A computed tomography-based deep learning model for non-invasively predicting World Health Organization (WHO)/International Society of Urological Pathology (ISUP) pathological grades of clear cell renal cell carcinoma (ccRCC): a multicenter cohort study.

Abstract

Background: Clear cell renal cell carcinoma (ccRCC) is the most common and aggressive subtype of kidney cancer, commonly exhibiting significant morphological heterogeneity in its pathological characteristics. The objective of this study is to develop a deep learning (DL) model for predicting pathological grades of ccRCC based on contrast-enhanced computed tomography (CECT).

Methods: Retrospective data were collected from 483 ccRCC patients across three medical centers. Arterial phase and portal venous phase computed tomography (CT) images from the dataset were segmented for renal tumors and kidneys. Three convolutional neural networks (CNNs) were employed to extract features from the regions of interest (ROIs) in the CT images across multiple dimensions including three-dimensional (3D), two-and-a-half-dimensional (2.5D), and two-dimensional (2D). Least absolute shrinkage and selection (LASSO) regression was used for feature selection. The models were evaluated using receiver operating characteristic (ROC) curves and decision curve analysis (DCA).

Results: Two types of 2.5D tumor DL models based on ResNet-34 and ShuffleNet_v2 were selected, both had area under the curves (AUCs) greater than 0.72 in the training set as well as in the internal and external test sets. The best model, resulting from the fusion of tumor and kidney models, achieved an AUC of 0.777 (95% confidence interval: 0.704-0.839, P<0.001) in the total test set, showing improved predictive ability compared to the tumor-alone models. DCA demonstrated the clinical utility of the model.

Conclusions: The DL model based on CT achieved satisfactory results in predicting the pathological grades of ccRCC.