Enhanced deep learning model enables accurate alignment measurement across diverse institutional imaging protocols

Area of Science:

• Orthopedic imaging analysis
• Artificial intelligence in radiology
• Medical image processing

Background:

• Radiographic measurements for leg alignment face challenges in consistency across diverse equipment and protocols.
• This study enhances a deep learning (DL) model to address these inconsistencies in full-leg radiographs.

Purpose of the Study:

• To evaluate an advanced DL model's proficiency in generating uniform and precise alignment measurements for full-leg radiographs.
• To assess the model's performance irrespective of institutional imaging differences and protocol variations.

Main Methods:

• An enhanced DL model was trained on over 10,000 full-leg radiographs using a segmented approach for hip, knee, and ankle regions.
• External validation involved 300 datasets from three institutes; seven key radiologic parameters were measured.
• Model measurements were compared to orthopedic specialist evaluations using intraclass correlation coefficients (ICCs) and absolute error analysis.

Main Results:

• The DL model demonstrated excellent performance with inter-observer ICCs (0.936-0.997) and perfect intra-observer ICC (1.000), matching specialist accuracy.
• Consistent and robust accuracy was observed across different institutional imaging protocols.
• Processing time was reduced 30-fold, from 10-11 seconds to 300 milliseconds, with minimal absolute error in hip-knee-ankle angle measurements.

Conclusions:

• The enhanced DL model provides accurate and rapid alignment measurements for full-leg radiographs.
• Its ability to overcome protocol variations indicates broad potential for clinical and research applications in orthopedic imaging.