BW-Dice: A Boundary-Weighted Dice Loss for Brain Tumor Sub-Region Segmentation

S. U. Ravi Kumar Chavali; P. V. Kumar

doi:10.36948/ijfmr.2026.v08i04.85453

BW-Dice: A Boundary-Weighted Dice Loss for Brain Tumor Sub-Region Segmentation

Author(s)	Mr. S. U. Ravi Kumar Chavali, P. V. Kumar
Country	India
Abstract	Brain tumor segmentation from multi-modal MRI is dominated by U-Net family architectures paired with Dice-based losses. In prior matched-condition work on BraTS 2020, it was established that the top five configurations of decoder architecture and standard loss function are statistically indistinguishable on aggregate mean Dice, and three distinct failure patterns were identified — edema over-segmentation, NCR/edema boundary confusion, and missed multi-focal disease. Motivated by the NCR/edema boundary failure mode, a Boundary-Weighted Dice (BW-Dice) loss is proposed, a generalization of soft multiclass Dice that applies per-pixel Gaussian weighting near ground-truth class boundaries, with a class-specific multiplier emphasizing the NCR-edema interface. BW-Dice reduces to standard multiclass Dice when the boundary emphasis strength λ = 0, providing a principled generalization rather than a competing alternative. BW-Dice is evaluated on the UNet++ architecture under identical training conditions to the prior baseline, spanning four hyperparameter configurations and three random seeds, with a controlled ablation removing the class-specific term. Across all configurations, BW-Dice significantly improves Tumor Core segmentation (Δ = +0.012, p < 10⁻³) while significantly degrading Enhancing Tumor segmentation (Δ = −0.018, p < 10⁻³), with aggregate mean Dice statistically indistinguishable from vanilla Dice (Δ = −0.001, p = 0.97). The controlled ablation demonstrates that this regional pattern persists without class-specific weighting, indicating it is a general property of boundary-emphasis losses rather than an artifact of the design choice. Class-specific NCR-edema emphasis significantly reduces the magnitude of ET degradation but does not significantly change TC improvement. Implications for future loss-function design in medical image segmentation with hierarchical class structure are discussed.
Keywords	Brain Tumor Segmentation, BraTS 2020, Boundary-Weighted Loss, Dice Loss, U-Net, Medical Image Analysis, Statistical Significance Testing, Class Imbalance
Field	Computer > Artificial Intelligence / Simulation / Virtual Reality
Published In	Volume 8, Issue 4, July-August 2026
Published On	2026-08-09
DOI	https://doi.org/10.36948/ijfmr.2026.v08i04.85453

View / Download PDF File

About IJFMR Fees & Payment Current Issue Publication Archive	Submit Research Paper Track Submission Status Publication Guidelines Publication Ethics Peer Review & Plagiarism	Join as a Reviewer Editors & Reviewers Reviewer Referral Program Get Reviewer Membership Certi.	Website/Journal Policies Usage Policy Content Policies Privacy Policy

Contact Us		+91-9687-828-838	editor@ijfmr.com

International Journal For Multidisciplinary Research

A Widely Indexed Open Access Peer Reviewed Multidisciplinary Bi-monthly Scholarly International Journal

BW-Dice: A Boundary-Weighted Dice Loss for Brain Tumor Sub-Region Segmentation

Share this