Title:Pathological MRI Segmentation by Synthetic Pathological Data Generation in Fetuses and Neonates

Abstract:Developing new methods for the automated analysis of clinical fetal and neonatal MRI data is limited by the scarcity of annotated pathological datasets and privacy concerns that often restrict data sharing, hindering the effectiveness of deep learning models. We address this in two ways. First, we introduce Fetal&Neonatal-DDPM, a novel diffusion model framework designed to generate high-quality synthetic pathological fetal and neonatal MRIs from semantic label images. Second, we enhance training data by modifying healthy label images through morphological alterations to simulate conditions such as ventriculomegaly, cerebellar and pontocerebellar hypoplasia, and microcephaly. By leveraging Fetal&Neonatal-DDPM, we synthesize realistic pathological MRIs from these modified pathological label images. Radiologists rated the synthetic MRIs as significantly (p < 0.05) superior in quality and diagnostic value compared to real MRIs, demonstrating features such as blood vessels and choroid plexus, and improved alignment with label annotations. Synthetic pathological data enhanced state-of-the-art nnUNet segmentation performance, particularly for severe ventriculomegaly cases, with the greatest improvements achieved in ventricle segmentation (Dice scores: 0.9253 vs. 0.7317). This study underscores the potential of generative AI as transformative tool for data augmentation, offering improved segmentation performance in pathological cases. This development represents a significant step towards improving analysis and segmentation accuracy in prenatal imaging, and also offers new ways for data anonymization through the generation of pathologic image data.