An interpretable deep learning framework for medical diagnosis using spectrogram analysis

Abstract

Convolutional Neural Networks (CNNs) are widely utilized for their robust feature extraction capabilities, particularly in medical classification tasks. However, their opaque decision-making process presents challenges in clinical settings, where interpretability and trust are paramount. This study investigates the explainability of a custom CNN model developed for Covid-19 and non-Covid-19 classification using dry cough spectrograms, with a focus on interpreting filter-level representations and decision pathways. To improve model transparency, we apply a suite of explainable artificial intelligence (XAI) techniques, including feature visualizations, SmoothGrad, Grad-CAM, and LIME, which explain the relevance of spectro-temporal features in the classification process. Furthermore, we conduct a comparative analysis with a pre-trained MobileNetV2 model using Guided Grad-CAM and Integrated Gradients. The results indicate that while MobileNetV2 yields some degree of visual attribution, its explanations, particularly for Covid-19 predictions are diffuse and inconsistent, limiting their interpretability. In contrast, the custom CNN model exhibits more coherent and class-specific activation patterns, offering improved localization of diagnostically relevant features.