Dual-domain explainability-driven data augmentation for enhanced COVID-19 detection in chest X-rays

Houda El Mohamadi, Mohammed El Hassouni, Rachid Jennane

Abstract: The classification and detection of COVID-19 from chest X-ray (CXR) images using deep learning, particularly transfer learning models, have shown promising results. However, the limited amount of available data often leads to performance degradation and an increased risk of overfitting. While data augmentation techniques provide an effective solution to these challenges, they frequently suffer from training instability and fail to capture essential visual features, resulting in poor generalization and detection capabilities. In this paper, we propose a novel data augmentation strategy that leverages explainable deep learning for COVID-19 diagnosis using CXR images. Our approach utilizes explainability in both the spatial and frequency domains, employing the Dual-Tree Complex Wavelet Transform (DT-CWT) to extract the most critical regions in the images, ensuring that the model focuses on essential areas for more accurate detection. To augment our original dataset, we generate saliency maps using explainable methods that highlight the most important features, serving as an importance filter. These saliency maps guide the data augmentation process, resulting in the generation of new images enriched with relevant information for precise diagnosis. Our approach improves training, leading to better accuracy (2-22%), reduced false positives, minimized overfitting, and smoother learning curves, enhancing the model’s reliability.

External IDs:dblp:journals/mta/MohamadiHJ25