Multi-Task Learning U-Net for Functional Shoulder Sub-Task Segmentation.

Area of Science:

• Biomedical Engineering
• Medical Imaging Analysis
• Rehabilitation Technology

Background:

• Accurate assessment of frozen shoulder (FS) is vital for treatment evaluation.
• Manual labeling of functional shoulder sub-tasks is time-consuming and error-prone.
• Automated tools are needed for reliable clinical assessment of FS.

Purpose of the Study:

• To develop an automatic and reliable functional shoulder sub-task segmentation (STS) tool for frozen shoulder (FS) clinical evaluation.
• To propose a deep multi-task learning (MTL) U-Net integrating STS and transition point detection (TPD).
• To enhance the accuracy of shoulder sub-task segmentation using auxiliary TPD information.

Main Methods:

• A deep multi-task learning (MTL) U-Net architecture was designed for STS.
• An auxiliary task of transition point detection (TPD) using lightweight CNNs was incorporated.
• Joint optimization of STS and TPD objective functions was performed.
• Wearable inertial measurement units (IMUs) collected 815 shoulder task sequences from 20 healthy subjects and 43 FS patients.

Main Results:

• The deep MTL U-Net demonstrated superior performance compared to single-task models.
• The proposed method achieved effective functional shoulder sub-task segmentation (STS).
• The integration of TPD improved the boundary detection accuracy for STS.

Conclusions:

• The developed deep MTL U-Net provides an automatic and reliable tool for functional shoulder STS.
• This approach offers significant potential for improving clinical evaluation in frozen shoulder (FS) patients.
• The study highlights the effectiveness of multi-task learning for complex biomechanical analysis.