A comparison of CNN-based visual explanation methods inmedical images

Artificial intelligence (AI) is becoming more vital in modern life, but as AI approaches get more complicated, they may become more difficult to explain and leading to lack of trust in their result specially in the medical field. In medical AI applications, eXplainable Artificial Intelligence (XAI) approaches are critical for ensuring that machine learning models are interpretable and trustworthy, which is critical for improving patient outcomes. In medical imaging, visualization methods such as heat maps are the best way to create interpretable and trustworthy representations of machine learning outputs, enabling the identification of regions of interest that could represent signs of sickness or abnormalities. This research examines and compares multiple Convolution Neural Network based (CNN-based) visual explanation approaches for creating accurate and trustworthy heat maps for medical image analysis, with the goal of improving the reliability and utility of medical AI. This thesis compares several XAI visualization approaches on the Kvasir endoscopy dataset with two classes and the HAM10000 skin cancer dataset with 7 classes. On the Kvasir dataset, an efficientnet-v2-s model is trained which achieved validation accuracy of 0.9, and several Activation Map (CAM) methods such as Gradient Activation Map (GradCAM), Gradient Activation Map ++ (GradCAMpp), techniques are used to create heatmaps. On the HAM10000 dataset, a VGG16 model is trained which achieved validation accuracy of 0.74 and using both the Layer-wise Relevance Propagation (LRP) and GradCAM approaches to create heatmaps. The Intersection Over Union (IOU) technique is used as an evaluation metric. This thesis tries to compare the visualization method and enhance the heatmaps and boost coverage of the regions of interest by combining the techniques. Three different methods, that is maximum voting, union, and intersection, are used to combine the heatmap images. The findings of the thesis indicate that the combination of GradCAM methods produces highly effective heatmaps that capture the most relevant features in medical images. This experiment is confirmed by the IOU method, with the combination of GradCAM++ and LayerCAM techniques generating the greatest results for HAM1000 dataset, for Kvasir dataset the method of GradCAM and XGradCAM have greatest results. Furthermore, the findings of thesis indicate that GradCAM techniques produce more accurate and trustworthy heatmaps than LRP, with important variables like the number of images, the disease type (in this thesis, data on polyps and skin cancer) and the model selection for training being crucial for achieving the best results.