Introduction

Knee osteoarthritis (KOA) is a degenerative joint disease with a prevalence ranging from 4 to 12% [1,2,3]. Currently, for orthopaedic surgeons, knee weight-bearing standing X-ray radiographs as a standard method for evaluating KOA remain the most common radiology examination method due to their safety, popularity and low cost [4,5,6]. Currently, accurate KOA diagnosis and assessment are highly based on radiographic evidence [7, 8].

Currently, the Kellgren–Lawrence scale (K-L) is most commonly used to diagnose and determine the severity of KOA based on joint space narrowing, osteophytes, sclerosis, and definite bony deformities on X-rays in radiology examinations [9, 10]. However, the classification criteria of the K-L scales are subjective [11]. In clinical use, different doctors or the same doctor at different times may often obtain similar results rather than identical results on the same X-ray. Many clinical studies involving KOA diagnosis have ensured reliability by increasing the number of repeated diagnoses [12,13,14,15,16]. Moreover, the total numbers of X-ray examinations are much higher in large hospitals, which is a heavy burden on radiologists and surgeons. Therefore, many rapid diagnosis and assessment models have been developed in collaboration with image analysis. The models can identify images via digital processing techniques to make the artificial intelligence process more accurate and cost-effective [13]. The technology includes knee joint recognition and image processing based on deep learning [14]. Swiecicki et al. [15] developed a diagnostic model based on the two-stage Faster R-CNN model to assess the severity of KOA from both posterior–anterior (PA) and lateral (LAT) views. Tiulpin et al. [16] also developed a diagnostic model based on the ResNet34 model to detect KOA from original PA views of knees. Similarly, Norman et al. [17] developed a KOA diagnostic model based on DenseNet, which uses the feature more effectively in deep technology. Current studies show that the existing diagnostic models can achieve satisfactory accuracy [15,16,17]. However, those models rely on preprocessed, highly optimized digital images in specific software and hardware, which may not be feasible in most clinical scenarios and affects the actual use value in some ways [18,19,20,32]. Second, although the study shows that this model has good reliability and reproducibility in radiology assessment, the specific diagnosis still needs other evidence, such as the patient’s clinical manifestations, laboratory tests, and other imaging findings [12]. Third, in the process of use, due to the restrictions of the algorithm, there may be some photos in which the entire knee joints cannot be identified, and the diagnostic model does not display a diagnosis if the algorithm detects implants or no object. Therefore, there may be a subset that is not identified as implants or knee joints, which requires rephotographing or further diagnosis by surgeons. Fifth, compared to similar studies, our sample size is relatively small and from a single hospital, and outcomes from a larger cohort and multiple countries may be different. Sixth, we propose the deep-learning-based RefineDet for object detection and classification, but the outcomes may be different based on other algorithms.

Conclusion

The deep learning-based diagnostic model can be used to assess the severity of KOA in portable devices. On the premise of improving the diagnostic efficiency, the results are highly reliable and reproducible.