Osteoarthritis in the Knee Joint: Detection and Prediction Using X-ray Image Analysis

Osteoarthritis (OA) of the knee is a chronic degenerative condition affecting synovial joints and is a leading cause of pain and disability globally. In 2019, an estimated 528 million individuals worldwide were living with osteoarthritis, representing a striking 113% increase compared to 1990 (1). Among those affected, approximately 73% were aged 55 years or older, and women accounted for nearly 60% of the cases (1). The knee joint is the most commonly involved site, with an estimated prevalence of 365 million cases globally, followed by the hip and hand (2). Of these, nearly 344 million individuals experience moderate to severe forms of the disease that could significantly benefit from rehabilitation interventions (3). The global burden of osteoarthritis is projected to rise further in the coming decades, driven by aging populations, increasing obesity rates, and higher incidence of joint injuries. Importantly, osteoarthritis should not be regarded as an inevitable outcome of aging but rather as a complex condition influenced by multiple risk factors that can be modified to reduce its impact. The disease is characterized by progressive deterioration of articular cartilage, formation of osteophytes, subchondral sclerosis, and joint space narrowing (JSN). These structural changes compromise joint function, leading to impaired mobility and reduced quality of life, particularly among older adults and individuals with obesity. According to the World Health Organization, more than 250 million people worldwide suffer from OA, with knee OA accounting for a substantial proportion of cases. Radiographic imaging, particularly X-rays, remains the primary diagnostic modality for assessing knee OA. The Kellgren–Lawrence (KL) grading system is the most widely adopted method for categorizing OA severity, ranging from Grade 0 (normal) to Grade 4 (advanced disease). Although widely used, manual interpretation of radiographs presents inherent limitations, including inter-observer variability and subjectivity, which may lead to diagnostic inconsistencies. Additionally, measuring joint space width (JSW) a critical indicator of cartilage loss manually is both time-intensive and prone to error (4). Artificial intelligence (AI) offers promising opportunities to overcome these limitations through automated and objective analysis of radiographic images (5). AI-based models, particularly convolutional neural networks (CNNs), have demonstrated high accuracy in medical image interpretation by learning complex patterns from large datasets. Automated JSW measurement can improve diagnostic reproducibility and facilitate early detection, which is essential for delaying disease progression and reducing the need for invasive treatments like total knee replacement. Despite these advancements, there is limited evidence on how AI compares with traditional manual methods for OA detection in real-world clinical settings. This study aims to fill this gap by comparing AI-based JSW measurements with manual measurements, evaluating their predictive ability for OA severity, and analyzing associations between radiographic severity and patient-reported pain levels.

MATERIALS AND METHODS

Study Design and Setting

This was a cross-sectional analytical study conducted at SCPM Hospital, Gonda. The study included patients diagnosed with knee OA who had undergone standard anteroposterior knee radiographs.

Study Population

A total of 39 patients aged 40 years and above with confirmed radiographic knee OA (KL grades 1–4) were enrolled. Patients with a history of knee surgery, trauma-induced arthritis, or rheumatoid arthritis were excluded.

Data Collection and Image Analysis

Standard knee radiographs were assessed both manually by experienced radiologists and through AI-based automated analysis. Manual JSW was measured using standard protocols, while AI-based JSW measurement utilized convolutional neural network (CNN) algorithms combined with preprocessing steps such as noise reduction, contrast enhancement, and region-of-interest segmentation.

Variables Assessed

• Independent variables: AI-based JSW, osteophyte presence, subchondral sclerosis.
• Dependent variables: OA severity (KL grade), pain score (1–10).

Statistical Analysis:

Statistical analysis was performed using SPSS software. The Paired Samples T-Test compared manual and AI-based JSW measurements. Correlation between KL grade and pain was analyzed using Spearman’s rank correlation. ROC analysis assessed the predictive ability of JSW_AI for OA severity. Ordinal logistic regression evaluated the role of age, gender, and AI-based features in predicting OA severity. A p-value <0.05 was considered significant.

RESULTS

Demographics and Baseline Characteristics

The study included 22 females (56.4%) and 17 males (43.6%), with a mean age of 57.8 ± 10.1 years. The majority of participants were in KL Grades 2 and 3 (moderate and severe OA), accounting for 61.6% of cases.

Table 1 Frequency Distribution of KL Grade (OA Severity)

Radiographic Features

Osteophytes were present in 64.1% of participants, and subchondral sclerosis was observed in 51.3%. The mean manual JSW was 3.249 mm (SD = 1.183), while the AI-based JSW was slightly lower at 3.079 mm (SD = 1.2327).

Table 2 Frequency Distribution of Osteophytes Presence

Comparison of AI vs. Manual Measurements

The Paired Samples T-Test revealed no significant difference between manual and AI-based JSW measurements (p = 0.507), indicating strong agreement between the two methods.

Table 3 Paired Samples Statistics for JSW (Manual vs. AI Measurements)

Association Between KL Grade and Pain

Pain scores ranged from 1 to 10 (mean = 4.97 ± 2.65). Spearman’s correlation showed a weak and non-significant relationship between KL grade and pain (r = 0.089, p = 0.588), suggesting that radiographic severity does not directly predict pain intensity.

Table 4 Spearman’s Correlation Between KL Grade and Pain Level

Predictive Analysis

ROC analysis for JSW_AI in predicting OA severity yielded an AUC of 0.517, indicating poor predictive capability when used as a standalone marker. Ordinal logistic regression found that age, gender, and AI-measured JSW were not significant predictors of OA severity (Nagelkerke R² = 0.117).

Figure