FreeHemoSeg: Label-Free Deep Learning Framework for Automated Segmentation of Fetal Brain Germinal Matrix and Intraventricular Hemorrhage
Mingxuan Liu, Yi Liao, Juncheng Zhu, Haoxiang Li, Hongjia Yang, Jialan Zheng, Zihan Li, Ziyu Li, Haibo Qu, Qiyuan Tian · Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025
Motivation: Antenatal GMH-IVH is a significant cause of infant mortality and morbidity. Deep learning model for automatically diagnosing GMH-IVH requires a large amount of brain data with GMH-IVH and labels for training, which are difficult to obtain due to the rarity of GMH-IVH. Goal(s): Deep learning model that can be simply trained on data from healthy subjects for segmenting GMH-IVH. Approach: FreeHemoSeg was proposed to synthesize pseudo slices with GMH-IVH from normal images for training a neural network. Results: FreeHemoSeg exhibited superior diagnostic and segmentation accuracy, outperforming unsupervised methods and networks trained on limited labeled data from patients. Impact: FreeHemoSeg provides accurate, automated segmentation and diagnosis of GMH-IVH without hemorrhage data and labels for training, substantially simplifying clinical workflows, aiding early diagnosis and prognosis, enabling hemorrhage volume measurement, supporting large-scale neuroscience research, and enhancing prenatal care and management strategies.