Development and validation of a predictive model for adherent perirenal fat based on CT radiomics and deep learning.
Authors
Affiliations (10)
Affiliations (10)
- Department of Radiology, First Hospital of Jilin University, Changchun, 130021, China.
- Department of Urology, First Hospital of Jilin University, Changchun, 130021, China.
- Department of Urology, First Hospital of Siping City, Siping, 136099, China.
- Department of Urology, Central Hospital of Liaoyuan City, Liaoyuan, 136200, China.
- Department of Urology, People's Hospital of Dongfeng County, Dongfeng, 136399, Jilin, China.
- Department of Clinical Research, First Hospital of Jilin University, Changchun, 130021, China.
- Imaging Institute, Cleveland Clinic, Cleveland, 44195, USA.
- Urology Unit, Azienda Ospedaliera Universitaria Integrata (AOUI) Verona, University of Verona, Verona, 37126, Italy.
- Department of Urology, First Hospital of Jilin University, Changchun, 130021, China. [email protected].
- Department of Clinical Research, First Hospital of Jilin University, Changchun, 130021, China. [email protected].
Abstract
The study aimed to develop and validate a predictive model for preoperative APF using computed tomography (CT) radiomics combined with deep learning, and validating the performance of the model in an independent cohort. The APF was defined as the need for a renal subcapsular dissection to separate the renal tumor for surgical procedures. This multicenter study collected data from renal tumor patients managed at tertiary institutions. The dataset was randomly split into a cross-validation group (70%) and an in-house test set (30%). Additionally, prospective data from three other tertiary hospitals were used as an external test set. A 3D-UNet deep learning model was employed for kidney segmentation. Intra-observer and inter-observer reproducibility was evaluated using the Dice similarity coefficient. Logistic regression models discriminated between APF and non-APF. This study enrolled a total of 460 patients, with 291 in the cross-validation set, 126 in the internal testing set, and 43 in the external testing set. The Dice similarity coefficients for both intra-observer and inter-observer assessments exceeded 0.80. Twenty-eight radiomics features and three clinical features were selected for modeling. The optimal model achieved an area under the curve of 0.95 (95% CI, 0.94-0.97) in the training set, 0.95 (95% CI, 0.92-0.97) in the validation set,0.872 (95% CI, 0.807-0.937) in the internal test set, and 0.805 (95% CI,0.676-0.935) in the external test set. CT radiomics combined with deep learning shows promise in predicting APF in surgical candidates affected with renal neoplasms.