Deep Learning-Based Topology-Preserving Inner Ear Subregion Segmentation in MRI

Motivation: Automatic inner ear subregion segmentation from MR images is challenging due to thin and tubular structures such as the semicircular canals. Goal(s): To develop a fully automated deep learning model for high-quality inner ear subregion segmentation, with a focus on improving connectivity in the semicircular canals. Approach: We adopted two topology-preserving methods: 1) a selective topology-focused loss applied to each subregion based on its morphological features, and 2) label-preserving data augmentation to maintain topology during training. Results: The proposed method enhanced the connectivity of the semicircular canals while maintaining volumetric overlap across all regions. Impact: The proposed inner ear subregion segmentation method may aid in diagnosing and planning treatment for auditory-related conditions, such as Meniere's disease, by enabling automatic quantification of contrast enhancement for each inner ear region.