GCSER-UNet achieves 94% dice score in brain tumor segmentation

Researchers Sourjya Mukherjee, Ananya Bhattacharjee, and R. Murugan have unveiled GCSER-UNet (Global Context-aware Squeeze and Excite Residual UNet), a deep learning architecture designed for automated brain tumor segmentation from multimodal MRI scans. The model addresses the limitations of manual segmentation—high cost, labor, and error risk—by integrating a fusion of spatial and channel-wise attention mechanisms. This allows GCSER-UNet to capture intricate spatial dependencies and contextual information more effectively than previous UNet variants. The architecture leverages squeeze-and-excite blocks and residual connections to enhance feature representation, enabling precise delineation of tumor boundaries even in complex MRI slices.

Evaluated on benchmark datasets, GCSER-UNet delivers remarkable results. On the TCGA LGG dataset, it achieves a dice score of 94%, surpassing the prior state-of-the-art of 91.8%. On the BraTS 2020 dataset, an ensemble approach yields dice scores of 95% for Whole Tumor (WT), 92% for Tumor Core (TC), and 90% for Enhancing Tumor (ET)—compared to previous bests of 94%, 93%, and 88% respectively. These improvements, particularly in the challenging enhancing tumor region, highlight the model's robustness. The authors note that GCSER-UNet can assist neurologists in more accurate brain cancer management and treatment planning, potentially improving patient outcomes through earlier and more reliable diagnosis.