Deep learning with a convolutional neural network model to differentiate renal parenchymal tumors: a preliminary study

Purpose: With advancements in medical imaging, more renal tumors are detected early, but it remains a challenge for radiologists to accurately distinguish subtypes of renal parenchymal tumors. We aimed to establish a novel deep convolutional neural network (CNN) model and investigate its effect on identifying subtypes of renal parenchymal tumors in T2-weighted fat saturation sequence magnetic resonance (MR) images. Methods: This retrospective study included 199 patients with pathologically confirmed renal parenchymal tumors, including 77, 46, 34, and 42 patients with clear cell renal cell carcinoma (ccRCC), chromophobe renal cell carcinoma (chRCC), angiomyolipoma (AML), and papillary renal cell carcinoma (pRCC), respectively. All enrolled patients underwent kidney MR scans with the field strength of 1.5 Tesla (T) or 3.0 T before surgery. We selected T2-weighted fat saturation sequence images of all patients and built a deep learning model to determine the type of renal tumors. Receiver operating characteristic (ROC) curve was depicted to estimate the performance of the CNN model; the accuracy, precision, sensitivity, specificity, F₁-score, and area under the curve (AUC) were calculated. One-way analysis of variance and χ² tests of independent samples were used to analyze the variables. Results: The experimental results demonstrated that the model had a 60.4% overall accuracy, a 61.7% average accuracy, and a macro-average AUC of 0.82. The AUCs for ccRCC, chRCC, AML, and pRCC were 0.94, 0.78, 0.80, and 0.76, respectively. Conclusion: Deep CNN model based on T2-weighted fat saturation sequence MR images was useful to classify the subtypes of renal parenchymal tumors with a relatively high diagnostic accuracy.