PRORED: a hybrid transformer framework with progressive refinement decoding for segmenting dynamic speech MRI.

Area of Science:

• Medical imaging
• Artificial intelligence
• Speech science

Background:

• Dynamic MRI of the upper vocal tract is crucial for speech analysis.
• Manual segmentation of speech MRI data is labor-intensive and time-consuming.
• Development of automatic segmentation methods is essential for efficient analysis.

Purpose of the Study:

• To introduce a novel hybrid transformer network for accurate segmentation of dynamic speech MRI.
• To improve the analysis of speech production and velopharyngeal closure.
• To develop a generalizable model for medical image segmentation.

Main Methods:

• A deep learning-based decoder model, Progressively Refinement Decoding (PRORED), was developed.
• A directional field (DF) module was incorporated to capture and refine feature contour details.
• The DF module was integrated at multiple decoder stages for progressive feature enhancement.

Main Results:

• The model achieved a 97.78% Dice coefficient and 6.84 mm Hausdorff distance on speech MRI data.
• It demonstrated superior efficiency in identifying closure patterns compared to baseline methods.
• The model achieved a 91.90% Dice score on a cardiac dataset, indicating generalizability.

Conclusions:

• The proposed model enables more accurate segmentation of speech MRI, particularly for velopharyngeal closure studies.
• The model's strong performance on a cardiac dataset highlights its generalizability across different medical imaging applications.
• This work presents the first model combining vision transformers and progressive refinement decoding for dynamic speech MRI segmentation.