Yan Zhuang, Boah Kim, Tejas Sudharshan Mathai, Pritam Mukherjee, Brandon Khoury, Ronald Summers
Automated segmentation of various structures and organs in MRI lags behind its CT counterpart due to the lack of an annotated multiparametric MRI dataset, which is labor-intensive, time-consuming, and requires expertise to build. A widely-used practice to reduce annotation workload and enhance annotation efficiency is to first obtain initial pseudo labels, and then refine them progressively in an iterative fashion. In this study, we propose a novel diffusion-based synthetic segmentation method to obtain pseudo labels for multiple organs and structures in abdominal MRI images without using any MRI annotations. Specifically, the proposed approach leverages the invariance of human anatomy across imaging modalities, and segments 13 different organs and structures in T1-weighted (T1w) abdominal MRI exams. We achieve this by learning with unpaired and publicly available annotated CT data. To validate the proposed method, we curated a multiparametric abdominal T1w MRI dataset containing pre-contrast and dynamic contrast-enhanced sequences, and conducted comprehensive comparisons against prior state-of-the-art approaches. Experiments revealed the proposed method surpassed other alternatives, especially for large abdominal organs and structures. Our approach reduces the annotation burden for building large-scale multiparametric MRI datasets, thereby facilitating the potential development of MRI-based multi-organ segmentation tools.