The predictive value of radiomics and deep learning for synchronous distant metastasis in clear cell renal cell carcinoma.

Objective: The objective of this research was to devise and authenticate a predictive model that employs CT radiomics and deep learning methodologies for the accurate prediction of synchronous distant metastasis (SDM) in clear cell renal cell carcinoma (ccRCC).

Methods: A total of 143 ccRCC patients were included in the training cohort, and 62 ccRCC patients were included in the validation cohort. The CT images from all patients were normalized, and the tumor regions were manually segmented via ITK-SNAP software. Radiomic features were extracted via the FAE toolkit. The least absolute shrinkage and selection operator (LASSO) algorithm was employed to select features and build various machine learning models. Additionally, the largest cross-section of the tumor was cropped to train the deep learning model. Multiple deep learning models were trained to predict SDM in ccRCC patients. The results of the best machine learning model were then fused with those of the deep learning model to create a combined model.

Results: Of the 944 radiomic features identified, 15 were closely associated with SDM. With these 15 features, the support vector machine (SVM) model emerged as the most effective, demonstrating areas under the curve (AUC) of 0.860 and 0.813 in the training and validation cohort, respectively. Among the deep learning models, ResNet101 performed optimally, achieving AUC of 0.815 and 0.743 in the training and validation cohort, respectively. The combined model yielded an AUC of 0.863. Decision curve analysis suggested that the combined model offers superior clinical applicability.

Conclusion: The model integrates radiomics and deep learning, showing significant potential in predicting SDM in ccRCC patients. It holds promise for supporting clinical decision-making, reducing missed diagnoses of SDM, and guiding patients in further enhancing their systemic examinations.