A deep learning-based approach to automatic proximal femur segmentation in quantitative CT images

Area of Science:

• Medical imaging
• Orthopedics
• Artificial intelligence

Background:

• Automatic segmentation of the proximal femur is crucial for assessing orthopedic diseases, particularly hip fracture risk.
• Differentiating cortical and trabecular bone compartments is essential for accurate bone health evaluation.

Purpose of the Study:

• To develop a fast and automatic deep learning approach for proximal femur segmentation.
• To extract periosteal and endosteal contours for differentiating bone compartments.
• To improve the efficiency of proximal femur segmentation compared to existing software.

Main Methods:

• A three-dimensional (3D) end-to-end fully convolutional neural network (CNN) was developed.
• The CNN was trained to segment periosteal and endosteal contours.
• Segmentation accuracy was validated against QCT software (MIAF-Femur) and ground truth measurements.

Main Results:

• The CNN achieved high accuracy with Dice Similarity Coefficients (DSC) of 97.82% for periosteal and 96.53% for endosteal contours.
• The developed model significantly reduced segmentation time from 30 minutes to a few minutes per case.
• Relative errors in femur volume measurements were less than 5% compared to ground truth.

Conclusions:

• The proposed deep learning approach enables fast and accurate automatic segmentation of the proximal femur.
• This method facilitates the differentiation of cortical and trabecular bone compartments.
• The approach is expected to aid in measuring bone mineral density and evaluating bone strength using Finite Element Analysis (FEA).