Potential of AI-Based Diagnostic Grading System for Knee Osteoarthritis (KOA)

Saman Shahid¹Aamir Wali¹Aatir Javaid²Shahid Zikria³Onur Osman⁴Jawad Rasheed^5,6,7,8*

• ¹National University of Computer and Emerging Sciences - Lahore Campus, Lahore, Pakistan
• ²The University of Lahore, Lahore, Pakistan
• ³Information Technology University, Lahore, Pakistan
• ⁴Istanbul Topkapi Universitesi, Istanbul, Türkiye
• ⁵Istanbul Sabahattin Zaim University, Istanbul, Türkiye
• ⁶Istanbul Nisantasi Universitesi, Sarıyer, Türkiye
• ⁷Istanbul Medipol University, Istanbul, Türkiye
• ⁸Applied Science Private University, Amman, Jordan

Background: Knee osteoarthritis (KOA) is a progressive musculoskeletal disorder and a leading cause of disability worldwide. Early and accurate diagnosis is crucial for timely intervention; however, conventional manual grading using radiographs is prone to variability. Artificial intelligence (AI)-based computer-aided diagnostic (CAD) systems offer potential to improve detection and grading accuracy. Objective: This study aimed to develop and evaluate an AI-based diagnostic grading system for KOA using X-ray imaging and transfer learning techniques, with the goal of assisting clinicians and medical trainees in early and precise diagnosis. Methods: An experimental cross-sectional study was conducted using 301 radiographs (602 knee images) collected from the Social Security Teaching Hospital, Lahore. The dataset included Kellgren–Lawrence (KL) grades 0–4, with labeling based on pain observation and expert orthopedic assessment. Image preprocessing involved binary thresholding, morphological operations, knee isolation, normalization, and zero-padding. Transfer learning with DenseNet-121 served as the base network, augmented by convolutional and fully connected layers. Performance was evaluated against other deep learning architectures (DenseNet201, ResNet50, MobileNet) and classical machine learning algorithms (SVM, decision tree, random forest). Metrics included accuracy, area under the curve (AUC), precision, and recall. Results: DenseNet-121 demonstrated the most robust performance among the tested models, achieving an accuracy of 68.85%, an AUC of 85.67%, a precision of 68.33%, and a recall of 67.21% on the independent test set. Comparative models, including DenseNet201 and MobileNet, exhibited lower accuracies (≈60–61%) and AUCs (≈80–83%). Machine learning approaches underperformed, with a maximum accuracy of 55.73%. The primary challenges included dataset imbalance and the difficulty in distinguishing between grade 0 and grade 1 due to overlapping radiographic features. Conclusion: The proposed AI-based CAD system shows promise for supporting KOA diagnosis and grading in clinical practice, particularly for training junior clinicians and radiologists. Despite limitations of dataset imbalance and restricted single-center data, transfer learning with DenseNet-121 achieved reliable performance. Future work should focus on expanding datasets to encompass diverse populations, incorporating multimodal inputs, and validating generalizability across various clinical settings. This approach highlights the growing role of AI in musculoskeletal imaging and its potential to enhance early disease detection and patient care.