Classification of pulmonary disorders using pulmonary network with explainable AI framework from chest multi-modal images.

Pulmonary disease (PD) is one of the most common diseases worldwide. To detect the PDs, several tests like RT-PCR, CBC, and antigen tests were used. But these are time-consuming processes. Therefore, chest radiographs are used to classify the PDs, but this may lead to false results due to similar lung lesions and nodules present in the chest radiograph. So, deep learning-based models were introduced, but these conventional models were less efficient. As a result, we propose the 30-layer novel PD classification network termed as PULMONET to classify the PDs from multi-modal chest images. In this, we have used 9917 chest multi-modal images to classify PDs into binary and ten classes. The model is trained with 50 epochs, 64 batch sizes, Adam optimizer and 5-fold cross-validation is used. In 5 folds, four folds were used for training and remaining one fold was used for validation to avoid overfitting and provide efficient data for training. The proposed model provides the highest accuracy of 99.17% for binary and 90.70% for ten-class classification scenarios respectively. Finally, we applied heat maps to initial and deeper layers to study the network behavior to acknowledge the decision-making of the PULMONET. The experimental findings show significant increases across performance parameters and good classification accuracy, demonstrating the resilience of the technique. In the proposed 30-layer PULMONET, the diagnostic accuracy of medical picture analysis is generally increased. Lastly, we investigated network behavior and the issue of misclassification by applying the t-SNE and activation maps to the network’s initial and deeper layers. These activations show clear clusters of the correctly classified and misclassified classes. Also, we used a custom segmentation map to view the best discriminative features that are more useful for the decision made by the model. The datasets, codes, and figures that are used in this work is available in: https://www.kaggle.com/datasets/sahebgoudkaraddi/pulmonary-diseases-datasetsct-cxr