Abstract
BACKGROUND: Osteoarthritis (OA) of the knee is a prevalent chronic degenerative joint condition that is having a growing impact on a global scale., posing a challenge in diagnosis which is often reliant on time-consuming and error-prone visual analysis by physicians. There is a critical need for an automated, efficient, and accurate diagnostic method to improve early detection and treatment. METHODS: We developed a novel Convolutional Neural Network (CNN) module, Dense Multi-Scale (DMS), an advancement over Multi-Scale Convolution (MSC). This module utilizes dense connections in convolutions of varying sizes (1 × 1, 3 × 3, 5 × 5) and across layers, enhancing feature reuse and complexity recognition, thereby improving recognition capabilities. Dense connections also facilitate deeper network architecture and mitigate gradient vanishing problems. We compared our model with a standard baseline model and validated it using an unseen-data test set. RESULTS: The DMS model exhibited exceptional performance in unseen-data tests, achieving 73.00% average accuracy (ACC) and 92.73% area under the curve (AUC), surpassing the baseline model's (DenseNet) 63.52% ACC and 88.76% AUC. This highlights the DMS model's superior predictive capability for knee OA. CONCLUSION: The DMS model presents a significant advancement in predicting and grading knee OA, holding substantial clinical importance. It promises to aid radiologists in accurate diagnosis and grading, and in choosing appropriate treatments, thereby reducing misdiagnosis and patient burden.