Multi-atlas-based segmentation of vocal tract anatomy for children using label fusion techniques from dynamic MRI

Segmentation of oral and velopharyngeal anatomy is essential for quantifying anatomical structures and analyzing the dynamic and discrete movements of these structures during speech production. This aids in the diagnosis and treatment of speech-related disorders as the different speech articulators are able to be analyzed independently and in relation to their co-articulatory movements. Direct visualization of structures and musculature important for speech production is best attained using MRI. Manual segmentation, however, is labor-intensive and often suffers from poor reproducibility, requiring automatic segmentation techniques to enhance efficiency. To address this, we aim to compare multi-atlas-based segmentation techniques to delineate the tongue, velum, and adenoid. We first constructed five spatiotemporal atlases of speech tasks, each with over 80 time frames. We treated each time frame as a single atlas, with diffeomorphic registration of 10 to 40 time frames in steps of 10 from the anchor spatiotemporal atlas with the other four atlases. Manual labels were propagated across the four atlases. Two label fusion techniques were compared: majority voting and joint label fusion, evaluated using the Dice Similarity Coefficient (DSC). Our results demonstrate that joint label fusion outperformed majority voting, achieving a DSC of over 0.84 on all four atlases, with the highest reaching a DSC of 0.87. These findings indicate that joint label fusion offers a promising tool for velopharyngeal and vocal tract motion analysis.