CT radiomics with transfer learning features for detecting DECT−positive periarticular monosodium urate crystal deposition: a single−center retrospective study

Objective To develop and validate single−energy CT (135 kVp)−based radiomics and deep learning models for the non−invasive detection of periarticular monosodium urate (MSU) crystal deposition. Methods This retrospective study included 605 patients with suspected periarticular MSU deposition, randomly split into a training cohort (n=425) and a validation cohort (n=180). Clinical variables and CT values were collected. Hand−crafted radiomics features were extracted from lesion ROIs and selected using t−test, Pearson correlation, LASSO, and mRMR. Deep features were derived from the maximum cross−sectional ROI using a ResNet50 transfer learning framework, and fused features underwent the same selection. A multilayer perceptron was used to construct the radiomics, deep learning radiomics (DLR), and combined (clinical + DLR) models, as well as a clinical−only model. Results Serum uric acid (OR 1.003), age (OR 1.017), bone erosion (OR 3.476), and CT value (OR 0.993) were independently associated with MSU deposition (all P < 0.05). AUCs for the validation cohort were 0.820 (clinical), 0.912 (radiomics), 0.940 (DLR), and 0.942 (combined). The combined model achieved accuracy 0.889, sensitivity 0.905, and specificity 0.837. The DLR and combined models significantly outperformed the clinical model (DeLong P < 0.05), with no significant difference between them. Conclusions The single−energy CT−based radiomics and deep learning radiomics models showed comparable performance for identifying periarticular MSU deposition (no statistically significant difference). The combined clinical−imaging model achieved numerically higher performance but did not significantly outperform the deep learning radiomics model.