A radiology-based differential diagnostic model for pelvic chondrosarcoma using random forest algorithms

To develop a non-invasive diagnostic method for pelvic chondrosarcoma using clinical and radiological features, aiming to improve early diagnostic accuracy and reduce reliance on invasive biopsies.

Two hundred and thirty-eight patients (39.50% with pelvic chondrosarcoma; median age, 42.87 years; 53.78% male; all Chinese) were enrolled and assigned to training (n = 167) and testing (n = 71) cohorts. Seven clinical and radiological features were evaluated for their potential associations with pelvic chondrosarcoma diagnosis. A predictive model was developed using a random forest algorithm and validated in the testing cohort, with performance assessed by partial dependence plot analysis and additional accuracy metrics.

The likelihood of a positive diagnosis of pelvic chondrosarcoma was significantly associated with age around 50 years (p < 0.001), the presence of high-intensity signals on T2-weighted magnetic resonance imaging (p = 0.021), a ring-and-arc enhancement pattern on contrast-enhanced T1-weighted magnetic resonance imaging (p < 0.001), and intratumoral calcification (p < 0.001). Tumor location was also a critical determinant (p = 0.009), with acetabular tumors exhibiting higher diagnostic probability. Among all variables, the random forest model identified the ring-and-arc enhancement pattern and patient age at diagnosis as the two most influential predictors. The model achieved excellent diagnostic performance, with a sensitivity of 96.55%, specificity of 90.48%, and overall accuracy of 92.96%.

Our random forest-based model provides a reliable and practical non-invasive diagnostic tool for pelvic chondrosarcoma, improving diagnostic accuracy while potentially reducing the reliance on invasive biopsy procedures and supporting clinical decision-making.