Session: 03-04-03: Artificial Intelligent Applications in Manufacturing III

Paper Number: 165680

Deep Learning in Orthopedic Imaging: Detectron2 for Knee Osteoarthritis Detection and Grading 

The application of Detectron2 for knee osteoarthritis (KOA) classification represents a significant advancement in medical imaging diagnostics. This research addresses a critical healthcare challenge affecting elderly populations, where traditional manual classification using the Kellgren-Lawrence system has long been hampered by subjectivity, time constraints, and experience-dependent variability.
The study's methodological approach is noteworthy for its use of Detectron2, an instance segmentation framework that surpasses traditional Convolutional Neural Networks (CNNs) in capability. By enabling precise localization of knee structures, Detectron2 facilitates more granular feature extraction, potentially capturing subtle radiographic indicators of osteoarthritis that might be overlooked in conventional approaches. This architecture choice aligns with the clinical need for anatomically aware analysis in orthopedic imaging.
The dataset composition from the Osteoarthritis Initiative (OAI) represents both a strength and limitation of the study. While the images are likely of high quality and clinically validated, the relatively modest sample size (210 training, 60 validation, and 30 test images) raises questions about the model's performance across diverse patient demographics, varying imaging protocols, and equipment specifications. Deep learning models typically benefit from larger training datasets to ensure robust generalization.
Performance metrics across both classification tasks are remarkably high. The binary classification results (98.6% accuracy, 97.8% precision, 98.4% recall, and 98.1% F1-score) suggest near-perfect discrimination between healthy knees and those affected by osteoarthritis. Similarly, the multi-class classification accuracy of 94.2% with class-wise F1-scores exceeding 92% demonstrates strong performance in grading KOA severity. The 89% Intersection over Union (IoU) for knee structure segmentation further validates the model's spatial awareness and anatomical precision.
These metrics collectively suggest a model that significantly outperforms conventional diagnostic approaches and potentially rivals experienced radiologists in accuracy. However, the exceptional performance warrants careful consideration of potential overfitting, particularly given the limited test set size. External validation across multiple institutions and diverse patient cohorts would strengthen the clinical validity of these findings.
The integration of explainable AI techniques represents a crucial advancement for clinical adoption. The interpretability component addresses the "black box" criticism often leveled at deep learning systems in healthcare, potentially increasing physician trust and facilitating regulatory approval. By providing transparency into decision-making processes, the model may serve as an educational tool for training radiologists while simultaneously augmenting expert diagnosis.
From a clinical impact perspective, the high confidence and accuracy in KOA detection could streamline diagnostic workflows, reduce inter-observer variability, and potentially enable earlier intervention. The system's ability to consistently grade osteoarthritis severity could also enhance treatment planning, disease monitoring, and outcomes assessment in clinical trials.
While promising, further validation through prospective clinical studies will be necessary to establish the real-world utility and implementation feasibility of this approach in routine radiological practice.

Presenting Author: Moneesh Rajaram Kennesaw State University

Presenting Author Biography: A freshman student who is interested in the potential of Artifical Intelligence. Studying Information Technology at Kennesaw State University.

Authors:

Moneesh Rajaram Kennesaw State University
Joshua Daniel Kennesaw State University
Xavistin C. Arul Arasu Georgia Institute of Technology
Sathish Kumar Gurupatham Kennesaw State University