Abstract
Deep infiltrative endometriosis (DIE), a subtype of endometriosis, is a chronic inflammatory disease where endometrial tissue (\textit{i.e.}, lesions) grows outside the uterus and infiltrates pelvic organs such as the bladder or rectum. Accurate lesion mapping is crucial for diagnosis and surgical planning, notably for computing standardized severity scores such as the deep Pelvic Endometriosis Index (dPEI). Yet DIE lesions are small, low-contrast, and ill-delineated, making their detection on T2-weighted MRI challenging. On a retrospective, lesion-level-annotated cohort, we systematically compare four deep-learning formulations under a common instance-level protocol: dense segmentation (nnU-Net), bounding-box detection (YOLO), a latent token grid, and a patch-level classifier. We show that standard baseline formulations are insufficient, and the token-grid formulation yields only a partial, unstable spatial signal. The patch-level approach performs best, achieving a high local classification score in cross-validation but degrades significantly under full-slice sliding-window inference. An oracle analysis, in which windows are centered on lesions, shows that the local lesion signal is in fact highly separable. Taken together, these results indicate that the main bottleneck is not recognizing a lesion within a region, but generating anatomically relevant candidate regions across complete MRI slices. We argue that anatomically constrained candidate generation is the most promising direction towards a usable lesion mapping pipeline.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CAPI_WOMEN_006.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=sIgEr17oFn
BibTex
@InProceedings{LegEli_Comparing_MICCAISAT2026,
author = { Leguy, Eliot AND Nyangoh-Timoh, Krystel AND Jannin, Pierre AND Baxter, John S. H. AND Germani, Elodie},
title = { { Comparing Problem Formulations for Endometriosis Lesion Detection on T2-weighted Pelvic MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}