Related Experiment Video
Updated: Jul 29, 2025
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
2.5K
Automatic Multiple Articulator Segmentation in Dynamic Speech MRI Using a Protocol Adaptive Stacked Transfer Learning
Subin Erattakulangara1, Karthika Kelat1, David Meyer2
1Roy J. Carver Department of Biomedical Engineering, University of Iowa, Iowa City, IA 52242, USA.
Bioengineering (Basel, Switzerland)
|May 27, 2023
Summary
This study introduces a novel stacked transfer learning U-NET model for segmenting the vocal tract in dynamic speech MRI scans. The model accurately analyzes speech production by leveraging pre-trained features, achieving expert-level segmentation with minimal protocol-specific data.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Speech Science
Background:
- Dynamic magnetic resonance imaging (MRI) is crucial for understanding speech production by analyzing vocal tract airspace and articulator movement.
- Fast speech MRI protocols generate high-frame-rate datasets (80-100 fps), enabling detailed functional analysis.
- Accurate segmentation of the deforming vocal tract in these dynamic images is essential for quantitative analysis.
Purpose of the Study:
- To propose and evaluate a stacked transfer learning U-NET model for segmenting the vocal tract in 2D mid-sagittal dynamic speech MRI slices.
- To demonstrate the model's adaptability across different fast speech MRI protocols with limited labeled data.
- To compare the model's segmentation performance against expert human annotators and a conventional U-NET model.
Main Methods:
- A stacked transfer learning U-NET architecture was developed, integrating low-, mid-, and high-level features.
- Low- and mid-level features were initialized using models pre-trained on brain tumor MR, lung CT, and airway datasets.
- High-level features were fine-tuned on protocol-specific dynamic speech MRI data from three distinct acquisition protocols.
Main Results:
- The proposed model achieved accurate vocal tract segmentations comparable to expert human annotators.
- Performance was validated across three different fast speech MRI protocols (radial, uniform spiral, variable density spiral) with varying regularization techniques.
- The transfer learning approach required only a small number of protocol-specific images (approx. 20) for successful adaptation.
Conclusions:
- Stacked transfer learning offers an effective method for segmenting dynamic speech MRI data, even with limited labeled samples.
- The model demonstrates robustness and adaptability to diverse speech MRI acquisition protocols and speech tasks.
- This approach significantly enhances the ability to analyze vocal tract dynamics during speech production using MRI.
