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

Endometriosis is a prevalent gynecological condition affecting approximately 190 million women of reproductive age globally. Due to its complex presentation and the invasive nature of diagnostic procedures, Magnetic Resonance Imaging (MRI) has emerged as the preferred non-invasive approach for diagnosing this condition. Recognizing the critical need for advanced diagnostic tools, this study introduces a novel multi-modal pelvic MRI dataset specifically curated for deep learning applications aimed at segmenting pelvic organs in endometriosis patients. The dataset comprises MRI scans collected from two distinct clinical institutions, encompassing a total of 132 female subjects. The first component is a multicenter dataset including 51 subjects, which features manual annotations of multiple pelvic structures provided independently by three expert raters to assess interrater reliability. The second component is a single-center dataset with 81 subjects, annotated by a single rater, which was used to develop and refine automated ovary segmentation pipelines. To benchmark segmentation performance, two auto-segmentation methods were evaluated: the self-configuring nnU-Net and RAovSeg, a custom-designed convolutional neural network. Both models were tasked with delineating critical pelvic organs, primarily the uterus and ovaries, which were annotated for all subjects. Additionally, the dataset includes segmentations of endometriomas for all subjects where such lesions were visible. This comprehensive annotation scope supports robust training and evaluation of segmentation algorithms tailored to the complexities presented by endometriosis. A significant challenge highlighted by the study is the difficulty in manual ovary segmentation within endometriosis MRI scans. Variability among raters underscores the subjective nature of manual delineations and reinforces the necessity for reliable, reproducible automated segmentation methods. The availability of this dataset aims to facilitate further research and development of deep learning models, ultimately improving diagnostic accuracy and patient outcomes. Importantly, the dataset has been made publicly accessible, promoting transparency and enabling the wider scientific community to engage in pelvic MRI auto-segmentation research. The study's open-access nature encourages collaborative efforts toward refining segmentation algorithms, potentially accelerating advancements in both the diagnosis and treatment monitoring of endometriosis.