ENHANCED BRAIN TUMOR SEGMENTATION AND CLASSIFICATION VIA OPTIMIZED CONVOLUTIONAL NEURAL NETWORKS

Abstract

Accurate segmentation of brain tumors in MRI images is critical for diagnosis and treatment planning, yet it remains challenging due to the brain’s complex anatomy and tumor heterogeneity. Deep learning models, particularly U-Net, have shown strong potential due to their ability to process high-resolution images and delineate tumor regions effectively. This study uses the BraTS 2018 dataset and metrics including Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Absolute Volume Difference to assess the performance of four models: 3D U-Net, PSPNet, DeepLabV3+, and ResNet50.The 3D U-Net achieved the highest performance (DSC: 0.90, HD: 10.69mm, AVD: 11.15%), followed closely by PSPNet. DeepLabV3+ and ResNet50 yielded lower accuracy (DSC: 0.85 and 0.83, respectively). While transfer learning considerably enhanced DeepLabV3+ and ResNet50, data augmentation greatly increased the performance of 3D U-Net and PSPNet. For brain tumour segmentation using multi-contrast MRI, the 3D U-Net model, improved by augmentation and transfer learning, is often advised. These findings underscore the value of selecting and optimizing deep learning architectures for medical image analysis.