A Multi-Modal Pelvic MRI Dataset for Deep Learning-Based Pelvic Organ Segmentation in Endometriosis.

Area of Science:

• Medical Imaging
• Radiology
• Pelvic Anatomy

Background:

• Endometriosis impacts 190 million females globally, necessitating advanced diagnostic tools.
• Magnetic Resonance Imaging (MRI) is the preferred non-invasive method for endometriosis diagnosis.
• Accurate segmentation of pelvic structures in MRI is crucial for endometriosis assessment.

Purpose of the Study:

• To present novel multicenter female pelvic MRI datasets for endometriosis research.
• To evaluate the baseline performance of auto-segmentation pipelines (nnU-Net and RAovSeg) for endometriosis-related pelvic structures.
• To provide a publicly available resource for advancing pelvic MRI auto-segmentation.

Main Methods:

• Collected multi-sequence endometriosis MRI scans from two clinical institutions.
• Created a multicenter dataset (51 subjects) with manual labels for interrater agreement assessment.
• Developed ovary auto-segmentation pipelines using a single-center dataset (81 subjects).

Main Results:

• Publicly available datasets include uterus, ovary, and detectable endometrioma segmentations.
• Demonstrated baseline performance of nnU-Net and RAovSeg auto-segmentation pipelines.
• Highlighted challenges in manual ovary segmentation, underscoring the need for automated solutions.

Conclusions:

• The presented datasets are valuable for endometriosis research and pelvic MRI auto-segmentation development.
• Automated segmentation methods are essential for improving the efficiency and accuracy of endometriosis diagnosis.
• Public data sharing accelerates innovation in medical imaging for gynecological conditions.