A comparative study on state-of-the-art deep learning based vocal tract segmentation methods in volumetric sustained speech MRI

Motivation: This work is motivated by the need to improve MRI-based quantitative assessments of vocal tract postures in speech and voice studies. Goal(s): The goal is to compare state-of-the-art segmentation methods in volumetric vocal tract MRI segmentation, and provide insights into the their effectiveness. Approach: This comparative study examines four different U-Net architectures. All networks are trained and tested on an open-source French speaker database in a consistent manner to assess their performance with limited data. Results: Our findings indicate that transfer learning is particularly effective when training with small datasets. Additionally, we identified variability in dice coefficient between different segmenters. Impact: This study informs researchers about various state-of-the-art segmentation methods for upper airway MRI. It emphasizes the strengths and weaknesses of each method and identifies which methods work efficiently under specific conditions.