CT-based radiomics and nnU-Net deep features for predicting calcaneal spurs: Development and validation of a combined...

This study aimed to develop and validate a computed tomography (CT)-based model that combines radiomics features, deep features derived from nnU-Net, and clinical imaging variables for predicting the presence of calcaneal spurs. In this retrospective study, 803 calcaneal CT images were included and randomly divided into a training cohort (n = 642) and a test cohort (n = 161). The calcaneus was segmented with an iterative semiautomatic no new U-Net (nnU-Net) workflow, and conventional radiomics features and nnU-Net deep features were extracted from the reviewed regions of interest. Stable and discriminative features were selected using the intraclass correlation coefficient, between-group testing, and least absolute shrinkage and selection operator regression. Support vector machine models with ten-fold cross-validation and grid search were used to generate the radiomics score (Rad-score) and nnU-Net score. Logistic regression was used to develop the clinical and combined models. Model performance was assessed with receiver operating characteristic curves (ROC), calibration analysis, decision curve analysis, and interpretability analysis. Age and Gissane's angle were retained in the clinical model. The combined model included the Rad-score, nnU-Net score, and Gissane's angle. In the training cohort, the combined model achieved an area under the ROC curve of 0.916 (95% confidence interval: 0.894-0.937), exceeding the nnU-Net, radiomics, and clinical models. In the test cohort, the combined model also showed the highest area under the ROC curve (0.772; 95% confidence interval: 0.700-0.843), compared with the nnU-Net model (0.705), radiomics model (0.710), and clinical model (0.691). Calibration analysis showed acceptable agreement between predicted and observed probabilities, with Hosmer-Lemeshow P values of .218 and .463 in the training and test cohorts, respectively. Decision curve analysis suggested a higher net benefit for the combined model across clinically relevant threshold ranges. A CT-based combined model integrating the Rad-score, nnU-Net score, and Gissane's angle improved the prediction of calcaneal spur presence compared with clinical or imaging models alone. This model may provide a quantitative imaging tool for risk stratification of calcaneal spurs in patients undergoing foot or ankle CT.