A deep transfer learning radiomics nomogram using chest CT for differentiating anterior mediastinal cysts from...

The objective of this study was to develop a nomogram by integrating clinical features, radiomics predictive results, and deep transfer learning (DTL) predictive results to differentiate between anterior mediastinal cysts (AMCs) and low-grade thymomas.

The training set included 149 cases of AMCs and low-grade thymomas, whereas the test set included 49 cases of AMCs and low-grade thymomas. Contrast-enhanced chest CT images were used for analysis. A deep transfer learning radiomics (DTLR) nomogram was developed by integrating selected clinical features, radiomics predictive results, and DTL predictive results. Receiver operating characteristic (ROC) curves and decision curve analysis (DCA) curves were subsequently plotted.

This study identified clinical features, including maximum diameter, primary site, and relation to surroundings, to construct a clinical model. The DTLR nomogram demonstrated optimal predictive performance in the test set, achieving an area under the receiver operating characteristic curve (AUC) of 0.965, an accuracy of 0.857, a sensitivity of 0.895, and a specificity of 0.833. DCA demonstrated that DTLR is not optimal and is not significantly different from DTL. The DeLong test confirmed statistically significant differences in predictive performance between the DTLR nomogram and both the clinical model and the radiomics model, whereas no significant intermodal differences were observed among the other comparative approaches. DCA indicated that the predictive value of the DTLR nomogram was comparable to that of the DTL model. However, the DTLR nomogram demonstrated superior performance in the DeLong test, suggesting its potential as an effective clinical decision-support tool.

Although further validation is needed before clinical implementation, the DTLR nomogram demonstrated favorable predictive performance and showed promise as a practical tool for clinical decision-making.