Confidence-Driven Deep Learning Framework for Early Detection of Knee Osteoarthritis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Knee Osteoarthritis (KOA) is a common cause of disability, impacting mobility and quality of life.
• Current KOA diagnosis relies on subjective Kellgren-Lawrence (KL) grading, leading to diagnostic variability.
• Early and accurate KOA detection is crucial for timely intervention and management.

Purpose of the Study:

• To develop a confidence-driven deep learning framework for early KOA detection.
• To accurately differentiate between KL-0 (no OA) and KL-2 (mild OA) stages.
• To create an AI tool that assists radiologists and reduces diagnostic workload.

Main Methods:

• A Siamese-based deep learning framework with multi-level Global Average Pooling (GAP) for feature extraction.
• A hybrid loss strategy to manage annotation uncertainty by partitioning samples into confidence subsets.
• Training and validation on the Osteoarthritis Initiative (OAI) dataset.

Main Results:

• The framework achieved accuracy, sensitivity, and specificity comparable to expert radiologists.
• High Cohen's kappa values ($\kappa$ > 0.85) indicated substantial agreement with expert diagnoses.
• McNemar's test showed no significant difference ($p$ > 0.05) between the model and radiologists.
• Confidence distribution analysis demonstrated the model mimics radiologists' decision-making.

Conclusions:

• The proposed deep learning framework offers a robust and reliable method for early KOA detection.
• This AI tool has the potential to serve as an auxiliary diagnostic aid, enhancing clinical efficiency.
• The confidence-driven approach improves model robustness in handling diagnostic uncertainty.