Deep Learning Fusion: CNN-GRU Model for Tuberculosis Detection Enhanced with LIME Interpretability

Substantial technological advances and extensive accessibility of medical data such as CT scans, X-rays and MRI images, have significantly driven the growth of automated detection systems for disease diagnosis. These Computer-Aided Diagnosis (CAD) systems help healthcare providers make accurate diagnoses, particularly for infectious diseases that pose potentially fatal risks to patients. Extensive research endeavors have been undertaken, employing a combination of deep learning techniques alongside chest X-rays (CXRs) and CT scan images, to address a spectrum of pulmonary disorders, including tuberculosis (TB). This research study investigates the fusion of a traditional Convolutional Neural Network (CNN) with a Recurrent Neural Network (RNN), specifically the Gated Recurrent Unit (GRU), to improve the categorisation of CXRs into tuberculosis and normal categories. The GRU’s representation learning capability is leveraged as it learns from the attributes identified by the CNN and passed on to the model. This capability plays a crucial role in our study, as it enables the model to capture sequential patterns and dependencies within the extracted features. The hypothesis is that this ability will optimize the model’s detection efficiency in identifying TB from CXRs, thereby enhancing the classification accuracy. To address the challenge of an unbalanced dataset, various data augmentation methods have been applied, followed by image enhancement techniques. Additionally, the LIME package has been utilized to obtain interpretable explanations for individual predictions by highlighting key features used by the model in classification. The proposed model achieved an impressive accuracy, precision, recall, F1 score and AUC of $\mathbf{9 8. 0 0 \%, 9 9. 4 3 \%, \mathbf { 9 6. 6 7 \% }, \mathbf { 9 8. 0 2 \% } \ 2 0}$ and 1.0 in correctly predicting target classes. The results establish that the proposed system performs well owing to the combination of CNN with GRU, which allows for the successful extraction of complex patterns from CXRs along with the capture of temporal dependencies within the data.