Automatic segmentation of trabecular and cortical compartments in HR-pQCT images using an embedding-predicting U-Net...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Orthopedics

Background:

• High-resolution peripheral quantitative computed tomography (HR-pQCT) is vital for in vivo bone microarchitecture quantification.
• Accurate image segmentation is crucial for extracting reliable microarchitectural parameters from HR-pQCT data.
• Current semi-automated segmentation methods are time-consuming, introduce bias, and hinder clinical adoption.

Purpose of the Study:

• To develop and validate a fully automated algorithm for segmenting HR-pQCT images of the radius and tibia.
• To overcome the limitations of manual segmentation protocols in HR-pQCT analysis.
• To enhance the efficiency and reproducibility of bone microarchitecture assessment.

Main Methods:

• A multi-slice 2D U-Net model was employed for initial segmentation predictions.
• Traditional morphological image filters were used for post-processing segmentation outputs.
• The algorithm was trained on a large dataset (1822 images, 896 participants) and validated on a separate dataset (386 images, 190 participants).

Main Results:

• The automated segmentation demonstrated excellent agreement with reference standards for morphological parameters (R² > 0.938).
• Precision of parameter quantification was significantly improved, particularly for cortical porosity.
• The algorithm achieved high accuracy on a large, diverse dataset.

Conclusions:

• The proposed automated segmentation algorithm is robust and accurate for HR-pQCT images.
• This innovation significantly improves the feasibility of HR-pQCT in both research and clinical settings.
• Automated segmentation facilitates more streamlined and reliable bone microarchitecture analysis.