Automatic hip osteoarthritis grading with uncertainty estimation from computed tomography using...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Hip osteoarthritis (OA) progression causes pain and disability, often necessitating hip arthroplasty.
• Current Crowe and Kellgren-Lawrence (KL) classifications for hip OA severity are subjective.
• Accurate grading is crucial for treatment planning and disease progression monitoring.

Purpose of the Study:

• To develop an automated deep learning approach for classifying hip OA severity.
• To classify hip OA severity using both Crowe and KL grading schemes.
• To assess model uncertainty as a predictor of classification accuracy.

Main Methods:

• Deep learning models were trained for automatic hip OA severity grading.
• Models predicted Crowe and KL grades separately and a combined ordinal label.
• Model uncertainty was estimated and validated against classification accuracy.
• Training and validation involved 197 hip OA patients, with external validation on 52 patients.

Main Results:

• Deep learning models achieved comparable accuracy in classification and regression settings (approx. 0.65 ECA, 0.95 ONCA).
• Model uncertainty was significantly higher for cases with larger classification errors.
• The developed models demonstrated robust performance on both internal and external datasets.

Conclusions:

• An automated method for grading hip OA severity from CT images was successfully developed.
• The models achieved high one-neighbor class accuracy (ONCA), enabling automated grading in large datasets.
• Model uncertainty correlates with classification accuracy, allowing for error prediction and potential for further disease progression analysis.