Recent advances in imaging techniques and deep learning applications for early diagnosis of knee osteoarthritis: A narrative review

Literature search strategy

A comprehensive literature search was conducted in PubMed to identify studies on EKOA diagnosis, focusing on imaging techniques and deep learning (DL) applications. The search combined Medical Subject Headings (MeSH) terms (e.g., “osteoarthritis, knee,” “osteoarthritis/diagnosis, early,” “imaging, medical,” and “deep learning”) and free-text keywords (e.g., “EKOA,” “Kellgren–Lawrence grade 1–2,” “Radiograph,” “US,” “MRI,” and “CNN”) with Boolean operators (AND, OR) to ensure coverage of relevant concepts, including synonyms (e.g., “US” OR “ultrasonography”). Articles published between 2017 and 2025 were prioritized. The inclusion criteria comprised studies addressing EKOA (Kellgren–Lawrence [KL] grade 1–2 or equivalent), imaging/ DL applications in EKOA, and original research/reviews/ clinical guidelines. Exclusion criteria included non-English literature, studies on advanced KOA (KL ≥ 3)/other joint diseases, animal studies, case reports, non-peer-reviewed conference abstracts, duplicates, and studies with insufficient data.

Literature search

A total of 112 studies were identified from the initial PubMed search using MeSH terms and free-text keywords related to EKOA diagnosis, imaging techniques, and DL. Of these, 15 duplicates and 7 non-English language articles were excluded, leaving 90 studies for title/abstract screening. During the screening, 30 studies were excluded because they focused on advanced KOA (KL ≥ 3), other joint diseases, animal research, case reports, non-peer-reviewed conference abstracts, or insufficient data. The remaining 60 articles underwent critical review for adherence to inclusion criteria (EKOA definition as KL grade 1–2, imaging/DL applications, and original research/ reviews/clinical guidelines).

After assessing methodological rigor and evaluating data sufficiency, 40 studies were ultimately included in this narrative review. Specifically, 12 studies focused on traditional imaging techniques, 11 on DL applications, 5 on emerging technologies, 7 on basic pathology and definitions, and 5 on review articles and methodology research. While some studies may involve multiple categories, they were classified according to their core research direction. Figure 1 illustrates the literature selection flowchart, detailing the progression from initial search to final inclusion.

Close

Figure 1:

Literature selection flowchart. KOA: Knee osteoarthritis, KL: Kellgren–Lawrence grade.

Export to PPT

Pathological mechanism and definition of EKOA

KOA is a complex disease. Knee joint tissues, such as cartilage, subchondral bone, and others, are involved.^[8] Affected by factors like mechanical injury, it develops and has pathological changes like synovitis.^[9] Synovitis can trigger cartilage degeneration at any OA stage, which is related to KOA symptoms and grade.^[10] Chronic mild synovitis causes a vicious cycle of damage between the synovium and cartilage.^[8,9]

At present, EKOA lacks a unified and effective definition, and its radiological criteria vary. Among various KOA classification systems, the KL scale [Table 1] is the most commonly used.^[11,12] In general, EKOA is defined as KL scale grades 1 and 2.^[7]

The application and limitations of traditional imaging diagnostic techniques

Plain radiograph films, which are widely used and cost-effective, require both anteroposterior and lateral views. They can identify KOA-related changes such as osteophytes and subchondral sclerosis [Figure 2],^[11,13,14] with joint space narrowing and osteophytes being the most notable features.^[15] In addition, they can reveal cartilage calcification, dislocations, fractures, and deformities. However, at present, the correlation between radiograph features and symptoms is unclear, particularly in the early stages when there is no obvious cartilage deterioration,^[12] making it difficult to determine. Moreover, plain radiograph analysis relies on experienced doctors, resulting in subjective diagnosis and inconsistent results,^[12,16,17] while artificial intelligence (AI) may offer solutions.^[18]

Close

Figure 2:

Weight-bearing left knee and full-length standing lower-limb plain radiograph films. (a) Normal knee joint structure, with bone hyperplasia at femoral condyle and tibial plateau articular margins and mild narrowing of the knee joint space. (b) Bone hyperplasia at the patellar articular margins and slight sharpening of the tibial intercondylar eminence. (c) A full-length view of the lower limbs, confirming normal lower limb alignment, with no evidence of bone destruction or abnormal soft tissue masses in the visualized regions.

Export to PPT

Although ultrasonography (US) is sensitive, it has a low negative predictive value for radiological KOA, rendering it unsuitable for EKOA. Nevertheless, its cost-effectiveness and safety continue to drive research.^[19] Ishibashi et al. discovered that knee joint effusion might be an indicator of EKOA progression, suggesting that the US could be employed to explore its role and potential as a target.^[10] Shi et al. developed a US-based grading system, demonstrating that the US can reflect early cartilage biological features, which are essential for OA screening.^[20] Vasileva et al. investigated the associations between chemerin, resistin, inflammatory markers, and US scores in KOA, highlighting their potential role in the inflammatory process.^[21] Since knee joint tissue nutrition is derived from the lower limb arteries, poor circulation can lead to soft tissue degeneration.^[22] Wu et al. utilized color Doppler ultrasound to demonstrate that in EKOA, the hemodynamics of the main arteries remain stable, whereas the flow velocity and volume of branch vessels increase, making it a valuable tool for early screening.^[22] Takemoto et al. proposed that US-detected bone cortex blood flow signals can assist in screening magnetic resonance imaging (MRI)-detected bone marrow lesions, thereby reducing the need for unnecessary MRIs in EKOA patients [Figure 3].^[23]

Close

Figure 3:

(a-e) Color Doppler ultrasound. Examination of blood vessels within the bone marrow. The white boxes indicate regions of interest for detecting and visualizing bone marrow vascular structures via color Doppler signals.

Export to PPT

Clinically, radiography has limitations in detecting EKOA changes and soft-tissue problems.^[24] In contrast, MRI excels by being able to image all joint tissues and identify early cartilage damage, bone marrow injury, synovitis, as well as pathologies of the meniscus, ligaments, and tendons [Figure 4].^[25] With its excellent soft tissue contrast and high spatial resolution, MRI enables the clear observation of the cartilage’s shape, size, and thickness, manifesting mainly in different types of imaging. Morphological MRI, including T1, T2, and proton density-weighted imaging, can effectively assess KOA-related changes. Compositional MRI, such as T2/T2* relaxation time imaging and T1ρ relaxation time imaging, quantifies collagen, glycosaminoglycan, and water, thereby revealing cartilage biochemical and microstructural alterations. Multidimensional MRI, such as hybrid multidimensional MRI and diffusion-relaxation correlation spectroscopy, depicts initial cartilage changes at the subvoxel level and provides quantitative information about the microenvironment. Furthermore, the combination of MRI techniques with AI, especially DL, accelerates image processing, standardizes procedures, and enhances the efficiency and accuracy of cartilage analysis, which is crucial for the detection and monitoring of EKOA.^[26]

Close

Figure 4:

Left knee plain and diffusion-weighted images. (a) Bone: Local cartilage abrasion and thinning occur at the femoral condyle, tibial plateau, and patellar posterior edge, with subarticular patchy bone marrow edema shadows, and articular margin bone hyperplasia. (b) Soft tissue: Joint effusion, and the surrounding soft tissues of the left knee joint are swollen, with increased signal.

Export to PPT

Exploration of new technology

In KOA diagnostic grading, although the grayscale US is objective and comprehensive, it has limitations and can only score KOA patients with obvious degenerative changes.^[19,27] As a non-invasive, quantifiable, real-time imaging detection technology, shear wave elastography (SWE) can detect hardness changes in cartilage and surrounding tissues. Compared with compression US elastography, it is more sensitive, objective, and reproducible. SWE can assist in the early detection of soft tissue damage in KOA patients and identify early strain points, providing support for KOA-graded diagnosis [Figure 5].^[27] A study involving 910 patients demonstrated that the medial meniscus projection distance and its elasticity measured by SWE can serve as reliable OA diagnosis indicators, particularly for elderly individuals and those with a high body mass index (BMI). However, the study has limitations, including a small sample size and a lack of samples from younger age groups. Larger-scale studies are needed to verify these findings and explore the potential of SWE in evaluating other joint structures.^[28]

Close

Figure 5:

(a-d) Shear wave elastography. Musculoskeletal system examination.

Export to PPT

Fluorine-18-labeled sodium fluoride positron emission tomography (PET)/MRI, an emerging technology, uniquely enables synchronous evaluation of metabolic and structural markers in KOA. Compared with MRI alone, it can detect subtle knee joint abnormalities and demonstrates significant potential in identifying early OA metabolic changes. Studies show that higher fluorine-18-labeled sodium fluoride uptake correlates with more severe adjacent cartilage degeneration, revealing a spatial association between bone remodeling and cartilage health. In addition, PET/MRI can simultaneously assess various early metabolic and biochemical markers across all joint tissues, offering new insights into OA pathogenesis and potential therapeutic targets.^[29]

Ultrashort echo time (UTE) T2* mapping, an MRI technique using a UTE sequence, measures tissue T2* relaxation time and excels in detecting short T2 tissues (e.g., deep cartilage). Its T2* relaxation time correlates with histological cartilage degeneration, suggesting utility for monitoring early cartilage degeneration.^[30]

Two-dimensional radial T2* mapping, an advanced MRI approach, employs radial sampling for k-space data acquisition, enhancing sensitivity to short T2 signals and enabling better detection of deep cartilage signals. Compared with traditional T2 mapping, it demonstrates advantages in identifying early meniscal degeneration and exhibits a strong correlation with disease severity.^[31]

DL for KOA early diagnosis

KOA feature evolution is continuous, making grading subjective and diagnosis challenging. Automated early diagnosis methods are urgently needed. As a machine-learning subfield, DL uses multi-layer neural networks to learn complex data patterns. In medical imaging, DL can identify features undetectable by human radiologists from large datasets, demonstrating great potential in diagnosis and disease progression prediction.^[32]

DL based on radiographs

In recent years, DL has been applied to radiographic images for KOA diagnosis, with several studies demonstrating varying model performances. In 2023, Alshamrani et al. utilized sequential convolutional neural network (CNN), visual geometry group-16 (VGG-16), and ResNet-50 to predict OA from radiographs, achieving over 90% accuracy across models, with VGG-16 showing the highest accuracy (99% training, 92% test), though the study lacked focus on real-world model stability.^[33]

Also in 2023, Nasser et al.’s “Discriminative Shape-Texture Convolutional Neural Network” outperformed existing DL models in classification tasks on two large databases, achieving superior accuracy and precision, but its generalizability to all KOA patients and reliability in special cases require verification.^[34]

In 2024, Rani et al. developed a 12-layer CNN-based DL method for KOA classification and severity assessment, achieving 92.3% binary and 78.4% multiclassification accuracy (outperforming prior methods), although it had limited interpretability.^[16]

A 2025 review by Zhao et al. found that DL showed good sensitivity (over 50%) for distinguishing advanced KOA grades (KL-0 and KL-4, with a maximum of 90.3% for KL-4) but noted suboptimal sensitivity for EKOA (KL-1 and KL-2). The review did not propose improvement strategies.^[17]

DL based on US

Following advancements in DL for radiographic KOA diagnosis, US-based DL research has emerged. In 2019, Kompella et al. used Mask R-CNN to segment femoral cartilage from knee US images, achieving optimal results (dice 0.88 max and 0.80 avg) with COCO 2016 pretraining and Gaussian-filtered 3D ultrasound data from healthy volunteers,^[35] offering a fast/accurate method with limited data exploration.

In 2022, Du Toit et al. proposed a DL-based 3D US femoral cartilage segmentation method using 2D U-Net for 2D US segmentation and 3D reconstruction, achieving an average dice score of 73.1% on 200 training/50 validation images,^[36] improving efficiency over manual segmentation but constrained by a small/non-diverse dataset and reconstruction limitations.

DL based on MRI

In KOA research, MRI-based DL methods bring new clinical diagnosis ideas. In 2019, Byra et al. applied attention U-Net and transfer learning to segment knee meniscus in 3D MRIs for KOA evaluation, developing a 2D attention U-Net with high dice scores and a strong correlation between automatic/ manual segmentation using 61 subjects’ images;^[37] however, the study faced limitations such as a small dataset, 3D transfer-learning challenges, and subjective image selection.

In 2022, Panfilov et al. proposed a DL-based knee MRI automatic segmentation method combining DL segmentation and non-rigid registration, which accurately extracts cartilage features (r > 0.93, volume difference < 116.00 mm³) and relates to OA progression,^[38] through its generalizability due to equipment reliance needs verification, helping with diagnosis and treatment planning.

In 2024, Panwar et al. found ResNet50 the most accurate among CNN, AlexNet, ResNet34, and ResNet50, and after integrating with DeepStack, achieved 99.71% accuracy,^[24] improving prediction but increasing complexity.

Also in 2024, Alyami used an integrated EfficientNet-B3 and ResNext-101 model on the osteoarthritis initiative dataset to predict KOA grade with 93.00% validation accuracy,^[39] yet its coverage of all patient types is uncertain, aiding in progression prediction and prevention.

Limitations of DL

DL has significant drawbacks. First, it requires large datasets for training, and insufficient data leads to poor generalization.^[40] Second, training demands substantial computing resources such as multicore CPUs and GPUs, which are expensive and burdensome for many. In addition, DL models are domain-specific, limiting their application scenarios and direct use in other data or applications.^[32] Finally, the complex and opaque internal operations of AI algorithms reduce interpretability, and enhancing DL model transparency remains a major challenge.^[25]

Recommendations

To advance EKOA diagnosis, several actionable steps are recommended. First, it is crucial to standardize EKOA diagnostic criteria. Instead of relying solely on KL grading, integrating pathological markers, such as synovitis and cartilage degeneration, with multi-modality imaging parameters is necessary. For example, combining MRI-derived cartilage thickness (e.g., T2 mapping) and SWE-measured tissue elasticity can help create a comprehensive grading system.

Second, developing large, diverse multi-modality datasets is essential. This involves prospectively collecting multi-modality data, including radiographs, US, MRI, SWE, and PET/MRI, from a wide range of populations with varying ages, BMIs, sexes, and ethnicities. In addition, ensuring that these datasets contain detailed clinical annotations, such as symptom severity and comorbidities, can enhance the generalizability of DL models and reduce bias.

Another important step is to enhance DL model interpretability. This can be achieved by mandating the use of explainability techniques, such as SHapley Additive exPlanations and gradient-weighted class activation mapping, in clinical DL applications. These techniques can clarify how models analyze imaging data, for instance, by highlighting osteophyte regions in radiographs, thereby increasing trust among clinicians and patients.

Finally, promoting the integration of multi-modality imaging and DL is vital. Investing in prospective studies to validate combined approaches, such as integrating MRI’s structural detail with PET/MRI’s metabolic data or SWE’s mechanical properties, should be a priority. For example, using DL to correlate MRI-detected cartilage defects with PET/MRI-derived bone remodeling markers can enable a holistic assessment of early OA pathogenesis. By taking these steps to address existing gaps, the field can bridge the divide between research and clinical practice, and leverage EKOA’s critical therapeutic window to slow disease progression and reduce the overall disease burden.