Introduction & Objectives
Treatment decisions for an incidental renal mass are mostly made with pathologic uncertainty. Improving the diagnosis of benign renal masses and distinguishing aggressive cancers from indolent ones is key to better treatment selection.
Materials & Methods
We analyzed 13261 pre-operative CT volumes of 4557 patients across six medical centers and five public imaging datasets. This multi-center cohort was divided into the training set, internal test set, external test set, prospective test set and TCIA test set. Convolutional neural networks were used for segmenting kidney structures and detecting renal masses. Two multi-view convolutional neural networks based on multi-phase CT images were developed to predict the malignancy and aggressiveness of renal masses, respectively. We compared survival differences between AI-predicted indolent and AI-predicted aggressive tumors using Kaplan-Meir analyses and Cox regression analyses. Genomic, transcriptomic, and immune landscapes were evaluated using bioinformatic analyses and immunohistochemistry.
Results
The first diagnostic model predicting malignancy of renal masses reached AUCs of 0.899 in the internal test set, 0.856 in the external test set, 0.871 in the prospective test set and 0.881 in TCIA test set, outperforming the average performance of seven experienced radiologists in the prospective test set. The second diagnostic model differentiating aggressive from indolent tumors had AUCs of 0.792 in the internal test set, 0.763 in the external test set, 0.784 in the prospective test set and 0.755 in the TCIA test set, higher than the R.E.N.A.L. nephrometry nomogram. In the external test set, AI-predicted aggressive tumors had significantly worse survival than AI-predicted indolent tumors [Disease-specific survival (DSS), p<0.001, HR= 20.81; Recurrence-free survival (RFS), p<0.001, HR=9.71; Overall survival (OS), p<0.001, HR=13.27]. The AI-aggressiveness score was an adverse independent risk factor (DSS, p=0.002; RFS, p<0.001; OS, p=0.002), and had higher C-index than TNM stage and ISUP grade in predicting survival outcomes. Aggressive renal tumors associated with a heavily infiltrated but immunosuppressive tumor microenvironment with increased infiltrations of CD8+T cells (p=0.051) and Tregs (p=0.041).
Conclusions
Deep learning models can non-invasively predict the likelihood of malignant and aggressive pathology of a renal mass based on preoperative multi-phase CT images. AI-predicted aggressive renal tumors are associated with worse survival outcomes.
