Comparison of Deep Learning Explainable Approaches for Medical Image Analysis

This article explores the application of Explainable Artificial Intelligence (XAI) methods to improve the interpretability and trustworthiness of neural network predictions for medical application. Our research uses two well-established convolutional neural network (CNN) architectures: VGG-16 and ResNet50. Subsequently, we optimized the networks, which led to a final accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$99.24 {\%}$</tex> for VGG-16 and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$98.48{\%}$</tex> for ResNet50. In the Python programming environment, we used the far-reaching XAI toolkit, Quantus. Using techniques such as saliency maps and integrated gradients offered by Quantus, we were able to pinpoint the critical pixels influencing each prediction made by the networks. While VGG-16 achieved slightly higher accuracy, it also had longer training times compared to the faster ResNet50.