SFX-SwinUNETR: selective fusion and prototype-based explanation for multimodal brain tumor segmentation
Year: 2026
Type: article
Abstract: Transformer-based architectures have become standard for volumetric brain tumor segmentation; however, three difficulties persist: substantial computational cost, degraded performance on small subregions such as the enhancing tumor, and predictions that cannot be traced to interpretable evidence. This study presents SFX-SwinUNETR, a Swin-UNETR-based architecture incorporating three components. A selective complementary feature fusion (SCFF) block recalibrates the four MRI channels through squeeze-excitation attention combined with an identity path. A prototype-based self-explainable (ProtoSX) decoder derives the segmentation logits from per-voxel L2 similarities to learned tissue-appearance prototypes, so attribution is intrinsic to the forward computation; we claim this as an architectural property and not as a demonstration of explanation quality. A contour-weighted compound loss (CWCL) augments soft-Dice with a Sobel-based boundary penalty targeting irregular tumor margins. Experiments used a 125-case subset of the BraTS 2021 training cohort with a held-out partition of five volumes; the official BraTS validation and test partitions were not accessed, and no leaderboard ranking is claimed. Under a cold-start regime without external pre-training, the Swin-UNETR baseline reached a mean Dice of 0.651 (tumor core 0.616, enhancing tumor 0.506), and the recalibrated configuration reached 0.902 (tumor core 0.919, enhancing tumor 0.888). These are single-run values at small n, so no dispersion estimate accompanies them, and they are an existence demonstration rather than a cohort-level estimate. The reported figures were obtained with a preliminary implementation in which the prototype decoder reduced to an identity mapping and the objective was soft-Dice without the contour term, so they do not evidence a contribution from either component (Section 4.5). No component-wise attribution is claimed, because the configurations that would isolate each module have not been trained. Differences are reported descriptively; distance-based surface metrics and paired per-patient hypothesis testing are implemented and await the five-fold evaluation. The architecture adds 5,192 parameters (+0.008%) relative to the baseline. The explanations were not assessed for fidelity, stability, or clinical usefulness. The architecture specified in Section 3 is released and unit tested, and its evaluation is outstanding.
Source: Frontiers in Neuroscience
Authors: Hardik Kaushik(Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing), Ganesh Khekare(Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing), Khushi Bhatnagar(Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing)
Institution: Vellore Institute of Technology University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7236722249
