Comparing Encoder Fusion Strategies in Hybrid Cnn–transformer Networks for 3d Brain Tumor Segmentation: a Controlled Ablation on Brats 2021
Year: 2026
Type: article
Abstract: Hybrid CNN–Transformer architectures have become the dominant paradigm for volumetric medical image segmentation, combining the local inductive bias of convolutions with the long-range context modeling of self-attention. However, the point at which the two encoder streams are fused and the mechanism used to fuse them are rarely studied in isolation: most published gains conflate fusion design with unrelated changes in depth, patch size, or training schedule. This paper presents a controlled ablation that isolates the fusion strategy as the sole independent variable. Using a shared CNN and Swin-Transformer backbone, we compare six fusion strategies—early, late, additive, concatenation, cross-attention, and a proposed gated multi-scale cross-attention fusion—on the BraTS 2021 dataset (1,251 training cases) under identical training conditions. The proposed gated multi-scale cross-attention fusion achieves the best overall performance, with mean Dice scores of 94.2%, 89.1%, and 82.5% for the whole tumor (WT), tumor core (TC), and enhancing tumor (ET) sub-regions, respectively, and the lowest HD95 of 4.68 mm, outperforming the strongest single-branch baseline by 2.4–4.1 Dice points while adding only 12 million parameters.
Institution: Fergana State Technical University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213884260