Unpaired multi-modal multi-label learning for detecting endometriosis signs
Yuan Zhang1, Hu Wang2, Yutong Xie2
1Robinson Research Institute, The University of Adelaide, North Adelaide, SA 5005, Australia; Australian Institute for Machine Learning (AIML), The University of Adelaide, Adelaide, SA 5000, Australia.
Artificial Intelligence in Medicine
|August 13, 2026
Summary
Endometriosis diagnosis can be improved using a new AI framework, EndoFusion, that combines MRI and TVUS scans. This multi-modal approach enhances detection of key endometriosis signs like Pouch of Douglas obliteration and bowel nodules.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Gynecology
Background:
- Endometriosis is a prevalent gynecological disorder causing pain and infertility.
- Current diagnostic methods like laparoscopy are invasive, slow, and costly.
- Non-invasive imaging, including transvaginal ultrasound (TVUS) and magnetic resonance imaging (MRI), is crucial but faces challenges in integrating data from single modalities.
Purpose of the Study:
- To develop a novel framework, EndoFusion, for non-invasive, multi-modal detection of endometriosis imaging signs.
- To enable accurate detection of Pouch of Douglas (POD) obliteration and bowel nodules (BN) using unpaired TVUS and MRI data.
- To leverage complementary strengths of different imaging modalities and model correlations between endometriosis signs.
Main Methods:
- Proposed EndoFusion, an unpaired multi-modal, multi-label learning framework for detecting POD and BN from TVUS and MRI.
- Introduced label-based pairing, mixup, and cross-modal feature exchange for robust single-modality inference.
- Implemented Dynamic Mutual Knowledge Distillation (DMKD) for adaptive cross-modal knowledge transfer and modeled label correlations using multi-head attention.
Main Results:
- EndoFusion significantly outperformed comparison methods in detecting endometriosis imaging signs.
- Achieved an average Area Under the Curve (AUC) of 0.827 (95% CI: 0.790-0.861) using single-modality inference.
- Demonstrated proof-of-concept for multi-modal, non-invasive assessment of endometriosis signs.
Conclusions:
- EndoFusion offers a robust approach for integrating TVUS and MRI data for endometriosis diagnosis.
- The framework effectively transfers knowledge across modalities and handles imbalanced data and label correlations.
- This work advances non-invasive diagnostic capabilities for endometriosis, improving patient care.
Keywords:
Bowel noduleEndometriosisKnowledge distillationMulti-label classificationPouch of Douglas obliterationUnpaired multi-modal learning
