Authors: J.Daisy Selva Rani, Anjitha Ajayan
Abstract: Deep learning applications in medical imaging have proved to be accurate; however, limited usage is due to the black-box issue of such systems. We propose an end-to-end XAI approach for medical images. Our work aims to cover the necessity of Explainable Artificial Intelligence solutions in medical imaging and show that they improve classification performance. To build such a solution, a combination of methods like SHAP, LIME, and Grad-CAM with CNN architecture are suggested. To test its performance, our experiments are conducted with X-ray and mammograms data, where we have observed a dependence of the interpretation quality on the architecture, specifically, the importance of spatial information preservation. In the result, we obtain improved classification quality, together with interpretability.
