A Combination of Integrated Gradients and SRFAMap for Explaining Neural Networks Trained with High-Order Statistical Radiomic Features

This research tackles the problem of high-order statistical radiomic features’ visual explainability. While methods like Radiomic Features Activation Maps exist to solve this problem, they have important limitations. This includes the inability to produce a single explanation for all features and a lack of direct connection between classification results and generated explanations. This study contributes to the body of knowledge with a new explanatory saliency map generation approach for models trained with high-order statistical radiomic features. It extends the existing SRFAMap method using the Integrated Gradients method from Explainable AI. In detail, it exploits the integrated gradients of high-order statistical radiomic feature functions. Results with the tuberculosis classification dataset demonstrated better insertion and deletion correlation faithfulness metrics for saliency maps generated with the proposed approach than Radiomic Features Activation Maps.