Plagiarism is checked by the leading plagiarism checker
Volume 4 Issue 4
July-August 2026
| Author(s) | Mr. S.U. Ravi Kumar Chavali, Dr. V. Kumar P. |
|---|---|
| Country | India |
| Abstract | A companion matched-condition study found that a boundary-weighted Dice loss (BW-Dice) improves tumor-core (TC) segmentation but degrades enhancing-tumor (ET) segmentation — a trade-off with no net gain — and hypothesized a scale-mismatch mechanism in which the boundary-emphasis kernel is wider than the small ET region. We test the natural prediction of that hypothesis: shielding ET from the boundary emphasis should preserve the TC gain while removing the ET loss. Under matched conditions on BraTS 2020, we train a region-aware BW-Dice variant that applies boundary weighting to the necrotic-core and edema channels but trains ET with standard Dice, evaluated across three random seeds, a kernel-width sweep, and two additional decoders. Region-aware weighting does not improve on BW-Dice: in numerically stable settings it is equal to or worse than BW-Dice on both TC and ET, and a single-seed apparent gain proves non-reproducible. The core–enhancing trade-off is therefore not a fixable artifact of kernel scale. We also document seed- and architecture-dependent mixed-precision instability of weighted-Dice losses. Reporting this negative result narrows the space of promising loss-design interventions. |
| Keywords | Brain tumor segmentation; BraTS 2020; boundary-weighted loss; negative result; reproducibility |
| Discipline | Computer > AI / ML |
| Published In | Volume 4, Issue 4, July-August 2026 |
| Published On | 2026-08-15 |
| DOI | https://doi.org/10.62127/aijmr.2026.v04i04.1463 |

E-ISSN 2584-0487All research papers published on this website are licensed under Creative Commons Attribution-ShareAlike 4.0 International License, and all rights belong to their respective authors/researchers.