Integrated radiomics and deep learning model for identifying medullary sponge kidney stones

Abstract

Background: Medullary sponge kidney (MSK) is a rare congenital anomaly frequently associated with nephrolithiasis. Accurate preoperative differentiation between MSK stones and non-MSK multiple kidney stones remains challenging, yet it is essential for effective clinical decision-making. This study aims to develop a novel diagnostic model that integrates radiomics and deep learning features to improve the differentiation of MSK stones using CT imaging.

Methods: This single-center, retrospective study included patients who underwent surgical treatment for multiple kidney stones at Beijing Tsinghua Changgung Hospital between 2021 and 2023. All MSK and non-MSK cases were confirmed via endoscopic surgery. Radiomics features were extracted from manually delineated regions of interest (ROI) on nephrographic-phase CT images, while deep learning features were derived from a ResNet101-based model. Three diagnostic signatures-Radiomics (Rad), Deep Transfer Learning (DTL), and Deep Learning Radiomics (DLR)-were developed. A Combined model was constructed by integrating clinical variables with DLR features to further enhance diagnostic accuracy. Model performance was evaluated using AUC, calibration curves, Net Reclassification Index (NRI), and Integrated Discrimination Improvement (IDI) analyses. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) visualization was employed to identify imaging regions critical to classification, improving interpretability.

Results: A total of 73 patients with multiple kidney stones were analyzed, comprising 34 MSK cases and 39 non-MSK cases, encompassing 110 kidneys in total. The DLR signature demonstrated high diagnostic accuracy, with AUCs of 0.96 in both the training and test cohorts. The Combined model further enhanced diagnostic performance, achieving AUCs of 0.98 in the training cohort and 0.95 in the test cohort. Calibration curves indicated strong agreement between predicted probabilities and observed outcomes. Furthermore, NRI and IDI analyses highlighted the superior predictive power of both the DLR and Combined models compared to other approaches.

Conclusion: This study introduces an innovative approach for MSK stone diagnosis by integrating radiomics and deep learning features. The proposed model offers high diagnostic accuracy and promising clinical utility.

Keywords: artificial intelligence; deep learning; kidney stone; medullary sponge kidney; radiomics.

Figure 1

Workflow for developing a model integrating radiomics and deep learning to differentiate MSK stones from non-MSK stones. MSK medullary sponge kidney.

Figure 2

Performance of the DLR signature for differentiating MSK stones from non-MSK stones. (A) The ROC curves of the DLR signature in the training and testing cohorts. (B) Predicted scores for individual samples generated by the DLR signature. DLR deep learning radiomics, MSK medullary sponge kidney, ROC receiver operating characteristic.

Figure 3

Performance comparison of different models (clinical, Rad, DTL, DLR, and Combined) in the training and testing cohorts. (A) ROC curves in the training cohort. (B) ROC curves in the testing cohort. (C) Calibration curves in the training cohort. (D) Calibration curves in the testing cohort. (E) DeLong test results for the training cohort. (F) DeLong test results for the testing cohort. Rad radiomics, DTL deep transfer learning, DLR deep learning radiomics, ROC receiver operating characteristic.

Figure 4

Nomogram and DCA for the combined model. (A) Nomogram for the combined model. (B) DCA for the training cohort. (C) DCA for the testing cohort. DCA decision curve analysis.