A Staged Deep Learning Framework for Endometriosis-Aware Pelvic MRI Analysis Using UT-EndoMRI

Abstract

Endometriosis remains challenging to assess on pelvic MRI because lesions are often small, heterogeneous, and poorly distinguished from surrounding anatomy. This study developed a reproducible two-stage framework using the public UT-EndoMRI dataset to investigate whether lesion-aware segmentation can provide informative representations for subject-level endometriosis-aware prediction. In Stage 1, SegResNet, SwinUNETR, and SwinUNETR with lesion-aware refinement were compared for multi-class pelvic MRI segmentation. The refined SwinUNETR achieved the best test macro-Dice of 0.0786 and macro IoU of 0.0447, although ovary and lesion segmentation remained weak. The selected model was subsequently frozen and used to generate segmentation-informed features for Stage 2. A cohort of 121 subjects comprising 110 proxy/inferred positive and 11 proxy negative cases was evaluated using repeated five-fold stratified cross-validation over five repeats with L2-regularized logistic regression. The combined uncalibrated feature model achieved a balanced accuracy of 0.8215 ± 0.0798, MCC of 0.3953 ± 0.1172, and AUC of 0.8597 ± 0.0578. Subject-aggregated estimates were balanced accuracy 0.8364 (95% CI: 0.7955–0.8818), MCC 0.3968 (95% CI: 0.3407–0.4765), and AUC 0.8620 (95% CI: 0.7793–0.9306; permutation p = 0.00498). Isotonic calibration improved probability quality, reducing the Brier score to 0.0653 and ECE to 0.0380. These findings indicate that segmentation-informed representations can support internal subject-level prediction beyond coarse MRI summaries; however, the small negative class, weak lesion segmentation, proxy/inferred labels, possible target–predictor dependency, and absence of external validation preclude claims of clinical diagnostic validity.