Abstract
Background: Scribble-supervised LGE-CMR segmentation can overestimate myocardial scar. Methods: We introduced an asymmetric intensity prior that penalizes predicted scar darker than a scribble-derived reference. We evaluated 36 patient-disjoint test cases using reference myocardium-guided crops and performed separate reference-free central-crop and automatic-myocardial-wall analyses. The main prior-specific comparison matched the 2.5D backbone with and without the prior across three seeds. Results: On the pre-cropped images, mean Dice increased from to , while recall fell from 0.6924 to 0.6399. The predicted-to-reference scar ratio fell from to , and absolute scar-volume error fell from to mL on the common crop grid. However, restriction to the reference myocardial wall increased absolute percent-LV-scar error from to percentage points. With additional healthy-myocardium background points (seed 2022), the matched prior model reduced Pred/GT from 1.5789 to 1.1500 but Dice fell from 0.5441 to 0.5366. With one fixed-rule reference-free central crop and retraining (seed 2022), Dice was 0.4892 without and 0.5047 with the prior. In a paired automatic-wall analysis, the prior did not improve three-seed wall-restricted burden error (6.14 versus 8.16 percentage points). Conclusions: The prior reduced full-output overestimation on the pre-cropped images, but did not improve intramyocardial burden agreement or retain a Dice advantage in the single-seed healthy-background sensitivity analysis. Its effects depend on localization and wall quality and require independent prospective reference-free evaluation.