Slice, Subtract, Segment: Efficient 2D and Physiologically Informed 3D Brain Tumor Segmentation

Abstract

Background. Automated segmentation of glioma subregions on multiparametric MRI can be improved by changing either the network architecture or the information presented to the network. These two design levers are usually studied together, which makes their individual contributions difficult to see. We present an integrated analysis of two completed, independent studies, each of which isolates one lever. Methods. In the architecture-level study (Slice), SIBA-UNet - a 2D U-Net combining depthwise separable convolution, inverted-bottleneck blocks and attention-guided skip fusion - was trained on 128 x 128 multimodal slices from BraTS 2020 with a patient-level split, and a sigmoid decision threshold was selected on validation data. In the representation-level study (Subtract), CMSR-3D augmented the four standard MRI sequences with four voxel-wise cross-modal subtraction maps (T1-T1CE, T2-FLAIR, FLAIR-T1CE, T2-T1CE) and evaluated 4-, 6- and 8-channel inputs with an identical 3D U-Net on 1,251 BraTS-GLI 2023 cases (1,001/125/125 split). Results. SIBA-UNet reached an overall validation Dice of 0.8601 at a validation-selected threshold of 0.51, with reported subregion Dice of 0.8795 (necrotic core), 0.8501 (edema) and 0.9125 (enhancing tumor); the source ablation was only partially reported. With the architecture held fixed, the 8-channel CMSR-3D configuration gave the highest mean Dice among the evaluated inputs (0.8699 vs. 0.8534 for the 4-channel baseline) and the lowest mean HD95 (5.9859 vs. 8.2308), whereas either two-map subset alone changed mean Dice by at most 0.0037. Conclusions. Architecture design and input-representation design are separable levers for brain tumor segmentation. The controlled representation ablation indicates that the complete set of subtraction maps was more informative than either subset under a fixed 3D U-Net. Both studies were evaluated internally on their respective datasets, and their numerical results are not directly compared because the datasets, dimensionality and evaluation protocols differ.