Deep learning-based identification of spine growth potential on EOS radiographs

Area of Science:

• Radiology and Imaging
• Artificial Intelligence in Medicine
• Orthopedics

Background:

• Assessing spine growth potential is crucial for managing spinal conditions.
• Current methods rely on manual interpretation of EOS radiographs, which can be time-consuming.
• There is a need for automated, objective tools to aid clinicians.

Purpose of the Study:

• To develop and validate a deep learning (DL) algorithm for automatic assessment of spine growth potential and Risser sign from EOS radiographs.
• To compare the performance of the DL algorithm against human clinicians.

Main Methods:

• A DL algorithm was developed using 3383 EOS cases for training and testing.
• The algorithm was validated on an additional 440 cases.
• Performance was compared to four clinicians using metrics like kappa value, accuracy, and Youden index.

Main Results:

• The DL algorithm achieved a weighted kappa of 0.933 for Risser sign and 0.944 for spine growth potential, comparable to clinicians (0.909-0.930 and 0.911-0.934, respectively).
• The algorithm demonstrated higher accuracy (0.973) and Youden index (0.952) than the best clinician.
• The DL algorithm was significantly faster (15.2 s/40 cases) than clinicians (177.2–241.2 s/40 cases).

Conclusions:

• The DL algorithm shows comparable or superior performance to clinicians in assessing spine growth potential.
• The developed algorithm is stable, efficient, and convenient, offering a promising approach to assist clinicians.
• This automated method can quickly ascertain spine growth potential from EOS radiographs, assisting busy clinicians.