based self attention mechanism. Moreover, we are also planning to increase the number of data for the study.

In future work, ethical considerations, potential biases, and clinical deployment pathways must be carefully addressed. Although this study used a publicly available de-identified dataset where no patient consent was required, any clinical extension of this work will demand strict adherence to data privacy regulations (e.g., HIPAA, GDPR) and prospective data collection under informed consent. Another critical aspect is mitigating potential biases, as models trained on single-source datasets risk inheriting demographic, disease prevalence, or imaging protocol imbalances, which can lead to reduced generalizability and unequal performance across subgroups. Expanding validation to multi-institutional datasets, applying class imbalance handling strategies, and incorporating fairness assessments will help overcome these limitations. For clinical deployment, the proposed self-attention CNN with explainability should be positioned as a decision-support tool to aid radiologists rather than replace them. Its adoption will require prospective validation on real-world cohorts, seamless integration into hospital workflows, compliance with regulatory frameworks such as FDA and CE certification, and continuous post-deployment monitoring to ensure reliability and trust.

This study developed a SA-CNN architecture for classifiying Tuberculosis, Bacterial Oneumonia, Viral Pneumonia, and Normal Chest patients using Chest X ray data of the individuals. We reported with a Balanced Classification Accuracy of 84% on the testing set. After modeling, we also developed a local agnostic LIME explainable framework for explaining how the SA-CNN architecture developed the model. We are planning to implement more modified versions of SA-CNN’s to find out more granular region specific information about the Chest X ray images as a future study.