Open-Source Manually Annotated Vocal Tract Database for Automatic Segmentation from 3D MRI Using Deep Learning:...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Speech Science

Background:

• Accurate vocal tract segmentation from 3D MRI is crucial for voice, speech, and singing applications.
• Manual segmentation is time-consuming and prone to errors, necessitating automated solutions.

Purpose of the Study:

• To evaluate the efficacy of deep learning algorithms for automatic vocal tract segmentation from 3D MRI data.
• To compare the performance of different deep learning architectures in segmenting the vocal tract.

Main Methods:

• Four deep learning architectures were evaluated: 2D slice-by-slice U-Net, 3D U-Net, 3D U-Net with transfer learning, and 3D transformer U-Net (3D U-NetR).
• A dataset of 53 vocal tract volumes from 10 French speakers was used, with manual annotations serving as reference segmentations.
• Performance was assessed using Dice coefficient, Hausdorff distance, and structural similarity index measure.

Main Results:

• 3D U-Net and 3D U-Net with transfer learning achieved the highest Dice coefficients (0.896).
• Transfer learning models performed comparably to 3D U-Net using less training data and showed lower variability in Hausdorff distance.
• All models struggled with specific sounds like /kõn/ and segmentation near bony regions; however, oropharyngeal and laryngopharyngeal spaces were generally segmented accurately.

Conclusions:

• 3D convolutional networks, particularly with transfer learning, are effective for automatic vocal tract segmentation from 3D MRI.
• Future research should aim to improve segmentation of challenging vocal tract configurations and boundary delineations.