Enhancing Pathological Fetal MRI Segmentation through Generative AI: A Novel Approach to Synthetic Pathological Data Generation

Motivation: Pathological fetal brains are difficult to segment due to limited annotated MRI data. Additionally, privacy concerns often restrict data sharing. Therefore, we need alternative augmentation techniques for improved segmentation of the pathological brain. Goal(s): To generate realistic synthetic pathological fetal brain MRI data using generative AI and improve segmentation accuracy, specifically focusing on severe ventriculomegaly. Approach: We trained a stable diffusion model on fetal MRI-label pairs, generated synthetic pathological MRI-labels derived from healthy MRIs through morphological alterations, and evaluated segmentation performance. Results: The approach generated diverse high-quality synthetic pathological fetal brain MRIs and substantially improved segmentation performance, particularly for ventriculomegaly cases. Impact: Our approach overcomes challenges of limited annotated pathological MRI datasets, facilitating the training of robust segmentation models without the need for pathological data. This advancement is an important step towards addressing privacy issues while improving segmentation performance in prenatal imaging.