Abstract / Summary
Abstract Background/purpose Interstitial lung disease (ILD) is an important factor determining the course of systemic sclerosis (SSc). Quantitative high-resolution computed tomography (HRCT) radiomics can aid in the detection of ILD, particularly when pulmonary function tests cannot be performed. The aim of this study is to develop machine learning (ML)-based models using only radiomic, only clinical, and radiomic + clinical combinations for the detection of ILD in SSc and to compare their performance. Materials and methods This retrospective, single-center study included 67 SSc patients (38 ILD positive, 29 ILD negative). A total of 852 radiomic features were extracted from HRCT using whole-lung segmentation, and after feature selection, the performance of logistic regression (LR) models was evaluated using fivefold cross-validation and out-of-fold (OOF) probabilities to calculate AUC, accuracy, and F1-score (radiomic-only, clinical-only, radiomic+clinical); thresholds were determined using the Youden index. Calibration and decision curve analysis (DCA) were performed, and binary AUC differences were tested using the De Long method. Results The radiomics-only model achieved an OOF AUC of 0.819, accuracy of 0.821, and F1-score of 0.838. The clinical-only model achieved an OOF AUC of 0.760, whereas the combined model achieved an OOF AUC of 0.825. The DeLong test showed no significant AUC differences between the models (all p > 0.05). Decision curve analysis demonstrated positive net benefit for all strategies at low to moderate threshold probabilities, with largely overlapping curves. Conclusion HRCT radiomic features demonstrated good discrimination for detecting ILD in SSc, achieving numerically higher performance compared with clinical variables alone; however, these differences were not statistically significant according to DeLong testing. Adding clinical variables to the radiomic model slightly increased sensitivity without a significant improvement in AUC. HRCT-based radiomics may serve as a useful imaging biomarker for ILD detection. Although pulmonary function data were unavailable in the present cohort, the observed results suggest that HRCT-derived radiomics may provide complementary diagnostic information.