Classification models for arthropathy grades of multiple joints based on hierarchical continual learning

Area of Science:

• Artificial Intelligence in Medical Imaging
• Orthopedics
• Radiology

Background:

• Arthropathy classification is crucial for diagnosis and treatment planning.
• Current methods may lack efficiency and scalability for large-scale studies.
• A continually updatable model is needed for diverse anatomical structures.

Purpose of the Study:

• To develop a hierarchical continual arthropathy classification model for multiple joints.
• The model aims for continuous updates suitable for large-scale studies.
• To evaluate the model's performance and explainability.

Main Methods:

• A Hierarchical Dynamically Expandable Representation (Hi-DER) model was developed.
• Trained on 1371 radiographs (knee, elbow, ankle, shoulder, hip) from three hospitals.
• Evaluated at three hierarchical levels (L1, L2, L3) using five-fold cross-validation and external datasets.

Main Results:

• Achieved high weighted average AUCs: 0.994 (L1), 0.980 (L2), 0.973 (L3) on internal data.
• Demonstrated robust performance with AUCs above 0.800 even with 70% input region erasure.
• External validation on hip and knee joints showed strong performance (AUCs > 0.934 for L1).

Conclusions:

• The Hi-DER model accurately classifies arthropathy grades across multiple joints.
• It enhances diagnostic efficiency and supports early treatment initiation.
• The model's hierarchical and continual learning capabilities offer scalability for research.