Healthy versus pathological learning transferability in shoulder muscle MRI segmentation using deep convolutional...

Area of Science:

• Medical Imaging
• Machine Learning
• Orthopedics

Background:

• Accurate segmentation of pathological shoulder muscles from MRI is crucial for diagnosing musculoskeletal diseases and planning interventions.
• Manual segmentation is time-consuming and prone to variability, highlighting the need for automated solutions.
• Limited annotated pediatric data presents a significant challenge for developing robust deep learning models.

Purpose of the Study:

• Investigate the feasibility of automated pathological shoulder muscle segmentation using deep learning with limited pediatric data.
• Evaluate learning transferability from healthy to pathological muscle data to enhance model generalizability.
• Propose advanced deep convolutional encoder-decoder architectures, leveraging pre-trained encoders, to improve segmentation accuracy.

Main Methods:

• Utilized deep convolutional encoder-decoder architectures with encoders pre-trained on ImageNet for improved segmentation.
• Employed a leave-one-out cross-validation strategy on a dataset of 24 shoulder MRI scans from patients with obstetrical brachial plexus palsy.
• Focused on segmenting four key rotator cuff muscles: deltoid, infraspinatus, supraspinatus, and subscapularis.

Main Results:

• Achieved high Dice scores for deltoid (82.4%), infraspinatus (82.0%), and subscapularis (82.8%) muscles.
• Demonstrated a Dice score of 71.0% for the supraspinatus muscle.
• Maintained absolute surface estimation errors below 83 mm² for most muscles, with supraspinatus at 134.6 mm².

Conclusions:

• Deep learning models, particularly those leveraging transfer learning from non-medical datasets like ImageNet, show promise for automated pathological shoulder muscle segmentation.
• Jointly utilizing healthy and pathological data significantly improves model performance and generalizability.
• The developed automated segmentation method offers a valuable tool for clinical assessment and management of musculoskeletal disorders.