Year: 2026
Type: preprint
Abstract: Background:: Boundary-aware loss functions are intended to improve segmentation by penalizing spatial errors that may be insufficiently represented by region-based objectives. Whether they provide meaningful and reproducible benefits when incorporated into U-KAN for brain tumor segmentation remains unclear. Methods:: We compared five loss configurations in separate 2D and 3D U-KAN pipelines using multimodal MRI from the BraTS 2023 adult glioma dataset: a region-based baseline, fixed Boundary Loss, Staged Boundary Loss, a Generalized Surface Loss composite, and standalone Boundary Difference-over-Union loss. Each configuration was trained with three seeds in each pipeline, producing 30 runs. Performance was assessed using regional and macro Dice, the 95th-percentile Hausdorff distance (HD95), run-level consistency, and patient-level paired comparisons against the corresponding baseline. Results:: In 2D, Staged Boundary Loss achieved the highest mean macro Dice (0.8139 versus 0.8001) and reduced mean HD95 (22.55 versus 34.81~mm). Patient-level comparisons favored this configuration, but its overlap improvement was concentrated in whole-tumor segmentation and it exceeded the baseline in only one of three matched seed comparisons. In 3D, Staged Boundary Loss produced only small descriptive improvements in macro Dice (0.89794 versus 0.89698) and HD95 (3.645 versus 3.736~mm), with no significant favorable patient-level difference after multiplicity correction. The other boundary-aware configurations did not consistently outperform the baselines, and standalone Boundary Difference-over-Union exhibited substantial instability in one 2D run. Conclusions:: Boundary-aware training did not provide a consistent advantage over region-based objectives. Staged Boundary Loss was the most promising configuration, but its benefits were pipeline-dependent and not reliably reproduced across training runs.
Source: Research Square
Institution: Cadi Ayyad University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225416306