Recognition of Knee Osteoarthritis Using Deep Learning: A Review ^†

Abstract

Knee osteoarthritis is one of the most common disorders and afflicts millions of patients, particularly in older age groups. The degenerative joint disease significantly compromises the quality of life through disability. We explore the various deep learning and machine learning techniques to classify knee osteoarthritis using convolutional neural networks. We examined the validity and limitations of the recent studies with multivariate classification of knee osteoarthritis using magnetic resonance imaging and X-ray data. Diagnosis accuracy improves with machine learning techniques, and transfer learning in particular leads to better diagnosis and earlier detection, which subsequently yields better patient outcomes. There are challenges to be addressed, such as dataset bias and model interpretability, which need to be further investigated for more promising results.

1. Introduction

Knee osteoarthritis (KOA) is a widespread degenerative joint disease that impacts millions of individuals globally, with older age groups being most vulnerable [1]. KOA is defined as the slow deterioration of the cartilage of the knee joint, leading to pain, stiffness, decreased mobility, and a reduced quality of life for those who suffer from it [2]. According to the World Health Organization (WHO), osteoporosis affects approximately 300 million people worldwide, and rheumatoid arthritis is one of the leading causes of disability [3,4]. KOA is becoming common, particularly in people aged 50 and over, emphasizing how, in order to be effective, prompt diagnostic techniques need to facilitate timely diagnosis and management [5,6]. Based on the Kellgren–Lawrence (KL) grader system, KOA is classified into five categories: KL 0 (no osteoarthritis), KL 1 (doubtful), KL 2 (mild), KL 3 (moderate), and KL 4 (severe) [2]. KOA, however, may be asymptomatic in the early stages, making early detection more difficult [7]. Knee compartments and X-ray variations in osteophyte scores of osteoarthritis of the knee are depicted in Figure 1.

Medical imaging plays a vital role in diagnosing knee osteoarthritis (KOA) as it provides valuable information on changes in joint structure [8]. Commonly used imaging techniques, such as magnetic resonance imaging (MRI) and X-rays, are frequently employed to evaluate the severity of KOA [1]. Among these, X-rays remain the essential standard for KOA diagnosis due to their wide availability, low cost, and ability to reveal critical indicators such as osteophyte formation [3]. However, these methods highlight the need for advanced techniques that can enhance diagnostic capabilities and accuracy, particularly given the subjectivity of image interpretation, which may affect reliability [9].

Figure 1. X-ray examination of osteophytes and compartments allows for knee osteoarthritis grading [10].

Figure 1. X-ray examination of osteophytes and compartments allows for knee osteoarthritis grading [10].

Deep learning (DL), a branch of artificial intelligence, has emerged as a transformative tool in medical image analysis, enabling differentiation between healthy joints and KOA [11]. DL models can automatically learn and extract hierarchical features from medical images using convolutional neural networks (CNNs). This approach improves classification accuracy and reduces the potential for observer bias [12]. Applications of deep learning to X-ray images have been shown to enhance diagnostic efficiency, support KOA detection, and enable earlier recognition of the disease—critical for timely intervention and improved patient outcomes [13].

The remainder of this article is organized as follows. Section 2 discusses deep learning approaches for KOA recognition. Section 3 examines the datasets used in KOA research. Section 4 presents the results and discussion, while Section 5 addresses challenges and future directions. Finally, Section 6 provides the conclusions of the review.

2. DL for KOA Recognition

2.1. Overview of DL for KOA Recognition

Recent advances in deep learning techniques have improved the diagnostic accuracy of KOA. The EfficientNet B5 transfer learning algorithm, which relies on features learned from big data sets, has been used for classifying KOA at early stages. This approach has been reported to boost classification performance for the varying degrees of osteoarthritis levels [14]. Improved designs, such as the adversarial neural network (A-ENN), have enabled the longitudinal evaluation of KOA severity through x-ray imaging for accurate quantification of disease over time. Such advances underscore the significance of deep learning in enhancing diagnostic expertise, which ultimately provides better patient care and outcomes [15].

2.2. Types of Medical Imaging Methods

• X-ray imaging

Knee X-rays represent one of the fundamental techniques for assessing the severity of osteoarthritis and enable the physician to categorize the severity as ‘healthy’, ‘moderate’, and ‘severe’ stages [8]. This procedure offers important information on osteoarthritis, including subchondral sclerosis, development of osteophytes, and narrowing of the joint space [16]. X-ray diagnostic imaging is crucial for the patients as it helps to assist in coming up with prevention and intervention strategies on time. According to other researchers [8,17], this feature makes X-ray imaging common in clinical practice as it is easy to obtain, and that is why it has become part of the standard routine in the management of osteoarthritis (Figure 2) [11].

2.

MRI

Osteoarthritis of the knee joint is best diagnosed using magnetic resonance imaging (MRI). X-rays alone cannot provide an accurate view of the joint’s soft tissues, such as cartilage, menisci, and ligaments. MRI offers a superior option because it can capture detailed images of these structures [18,19]. As a result, diagnosis becomes clearer, with subtle changes that may be missed on X-ray being readily visible through MRI [20]. In addition, advanced techniques such as T2 mapping and dGEMRIC allow MRI to assess cartilage sensitivity [21]. This capability enhances patient care by supporting the development of targeted lifestyle recommendations and guiding subsequent therapies, thereby improving outcomes for individuals living with KOA [3]. MRI images of the knee joint, which reveal various structural features, are particularly valuable in determining cartilage status and identifying osteoarthritis [20].

3.

Computed tomography (CT)

CT scans generate detailed cross-sectional images of the knee joint, enabling thorough assessment of soft and bone tissues individually. This imaging modality is helpful in difficult scenarios where additional details are required beyond the source of MRI and X-rays [21]. To determine the extent of degenerative alteration and detect minor bone changes associated with KOA, CT can be utilized for bone marrow edema and cystic changes [21]. CT is beneficial for diagnostics and treatment planning in KOA since it provides high-resolution images that could also assist in surgical planning and postoperative assessments [22].

2.3. DL Architecture

CNNs are a specialized deep learning architecture increasingly applied to the diagnosis and identification of KOA. Because CNNs automatically learn hierarchical patterns of features from visual inputs, they are highly effective for interpreting medical images. As a result, CNNs are particularly well-suited for recognizing fine-grained features in knee X-ray images that reflect structural changes associated with osteoarthritis [12]. CNN comprises the following layers.

• Convolutional layers: These layers generate feature maps by applying filters to input images, highlighting important patterns necessary for classification. By reducing the spatial dimensions of the image while preserving key features, each convolution enhances diagnostic accuracy [23].
• Pooling layers: Following convolution, pooling layers downsample the feature maps, reducing spatial dimensions and improving the model’s tolerance to translations and minor distortions. This step also increases computational efficiency [24].
• Fully connected layers: In later stages, fully connected layers integrate the extracted information to perform higher-level reasoning. Activation functions, such as the Rectified Linear Unit (ReLU), enable nonlinear combinations that allow the model to learn complex relationships in the data [25].

Recent advances in CNN architectures (e.g., VGG16, ResNet, and DenseNet) have significantly improved the diagnostic performance of KOA detection, consistently outperforming traditional machine learning methods [25,26]. These innovations highlight the transformative potential of CNNs in KOA diagnosis, enabling greater accuracy and effectiveness, as illustrated in (Figure 3).

Recent advances in CNN architectures (e.g., VGG16, ResNet, and DenseNet) have improved the diagnostic performance of KOA detection, and model performances have been consistently better than traditional machine learning methods [25,26]. These advances highlight the transformative potential of CNNs in KOA diagnosis, enabling improved accuracy and effectiveness.

Recurrent neural networks (RNNs) are frequently employed in sequential data analysis and have exhibited promising performance in capturing temporal dependencies in medical evaluations. RNNs learn the patterns in the longitudinal data, which is appropriate for ongoing monitoring of how KOA is worsening or improving over time. To address this issue and assess KOA in a comprehensive way, hybrid models that utilize a CNN and an RNN have been developed to capture both temporal and spatial information [14,27]. Generative adversarial networks (GANs) address the limitations caused by insufficient datasets. GANs generate synthetic training data, improving model generalization and KOA classification performance through competitive training. Studies demonstrate that GANs can produce realistic knee X-ray images, which are particularly useful in low-data scenarios [28,29].

Transfer learning (TL) is another effective deep learning model that enhances performance with limited labeled data by leveraging knowledge from pretrained domains. TL benefits medical imaging tasks by utilizing features from large datasets, such as ImageNet, to detect KOA. TL reduces training time and improves diagnostic accuracy through knowledge transfer [8,30]. Consequently, pretrained models such as InceptionResNet and EfficientNet are widely adopted. EfficientNet is recognized for its scalable design, improving depth, width, and resolution while achieving high accuracy [9]. InceptionResNet, by integrating residual connections with Inception structures, addresses vanishing-gradient problems and enables deeper network training, further advancing KOA diagnosis [31,32].

3. Datasets

In KOA research, deep learning models for diagnosis and classification have been developed and validated using several well-established datasets.

Table 1. Datasets used in KOA research.

Reference	Application	Number of Images	Type of Image	Size of Image	N. of Classes
[15]	OAI	3294	X-ray	224 × 224	5
[20]	OAI- Zuse Institute Berlin	7200	MR	224 × 224	2
[17]	KOA	10,000	X-ray	224 × 224	5
[5]	OAI	200	X-ray	256 × 256	2
[28]	KOA (severity)	8260	X-ray	224 × 224	5
[3]	OAI	9786	X-ray	224 × 224	5
[3]	OAI	26,626,000	X-ray	224 × 224	5
[34]	OAI	4472	X-ray	224 × 224	5
[35]	OAI	4255	X-ray	224 × 224	11
[13]	OAI	4866	MRI	0.37 × 0.37 × 0.7	3
[33]	OAI	5778	X-ray	224 × 224	5
[29]	OAI	3615	X-ray	244 × 244 × 3	2
[21]	OAI	1100	MRI, X-ray	160 × 160 × 160	5
[36]	OAI	9786	X-ray	224 × 224	5
[22]	OAI	8260	X-rays	224 × 224	5
[37]	OAI	4130	X-ray	256 × 256	5
[38]	Radiograph	2579	X-ray	512 × 512	5
[39]	KOA	3836	X-ray	224 × 224	2
[10]	KOA	10,930	X-ray	224 × 224	2
[40]	OAI	1650	X-ray	224 × 224	5
[41]	KOA	4130	X-ray	2048 × 2560	5
[42]	OAI	1752	X-ray	227 × 227 × 3	2
[30]	OAI	9000	X-ray	224 × 224	5
[43]	OAI	8260	X-ray	224 × 224	5
[44]	OAI	5478	X-ray	224 × 224	2
[11]	KOA	6320	X-ray	224 × 224	5
[45]	OAI	9786	X-ray	224 × 224	5
[46]	KOA	10,000	X-ray	224 × 224	5
[47]	KOA	8260	X-ray	224 × 224	3

provides a summary of the primary datasets employed in KOA investigations. One of the largest datasets is the Osteoarthritis Initiative (OAI), which contains more than 26,626 images from both MRI and X-ray modalities, with X-ray scans typically captured at high resolution (2048 × 2560 pixels) [21]. Another important resource is the KOA dataset, comprising between 5000 and 10,000 standardized X-ray images, often scaled to 224 × 224 pixels to facilitate effective model training [8]. The Mendeley dataset, with over 1650 X-ray images, emphasizes visual features of KOA severity while maintaining image quality at approximately 162 × 300 pixels [33]. The National Health Service Digital Osteoarthritis Dataset includes data from about 2500 patients along with MRI and X-ray images, providing valuable real-world context for KOA research [9]. Additionally, the Radiological Society of North America bone age dataset offers around 12,000 high-quality radiographs that, while primarily intended for bone age assessment, also provide insights into joint health [34]. These datasets are critical for advancing machine learning techniques aimed at improving the diagnosis and classification of KOA.

Researchers have investigated a wide range of models for KOA classification, employing diverse methodologies and achieving varying levels of accuracy. For example, Bonakdari et al. [48] utilized support vector machines (SVMs) to predict KOA structural progression, reporting 80% accuracy. Khessiba et al. [49] applied BlazePose and Spatiotemporal Graph Convolutional Networks (GCNs) for gait analysis, achieving 93.75% accuracy in KOA categorization. Similarly, Yang et al. [38] implemented the RefineDet model on unprocessed knee X-ray images, attaining 95.7% accuracy and demonstrating strong potential in severity classification. These findings highlight the effectiveness of both classical and deep learning approaches, though performance varies considerably across architectures and datasets.

Machine learning and deep learning models have also been widely applied to KOA diagnosis using different data modalities. Janotheepan et al. [17] developed a CNN-based model trained on 10,000 knee X-ray images, achieving 89% accuracy through optimized preprocessing and hyperparameter tuning. Singh et al. [50] employed the OAI dataset to train a deep CNN, reporting 98% accuracy and improving early detection and treatment planning. Dharmani et al. [31] applied EfficientNetB1 with transfer learning on localized Indian X-ray datasets, achieving 89% accuracy and underscoring the importance of region-specific data in enhancing diagnostic performance.

The results illustrate the diversity of methodologies in KOA classification, differing in data modality, model architecture, and diagnostic accuracy. Ahmed et al. [28] investigated interpretable deep learning models using X-ray images, combining EfficientNetB7 with gradient-weighted class activation mapping (GradCAM) to improve transparency. Their approach achieved 99.13% accuracy in distinguishing normal from severe KOA, with GradCAM providing valuable visual cues that enhanced interpretability, emphasizing the role of explainable AI (XAI) in medical image classification. Parikh et al. [44] employed DenseNet-201 with transfer learning, achieving 82.48% accuracy in early KOA detection and demonstrating improved efficiency compared to manual methods. Sinha et al. [9] explored EfficientNet-B5 and DenseNet121, showing that deep neural networks outperform traditional diagnostic techniques in recognizing and grading KOA severity. Building on this, Navyeesh et al. [33] proposed an ensemble model combining DenseNet and EfficientNet, which achieved 94.76% accuracy, reinforcing the effectiveness of integrated deep learning models for early diagnosis and clinical decision-making.

Challa et al. [47] conducted a comparative study using MobileNet, Visual Geometry Group 16 (VGG16), EfficientNet Version 2 Large (EfficientNetv2l), and EfficientNetV2L to classify KOA from X-ray images. EfficientNetV2L outperformed the others, achieving 93.96% accuracy and demonstrating the capability of advanced CNN architectures in severity assessment. Kishore et al. [11] developed a computer-aided diagnosis (CAD) tool based on MobileNetV2, integrated into a web application using Flask. The tool analyzed radiographic images to assess KOA severity and achieved 98.7% accuracy, highlighting its potential to support healthcare professionals through early detection and timely intervention. Kumar et al. [22] examined the impact of image preprocessing using InceptionResNetV2, showing that sharpening techniques improved classification accuracy from 72% to 91.03%. Ahmed et al. [30] applied ResNet50 with transfer learning and data augmentation to address class imbalance across five severity levels, achieving 97.4% accuracy and demonstrating the strength of transfer learning in improving diagnostic precision with limited labeled data.

Rehman et al. [25] developed a hybrid CNN-VGG16 model to classify KOA severity stages, achieving 93.27% accuracy through data augmentation and feature integration. Marimuthu et al. [42] employed Faster RCNN with VGG16 and transfer learning to localize minimum knee joint space width (mKJSW), an important biomarker in KOA assessment, achieving 98.6% localization accuracy. Teoh et al. [35] introduced a hybrid model combining VGG16, Global Average Pooling, and K-Nearest Neighbors (KNN), achieving 89.75% accuracy, an ROC-AUC of 0.99, and Cohen’s kappa of 0.91, confirming the importance of joint center analysis in pain estimation. Sivakumari et al. [24] developed a deep learning approach using AlexNet and CNN architectures, achieving 98% accuracy in classifying KOA severity. Similarly, Chandu et al. [5] implemented a CNN framework enhanced with AlexNet, trained on a large dataset of knee radiographs, achieving 96.08% accuracy and demonstrating the potential of deep learning for automated KOA diagnosis. Singh et al. [14] employed EfficientNet-B5 with transfer learning to classify KOA into healthy, moderate, and severe categories, achieving 97% accuracy. Pandey et al. [6] applied EfficientNet-B0 to the OAI dataset, reporting 69.74% accuracy, highlighting the need for improved preprocessing. Singh et al. [8] further applied EfficientNet-B3 with transfer learning to classify KOA into five severity levels, achieving 92% accuracy, while emphasizing the importance of larger and more-diverse datasets for generalizability.

Guida et al. [21] demonstrated the advantages of MRI over X-rays by employing a 3D CNN, achieving 86.5% accuracy in distinguishing osteoarthritic knees, underscoring the benefits of volumetric data in early detection. Xin Wang et al. [20] used InceptionResNetV2 with transfer learning, fine-tuned using RMSprop, on the OAI-ZIB dataset, achieving 96.1% accuracy and outperforming previous MRI-based approaches. Panfilov et al. [13] introduced a predictive model combining a 2D CNN with a feature-pooling transformer to forecast KOA progression. Trained on 4866 knee MRI scans, the model achieved an AUC of 0.79, establishing a benchmark for MRI-based prognosis using deep learning. Several deep learning models have been proposed for knee osteoarthritis detection. A comparison of these techniques is presented in

Table 2. Comparison of DL techniques for KOA.

Reference	Dataset	Methodology	Accuracy	Result
[21]	OAI	3D CNN with MRI	86.5%	MRI improved diagnosis accuracy over X-ray
[48]	OAI	SVM	80%	Early identification of structural progressors for knee OA
[15]	OAI	A-ENN	62.7%	Beneficial to KOA grading over time
[20]	OAI-ZIB	InceptionResNetV2, transfer learning model	96.1%	Classifying knee OA effectively
[2]	OAI	Semi-supervised, active learning	64.13%	Improved severity rating using fewer samples that have been tagged
[34]	OAI	Confidence learning, CNN with hybrid loss	70.13%	Outstanding results on the OA assessment tasks
[13]	OAI	2D CNN + Transformer	pre- cision of 0.58 ± 0.03 and ROC AUC of 0.78 ± 0.01.	Accurately forecasts the course of KOA using MRI data
[37]	OAI	Local Center of Mass (LCM), Siamese CNN	93.2%	Effective identification and categorization of osteoarthritis in the knee
[14]	OAI	EfficientNet B5 Transfer Learning	97%	Excellent classification performance for knee OA
[1]	OAI	YOLOv5, OAFE, OADR blocks	78.93%	Joint extraction rate of 95.3%
[26]	OAI	Deep few-shot learning model	IoU 0.94	Effective segmentation with few annotated pictures
[6]	OAI	EfficientNet-B0	69.74%	Enhanced accuracy in classifying KOA severity
[49]	Gait dataset (KOA-PD-NM)	BlazePose, Spatiotemporal Graph Convolutional Network (STGCN)	93.75%	Improved KOA categorization based on gait patterns
[8]	KOA	EfficientNet B3, Transfer Learning model	92%	Five severity degrees are effectively classified
[3]	OAI	Ensemble of ResNet-34, VGG-19, DenseNet	98%	Better categorization for every KL grade
[51]	OAI	Improved CenterNet with DenseNet-201	99.14%	Strong identification and categorization across KOA levels
[35]	OAI	VGG16-GAP-KNN hybrid model	89.75%	Using radiographs to accurately estimate the level of discomfort
[29]	OAI	CNN, Random Forest, KNN, and 2D-CNN are CRK	99%	High predictive accuracy for diagnosing osteoarthritis
[25]	OA severity	A hybrid model that combines CNN and VGG16	93.27%	Improved identification of the phases of knee OA
[52]	OAI	ResNet-101, Deep CNN	94.24%	Enhanced feature extraction combined with efficient categorization
[42]	OAI	VGG-16, RCNN with transfer learning	98.6%	The suggested technique effectively determines the minimum width of the knee joint.
[30]	OAI	Transfer learning ResNet50	97.4%	Effective use of X-rays to categorize the degree of osteoarthritis in the knee
[44]	KOA	DenseNet-201, transfer learning	82.48%	Early detection of osteoarthritis in the knee
[28]	OAI	EfficientNetb7 with GradCAM, VGG Model, ResNet Model	99.13%	Using GradCAM insights for efficient categorization
[33]	OAI	DenseNet, EfficientNet Ensemble Learning	94.76%	Improved OA severity level prediction
[36]	OAI	Inception, ResNet50Ensemble model, Xception, and ResNetV2	68%	Enhanced KOA diagnosis using group methods
[53]	KOA	ROC curve, statistical analysis, and podoscope	AUC 0.74	Foot abnormalities are much more common in KOA patients
[54]	Osteoarthritis and Meniscus Pathologies	Radiography and MRI evaluation	84.2%	Meniscal abnormalities and the severity of OA are significantly correlated
[40]	OAI	ResNet152, CNN	91%	Assessing the severity of osteoarthritis effectively
[43]	OAI	ResNet-101, Evidential Deep Learning (EDL)	72%	EDL effectively measures the degree of ambiguity in the diagnosis of OA
[50]	OAI	Deep learning, CNN-based model	98%	Improves knee OA early identification and treatment
[45]	OAI	Ensemble of Xception, InceptionResNetV2	95.1%	Better identification and categorization of KOA severity
[46]	KOA	Machine learning (Decision Tree, KNN, Random Forest)	Random Forest: 38.35%	Early and efficient identification of osteoarthritis in the knee
[55]	KOA	Gradient Boosting Tree (GDBT)	89.29%	Accurately predicting KOA severity grades

.

4. Results and Discussions

Different methods demonstrate varying performance when applied to the OAI dataset. Figure 4 compares the accuracy of deep learning and machine learning models using OAI data, revealing discrepancies in performance. CenterNet combined with DenseNet-201 achieved the highest accuracy (99.14%), representing the strongest performance in osteoarthritis classification. EfficientNetB7 produced comparable results (99.13%) while offering the added advantage of interpretability through GradCAM. Ensemble approaches (ResNet34, VGG-19, and DenseNet) and models such as VGG-16 with RCNN and deep CNN achieved accuracies ranging from 98% to 98.6%. For scenarios with limited data, few-shot learning and LCM with Siamese CNN proved highly effective, delivering strong results despite small sample sizes. Hybrid models, including VGG16-GAP-KNN and ResNet-101 with deep CNN, also demonstrated high accuracy by integrating deep learning with traditional algorithms.

3D CNNs achieved 86.5% accuracy, while 2D CNNs performed moderately compared with state-of-the-art approaches. Traditional machine learning models, such as SVM, reached 80% accuracy but were clearly outperformed by deep learning techniques. Lower-performing models included A-ENN, Semi-Supervised Siamese, and confidence learning, with accuracies ranging from 62.7% to 70.13%. Similarly, Inception with ResNet50 Ensemble and ResNet-101 with Evidential Deep Learning achieved relatively poor results (68% and 72%), suggesting limited suitability for the complexity of OAI data.

5. Challenges and Future Directions

Despite significant progress in KOA classification using deep learning, several challenges remain. A major issue is dataset bias, which arises when training datasets lack diversity, limiting the generalizability of models across populations. Another challenge is the interpretability of complex models; for patients and physicians to trust and adopt these techniques in clinical practice, transparent explanations of predictions are essential. Future research should prioritize the development of robust datasets encompassing diverse clinical presentations and demographics. Incorporating multimodal data, such as imaging and clinical variables, enhances model applicability and performance. To build confidence among healthcare practitioners, explainable AI methodologies need to be required. Ultimately, collaboration among data scientists, healthcare providers, and patients is necessary to establish a comprehensive framework for KOA diagnosis that improves accuracy, fosters patient engagement, and strengthens trust in AI-powered treatments.

6. Conclusions

We examined how machine learning techniques, particularly CNNs, have advanced KOA diagnosis beyond traditional methods such as the KL grading system, which is limited by delayed detection and inter-rater variability. Although mainstream medical literature has been slow to adopt artificial intelligence in clinical practice, recent developments in deep learning have demonstrated substantial improvements in diagnostic efficacy and accuracy. These advances enable earlier detection and intervention within a learning paradigm, where each diagnosis contributes to model refinement and improved patient outcomes. Automated technologies streamline workflows by reducing reliance on manual examinations, minimizing human error, and delivering more consistent diagnoses. Such innovations underscore the disruptive potential of AI in KOA diagnostics, supporting physicians in refining diagnoses and optimizing treatment regimens tailored to individual patients. Sophisticated algorithms capable of analyzing large volumes of imaging data also provide deeper insights into disease progression. As the medical field continues to evolve, integrating these advanced methodologies will be crucial for improving patient care, health outcomes, and quality of life. This underscores the urgent need for further research and interdisciplinary collaboration to maximize the therapeutic applications of AI in osteoarthritis management.