Near-duplicate contamination inflates brain MRI benchmark performance: a size-matched, multi-seed MedNeXt-B study

Pre-partitioned brain tumor MRI collections may place exact or near-duplicate slices on both sides of a benchmark split, but the magnitude and uncertainty of the resulting performance difference are rarely measured.

A unified MedNeXt-B 2D model produced a four-class label and a tumor mask from a shared encoder. For each near-duplicate-screened fold, we constructed a replacement-based contaminated condition of identical size and tumor-class-by-plane composition; this condition was not the original unscreened training pool. Validation folds and the frozen 995-slice test set were identical. Constituent source-dataset matching was not possible because BRISC2025 provides no per-image source provenance. Both conditions were trained with seeds 42, 1,337, 2,024, 3,407, and 5,090. Primary uncertainty summaries used a 10,000-resample hierarchical bootstrap over seeds and inferred pseudo-patient clusters.

Across five seeds, screened accuracy was 0.9652 ± 0.0023 and mean class Dice was 0.7894 ± 0.0031 (mean ± sample SD). Accuracy was higher in the replacement-based contaminated condition by a mean difference of 0.0261 ± 0.0031 (range 0.0211-0.0281; hierarchical 95% CI, 0.0207-0.0318); corresponding mean differences were 0.0315 ± 0.0049 for mean class Dice (range, 0.0252-0.0375; hierarchical 95% CI, 0.0238-0.0394) and 0.0746 ± 0.0055 for meningioma recall (range, 0.0660-0.0792; hierarchical 95% CI, 0.0590-0.0911). The screening surrogate recovers only 5.8% of known same-patient slice pairs in an independent calibration dataset, so the screened condition may retain undetected related slices, and the intervals do not establish patient-level independence.

Retaining detected test-linked near-duplicates was associated with higher slice-level benchmark performance in this replacement-matched exploratory analysis.