Multi-label segmentation of carpal bones in MRI using expansion transfer learning

Area of Science:

• Medical Imaging
• Deep Learning
• Orthopedics

Background:

• Accurate segmentation of carpal bones from MRI is crucial for wrist therapy planning and dynamic analysis.
• Existing deep learning methods often require large datasets, which are challenging to acquire for specific anatomical structures like carpal bones.

Purpose of the Study:

• To develop a robust deep learning approach for multi-label segmentation of all eight carpal bones using a small in-vivo MRI dataset.
• To enable precise therapy planning and wrist dynamic analysis through accurate carpal bone segmentation.

Main Methods:

• A two-stage segmentation pipeline using a modified 3D U-Net architecture was proposed.
• The first stage employed expansion transfer learning (ETL) for pretraining on a focused region of interest (ROI) and subsequent expansion to the full field of view (FOV).
• The second stage utilized the ROI bounding box for high-accuracy, labeled segmentation of the eight carpal bones.

Main Results:

• The proposed approach achieved an average Dice Similarity Coefficient (DSC) of 87.8%, Average Surface Distance (ASD) of 0.46 mm, and Hausdorff Distance (HD) of 2.42 mm across all datasets.
• After applying exclusion criteria to 12 datasets, performance improved to 96.1% DSC, 0.16 mm ASD, and 1.38 mm HD.
• The method demonstrated superior performance compared to four state-of-the-art approaches.

Conclusions:

• This work presents the first Convolutional Neural Network (CNN)-based multi-label segmentation approach for human carpal bones in MRI.
• The novel ETL technique effectively improved the localization of small ROIs within large FOVs.
• The combination of a two-stage approach and ETL yielded highly accurate segmentation results despite a limited dataset size.