Decoding brain tumor insights: Evaluating CAM variants with 3D U-Net for segmentation

Abstract

Brain tumor segmentation is critical for effective diagnosis and treatment planning. While, conventional manual segmentation techniques are seen inefficient and variable, highlighting the need for automated methods. This study enhances medical image analysis, particularly in brain tumor segmentation by improving the explainability and accuracy of deep learning models, which are essential for clinical trust. Using the 3D U-Net architecture with the BraTS 2020 dataset, the study achieved precise localization and detailed segmentation with the mean recall values of 0.8939 for Whole Tumor (WT), 0.7941 for Enhancing Tumor (ET), and 0.7846 for Tumor Core (TC). The Dice coefficients were 0.9065 for WT, 0.8180 for TC, and 0.7715 for ET. By integrating explainable AI techniques, such as Class Activation Mapping (CAM) and its variants (Grad-CAM, Grad-CAM++, and Score-CAM), the study ensures high segmentation accuracy and transparency. Grad-CAM, in this case, provided the most reliable and detailed visual explanations, significantly enhancing model interpretability for clinical applications. This approach not only enhances the accuracy of brain tumor segmentation but also builds clinical trust by making model decisions more transparent and understandable. Finally, the combination of 3D U-Net and XAI techniques supports more effective diagnosis, treatment planning, and patient care in brain tumor management.

Keywords

• brain tumor segmentation
• 3D U-Net
• explainable AI
• class activation mapping
• deep learning
• medical imaging