Full-length radiograph based automatic musculoskeletal modeling using convolutional neural network.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Anatomy

Background:

• Full-length radiographs contain extensive anatomical data for pelvis, femur, and tibia.
• Current musculoskeletal modeling relies on limited parameters, potentially reducing accuracy.
• A need exists for automated methods to extract comprehensive anatomical data from radiographs.

Purpose of the Study:

• To develop a fully automatic algorithm for extracting anatomical parameters from full-length radiographs.
• To generate accurate musculoskeletal models surpassing those derived from linear scaling.
• To improve the efficiency and accuracy of musculoskeletal model creation.

Main Methods:

• A U-Net convolutional neural network was employed for segmenting pelvis, femur, and tibia from radiographs.
• Eight anatomical parameters (lengths, widths, angles) were automatically extracted from segmentation masks.
• Geometric accuracy was evaluated against CT bone models using distance errors and Jaccard index.

Main Results:

• High segmentation accuracy was achieved, with Sørensen-Dice coefficients of 0.9898 (pelvis), 0.9822 (femur), and 0.9786 (tibia).
• Algorithm-driven models showed significantly reduced maximum distance error (28.3% decrease) and RMS distance error (28.9% decrease) compared to scaled models.
• The automated modeling process was substantially faster (107 seconds) than manual methods (870 seconds).

Conclusions:

• The developed algorithm fully automates musculoskeletal model generation from full-length radiographs.
• This approach yields a ~30% reduction in geometric errors, enhancing model accuracy.
• The method holds potential for personalized musculoskeletal simulations in large populations, particularly for osteoarthritis.