Multi-Atlas-Based Segmentation of Pediatric Vocal Tract Anatomy in Dynamic Magnetic Resonance Imaging

Area of Science:

• Medical Imaging
• Speech Science
• Computational Anatomy

Background:

• Accurate segmentation of speech-related anatomy is vital for quantitative analysis.
• Manual segmentation is time-consuming and lacks reproducibility.
• Automated methods are needed to improve efficiency and consistency.

Purpose of the Study:

• To compare four multi-atlas segmentation methods for vocal tract structures.
• To determine the impact of temporal frame count on segmentation accuracy.
• To identify the optimal method for segmenting pediatric speech anatomy.

Main Methods:

• Developed five spatiotemporal atlases of speech tasks (approx. 80 frames each).
• Applied diffeomorphic registration to 10-40 frames per atlas.
• Evaluated four label fusion techniques: majority voting label fusion (MVLF), simultaneous truth and performance level estimation (STAPLE), multi-atlas label fusion (MALF), and corrective learning (CL).
• Assessed performance using Dice Similarity Coefficient (DSC) across increasing frame counts.

Main Results:

• Corrective learning (CL) achieved the highest segmentation accuracy, with average DSC increasing from 0.89 (10 frames) to 0.92 (40 frames).
• MVLF, STAPLE, and MALF showed less improvement, with DSC values between 0.82 and 0.87.
• Only CL demonstrated a significant positive correlation between the number of frames and DSC.

Conclusions:

• Corrective learning (CL) outperforms other methods for segmenting pediatric vocal tract structures in dynamic MRI.
• Increased temporal frame count significantly enhances CL's segmentation performance.
• CL shows promise for replacing manual segmentation in specialized, smaller datasets for speech anatomy analysis.