A machine learning-based radiomics approach for differentiating patellofemoral osteoarthritis from non-patellofemoral osteoarthritis using Q-Dixon MRI

Abstract

This prospective diagnostic study aimed to assess the utility of machine learning-based quadriceps fat pad (QFP) radiomics in distinguishing patellofemoral osteoarthritis (PFOA) from non-PFOA using Q-Dixon MRI in patients presenting with anterior knee pain. This diagnostic accuracy study retrospectively analyzed data from 215 patients (mean age: 54.2 ± 11.3 years; 113 women). Three predictive models were evaluated: a proton density-weighted image model, a fat fraction model, and a merged model. Feature selection was conducted using analysis of variance, and logistic regression was applied for classification. Data were collected from training, internal, and external test cohorts. Radiomics features were extracted from Q-Dixon MRI sequences to distinguish PFOA from non-PFOA. The diagnostic performance of the three models was compared using the area under the curve (AUC) values analyzed with the Delong test. In the training set (109 patients) and internal test set (73 patients), the merged model exhibited optimal performance, with AUCs of 0.836 [95% confidence interval (CI): 0.762-0.910] and 0.826 (95% CI: 0.722-0.929), respectively. In the external test set (33 patients), the model achieved an AUC of 0.885 (95% CI: 0.768-1.000), with sensitivity and specificity values of 0.833 and 0.933, respectively (p < 0.001). Fat fraction features exhibited a stronger predictive value than shape-related features. Machine learning-based QFP radiomics using Q-Dixon MRI accurately distinguishes PFOA from non-PFOA, providing a non-invasive diagnostic approach for patients with anterior knee pain.

Keywords: Q-Dixon MRI; anterior knee pain; fat fraction; machine learning; patellofemoral osteoarthritis; quadriceps fat pad; radiomics.

Figure 1

Patient selection flowchart.

Figure 2

Comparative sagittal MRI scans of patients with and without patellofemoral osteoarthritis (PFOA). Sagittal MRI scans illustrate quadriceps fat pad (QFP) differences between a 49-year-old male without PFOA and a 50-year-old male with severe PFOA. (A) Proton density-weighted turbo spin-echo imaging (PDWI) of the PFOA patient without PFOA shows uniform hypointensity in QFP. (B) T1-weighted imaging (T1WI) of the patient without PFOA shows uniform fat-like intensity in QFP. (C) Fat fraction (FF) mapping of the patient without PFOA showing high-fat content in QFP. (D) PDWI of the patient with PFOA showing an elevated signal in QFP. (E) T1WI of the patient with PFOA shows hypointensity in QFP. (F) FF mapping of the patient with PFOA shows reduced fat content in QFP.

Figure 3

Image-processing pipeline.

Figure 4

Performance of the PD model (A), the FF model (B), and the merged model (C) and their feature contribution of performance metrics generated using feature explorer software. Receiver operating characteristic (ROC) curves illustrate the model's performance in the training and internal test sets. The bar chart shows the contribution of selected features to each model's performance.