Interpretable chest X-ray localization using principal component-based feature selection in deep learning

Accurate identification and localization of diseases in chest X-ray (CXR) images are crucial for early diagnosis and timely medical intervention. Traditional localization techniques like Class Activation Mapping (CAM), depend on Global Average Pooling (GAP) layers, restricting their flexibility, while gradient-based methods like Grad-CAM involve computational overhead and limited interpretability. To address these limitations, this study introduces a novel Principal Component Analysis (PCA)-based localization method that eliminates reliance on GAP layers and gradient computations. Utilizing publicly available Kaggle datasets, namely the COVID-19 Radiography Dataset and Tuberculosis (TB) Chest X-ray Database. The proposed approach employs PCA to compress high-dimensional convolutional feature maps extracted from the pretrained VGG16 model into a lower-dimensional, spatially meaningful representation. This enables rapid, interpretable heatmap generation highlighting precise abnormal regions. Experimental results demonstrate that the proposed method achieved an average training loss of $0.0835\pm 0.1830$ and validation loss of $0.1385\pm 0.0741$ across 5-fold cross-validation. In addition, it achieved an impressive accuracy of 97.5%, sensitivity of 98.2%, specificity of 99.4%, a Dice Similarity Coefficient (DSC) of 97.5%, and an Intersection-over-Union (IoU) of 95.1%. Compared to CAM, and Grad-CAM, PCA-based localization significantly reduces inference time, enhances interpretability, and provides robust multi-class localization performance suitable for clinical deployment.