A method for brain tumor detection and localization by means of deep learning and reliable prediction explainability

Brain tumors, which are abnormal growths in the brain or central spinal canal, represent a serious global health threat. In 2023, they were responsible for 94,390 cases and 18,990 deaths in the US. Early detection with Magnetic Resonance Imaging is critical for effective treatment and higher survival rates. In this paper, we propose a method for the classification and localization of brain tumors in magnetic resonance imaging. Our method is based on the application of convolutional neural networks and for this task, we introduce and design a new network called the CBP_Net. The experimental analysis includes the Standard_CNN, a network designed by the authors, and other state-of-the-art networks, i.e., MobileNet, DenseNet, InceptionV3, ResNet50 and EfficientNetB0. This study emphasizes both quantitative and qualitative evaluations. The results demonstrate that the networks achieve optimal quantitative metrics: from 0.993 to 1.0 in accuracy. Qualitatively, we focus on visual explanations using Class Activation Mapping algorithms, specifically Grad-CAM and Score-CAM, to generate heatmaps highlighting relevant regions in the images. The Structural Similarity Index Measure is then employed to quantitatively measure the similarity between the heatmaps generated by the two CAM algorithms. This approach improves the reliability and trustworthiness of AI-driven brain tumor detection, helping clinicians diagnose.