An Improved Residual 3D U-Net for Brain Tumor Segmentation

Accurate brain tumor segmentation in MRI is essential for clinical diagnosis and treatment planning. This study proposes an improved deep learning model based on a Residual 3D U-Net architecture for glioma segmentation. The model integrates ResNet blocks to enhance feature extraction and employs skip connections to retain spatial detail across layers. The architecture features an encoder based on ResNet for extracting hierarchical features, followed by residual blocks that enhance feature representation while preserving spatial information. Skip connections are utilized during the decoding phase to handle structural scale variations within MRI images. Transposed convolutional layers in the decoder up-sample feature maps, retaining details and incorporating contextual information from earlier layers. Evaluated on the BraTS2020 dataset, the model achieved a Dice Score of 89.67% for the whole tumor, 90.59% IoU, 99.32% accuracy, and 85.00% F1-score, outperforming many state-of-the-art approaches. The proposed architecture holds promise for edge-deployable and real-time brain tumor analysis in clinical settings.