3D-sViT-UNET: An Effective Framework for Enhanced Brain Glioma Segmentation
Sadam Hussain, Usman Sadiq, Labiba Gillani Fahad, Ahmad Raza Shahid · 2024
Gliomas are the most common and severe type of brain tumor that causes higher mortality in adults worldwide. Accurate segmentation of brain tumors is crucial for treatment planning and patient survival. Given the heterogenetic nature of brain tumors, like indistinct boundaries and variability in size, shape, and location, this is a real challenge in medical imaging. In this research, we use an effective dual feature extraction strategy that integrates features from the U-Net encoder with the vision transformer encoder. UN et is effective at extracting local features, while Transformer excels at global feature extraction due to its higher receptive field. The final segmentation map is generated using the decoder part of the UN et. The proposed model generates satisfactory results. On BRATS 2020, dice similarity coefficients of 89.57%, 79.97%, and 82.44% for whole tumor, enhanced tumor, and tumor core, segmentation, respectively. Ablation results on BRATS 2021: Dice similarity coefficients of 92.5%, 84.75%, and 89.2% for the same tasks. With an average IOU/Jaccard coefficient of 80.80%, an average dice of 87.08%, a Hausdorff distance (95th percentile) of 4.68 mm, and an average surface distance of 0.90 mm. The model complexity is lower than other state-of-the-art methods which makes this more practical in this domain.