Comparative Analysis of Explainable AI Techniques: LIME, SHAP, and Grad-CAM for Medical Imaging

Abstract

90% of the AI developed today remains a "black box", which makes it not suitable in applications with strong responsibilities to operate with, such as in healthcare. This work presents a comparison on three most common Explainable Artificial Intelligence (XAI) methods, namely Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM) by testing on a Convolutional Neural Network (CNN) . This model has been built to be a classifier of brain tumours based on MRI images. We fine-tuned a ResNet50 CNN pretrained on ImageNet to an MRI brain tumour classification task with the Brain Tumor MRI Dataset (Nickparvar, 2021)[10], an open brain tumour MRI dataset of 7023 T1 images with four labels: glioma, meningioma, pituitary tumour or no tumour. It was split into 5,712 train and 1,311 test sets. Three general properties were discussed and compared: interpretability, stability and computational costs. The results suggest that, in visual interpretation, Grad-CAM produces heatmaps that are more focused and conceptually much clearer, while, in the aspects of model explanation stability, feature importance justification and cost, SHAP is the overall best choice for integration with diagnostic applications and acceptance as a future clinical verification dataset; LIME, while being the most flexible and provider of model-agnostic explanations, lags behind in explanation consistency between different runs and, with a runtime of 9.1 s to explain a single image, is too costly comparing to 2s of Grad-CAM and 2.3s of SHAP. Therefore, it can be concluded that the best choice of method is not the same under every scenario.