Combining the Transformers and CNNs for Renal Parenchymal Tumors Diagnosis

Abstract

Convolutional neural networks (CNNs) can be used to automatically classify the subtypes of renal parenchymal tumors. However, such approaches may overlook the long-range dependencies of features. In this study, we stacked Transformers and CNNs to efficiently capture both long-range dependencies of features and low-level spatial details. The optimal stack layout, the width, and the depth of the networks were determined according to the characteristics of the renal tumors. In addition, we adopted transfer learning for better performance. The experiments conducted on the T2-weighted magnetic resonance (T2W-MR) images from 199 patients presented an 82.9% overall classification accuracy and 0.96 average AUC. The results demonstrate that combining CNN with Transforms is an effective strategy for renal parenchymal tumors diagnosis.