Vision transformer-based automated segmentation of patient-specific Velopharyngeal Anatomy

Velopharyngeal dysfunction (VPD) caused by inadequate separation of oral and nasal cavities can have a substantial impact on the speech mechanism, communication abilities and quality of life. Diagnosis and treatment of VPD requires accurate and objective pre-operative assessment of Velopharyngeal (VP) anatomy. Recently, an innovative child-friendly magnetic resonance imaging (MRI) protocol has been developed to analyze and quantify the VP anatomy in children and young adults. However, previous approaches to quantify and measure the VP structures from the 3D MRI were based on time-intensive manual inspections, measurement, and segmentation. In this work, we are developing a deep learning-based image segmentation algorithm utilizing vision transformer to automatically identify and quantify the VP structures of normal and abnormal VP anatomy. This approach will enable automated, time-efficient, and improved pre-operative clinical assessment of VP anatomy for children born with VP anomalies, and design appropriate patient-specific surgical treatment plan to establish normal speech mechanism.