Contrastive Learning Approach for Assessment of Phonological Precision in Patients with Tongue Cancer Using MRI Data.

Area of Science:

• Medical imaging and biomechanics of speech production.
• Computational linguistics and machine learning applications in healthcare.

Background:

• Magnetic Resonance Imaging (MRI) visualizes vocal tract dynamics for speech analysis.
• Clinical speech assessment benefits from MRI, especially with phonological approaches.
• Acoustic data absence limits MRI-only speech sound recognition accuracy.

Purpose of the Study:

• To develop a contrastive learning approach for improved phonological class detection from MRI data.
• To enhance speech sound recognition accuracy when acoustic signals are unavailable during inference.
• To validate the clinical utility of the proposed method in assessing speech disorders.

Main Methods:

• Utilized Magnetic Resonance Imaging (MRI) data of speech production.
• Implemented a contrastive learning framework to analyze vocal tract dynamics.
• Evaluated frame-wise recognition of phonological classes using the developed approach.

Main Results:

• Contrastive learning improved frame-wise phonological class recognition from 0.74 to 0.85 F1-score.
• Demonstrated enhanced accuracy in detecting phonological classes from MRI data alone.
• Showcased promising results for clinical application in speech disorder assessment.

Conclusions:

• Contrastive learning effectively improves phonological class detection from MRI data without acoustic signals.
• The method shows significant potential for clinical applications, particularly in speech disorder assessment for patients like those with tongue cancer.
• This approach offers a viable solution for speech analysis when audio recordings are not feasible or available.