EXPLAINABLE ARTIFICIAL INTELLIGENT METHOD GRAD-CAM IN MEDICAL IMAGES PROCESSING

The reliability of modern deep learning models in the medical domain is frequently questioned due to their black-box nature. Post-hoc explainability techniques from the field of explainable artificial intelligence (XAI) offer a means to improve transparency and assess the reliability of predictions produced by convolutional neural networks. The research aims to investigate how XAI methods, specifically Gradient-weighted Class Activation Mapping (Grad-CAM), can provide reliable explanations for medical image classification. For this purpose, MRI images of brain were used to train a convolutional neural network to categorize the four stages of dementia in Alzheimer's disease. To make each prediction transparent, the areas of the brain which the trained network used to make the categorization on were highlighted using Grad-CAM. The resulting relevance maps, heatmaps, were evaluated using two approaches: spatial comparison with anatomically defined brain regions associated with Alzheimer’s disease using atlas overlay, and quantitative faithfulness assessment using a deletion-based metric, where highly influential regions identified by Grad-CAM were progressively removed and the impact on classification confidence was measured.