Measurement fidelity and external validation of automated breast DCE-MRI segmentation
Automated breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) segmentation could facilitate quantitative tumor measurement, but geometric overlap alone does not establish the fidelity of mask-derived measurements. We developed an enhancement-aware three-dimensional MedNeXt framework with gated skip fusion and boundary supervision and evaluated it in a retrospective multi-cohort design spanning locked internal testing, source-held-out assessment, retrospective external validation, and radiologist review. Measurement fidelity was the primary endpoint. Across 521 paired internal-test and DUKE cases, median errors were 9.7% for tumor volume, 1.8 mm for 3D maximum surface diameter, and 10.2% for operational enhancement-defined volume, with corresponding ICCs of 0.97, 0.80, and 0.97. Median Dice was 0.831 internally and 0.823 in the DUKE source-held-out stratum, and was 0.812 and 0.806 in the BC-MRI-SEG and public BreastDCEDL_AMBL retrospective external-validation cohorts, respectively, using the frozen checkpoint without external-label tuning. No major correction was required for 78 of 80 masks in the blinded two-reader review. In 276 paired response-analysis cases, AUROC was 0.732 with automatic geometry and 0.757 with expert geometry (difference, −0.025; 95% CI, −0.095 to 0.045); discrimination non-inferiority was not demonstrated. These findings support automated segmentation as a starting point for radiologist-supervised retrospective quantitative research within the evaluated cohorts, while distinguishing measurement fidelity from downstream predictive equivalence.
Authors
- 黄元夕
- Jian Zheng (ORCID: https://orcid.org/0000-0003-4824-9465)
- Xu Jin
- Chunlei Tan
- Shiyuan Zhang
- Li Chen
Institutions
- Harbin Medical University (CN)
- Chinese Academy of Medical Sciences & Peking Union Medical College (CN)
- Third Affiliated Hospital of Harbin Medical University (CN)
- Cancer Hospital of Chinese Academy of Medical Sciences (CN)
Publication Details
- Journal
- npj Digital Medicine
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1038/s41746-026-03446-z
- Primary Topic
- Medical Image Segmentation Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00