Dual-Path Convolutional Neural Network for Chronic Kidney Disease Classification in Ultrasound Echography

Abstract

Medical ultrasound (US) imaging is an important tool for the clinical diagnosis of chronic kidney disease (CKD). Since the interpretation of US images highly depends on the subjective experience of clinicians, artificial intelligence (AI) using convolutional neural networks (CNN) has been widely adopted to assist clinicians to improve their diagnoses. In this study, we retrospectively collected the renal US images and biopsy pathology reports in the past ten years from three hospitals affiliated with Taipei Medical University to realize the prediction of the stage of renal interstitial fibrosis and tubular atrophy (IFTA) by AI-assisted interpretation of US images. We will first input the renal US images into the Mask R-CNN model for ROI extraction with Intersection over Union (IoU) and Dice coefficient as quantitative metrics. Then, the proposed dual-path convolutional neural network (DPCNN) is used to simultaneously extract and integrate high-level and low-level features in the US image for IFTA prediction. With five-fold cross-validation, the proposed DPCNN for binary IFTA classification of non-diabetic nephropathy achieves the accuracy of 0.856 (0.818-0.876), the recall of 0.761 (0.715-0.817), the specificity of 0.927 (0.862-0.952), the precision of 0.887 (0.804-0.920), the F1 score of 0.819 (0.776-0.846) and the area under the receiver operating characteristic curve (AUC) of 0.922 (0.893-0.944). The results are all significantly better than other existing convolutional neural network models.