3D Glioma Segmentation Using SegFormer3d

Abstract

Abstract. Accurate segmentation of gliomas is essential for diagnosis and treatment planning but is challenged by their infiltrative nature and reliance on gadolinium-based contrast agents, which pose certain health risks. This study evaluates SegFormer3d, a transformer-based model, for efficient 3D glioma segmentation using the BraTS 2020 dataset, focusing on reducing dependency on T1-contrast (T1C) modality. With modifications including LayerNorm removal and a refined segmentation head, SegFormer3d achieved Dice scores of 0.8333 (Enhancing Tumor, ET), 0.8534 (Tumor Core, TC), and 0.9179 (Whole Tumor, WT) with all modalities, but ET performance dropped to 0.6178 without T1C. Weight Standardization and model scaling improved contrast-free results. 3D reconstructions using polygonal meshes and voxel-polygonal hybrid visualization aided surgical planning, though non-contrast ET accuracy requires enhancement. SegFormer3d offers a computationally efficient solution, but further advancements are needed for robust contrast-free segmentation.