Performance Evaluation of Deep Learning Models on Brain Tumor MRI Classification and Explainability

Abstract

Deep learning models often act as black boxes, making it difficult to understand their decision-making process. To understand how these models make decisions, this paper proposes a framework involving two phases. The first phase evaluates the performance of ten deep transfer learning models—ViT Transformer, EfficientNetB0, DenseNet121, Xception, GoogleNet, Inception V3, VGG16, VGG19, ResNet50, and AlexNet—for classifying brain tumors using MRI scans. The models are assessed based on metrics like accuracy, F1 score, recall, and precision, with EfficientNetB0 outperforming the other models with 98% accuracy and a balanced precision and recall, resulting in an F1 score of 98%. In the second phase, we use interpretability techniques such as Grad-CAM and Grad-CAM++, Integrated Gradient, and Saliency Mapping to investigate what these models learn within MRI images to make classification decisions. The results show that both Grad-Cam and Grad- Cam++ effectively identify the exact locations of tumor localization in the MRI images. This result enhances our understanding of the specific locations within the images where transfer learning models extract features to make classification decisions