A cascaded 3D-Unet with pre-activation residual SE attention for accurate brain tumor segmentation method in MRI imaging

Globally, brain tumors are among the most fatal malignancies in humans, and accurate segmentation of their intricate sub-regions remains a challenging and subjective task in clinical practice. This study proposes a cascaded 3D U-Net model with Pre-Activation Residual Squeeze and Excitation Attention to address these challenges. The model leverages ResNet34 and ResNext50 pre-trained encoders with ImageNet weights to enhance feature extraction, followed by a hierarchical segmentation strategy: coarse segmentation (whole tumor) and fine segmentation (tumor core and enhancing tumor). Two novel attention mechanisms: Cascaded Pre-Activation Residual Attention and Pre-Activation Residual SE Attention are introduced to focus the model on smaller tumor sub-regions. Experiments on the BraTS 2018 and 2019 datasets demonstrate the state-of-the-art performance. On BraTS 2018, the model achieves Dice scores of 91.9% (whole tumor), 95.2% (tumor core), and 98% (enhancing tumor), with corresponding IOU scores of 90.8%, 94.5%, and 97.5%. On BraTS 2019, it attains Dice scores of 90.2% (whole tumor), 95% (tumor core), and 95.6% (enhancing tumor), alongside IOU scores of 90%, 94.7%, and 99.6%. These results significantly outperform existing approaches, particularly in segmenting small and heterogeneous tumor sub-regions.