PASnet: A Joint Convolutional Neural Network for Noninvasive Renal Ultrasound Pathology Assessment

Abstract

Nephropathy is a worldwide clinical and health problem that is getting more and more attention from the public. The gold standard for the diagnosis of nephropathy is still renal puncture biopsy, which is an invasive examination and has many contraindications. We proposed to analyze renal ultrasound images using deep learning method to achieve noninvasive assessment. However, the kidney ultrasound images with accurate pathological diagnosis are relatively difficult to collect, which belongs to the category of few-shot learning. To mitigate the impact of few data on performance, this paper proposed a conceptually simple, flexible, and mixed framework for aided diagnosis of nephropathy. Our method, called the PASnet, consists of pretrained network and siamese network. Pretrained network trained by abundant samples from ImageNet can achieve fast convergence and better performance on a new data set. Siamese network learns to converge or disperse image pairs in distance space according to whether it comes from the same class or not. PASnet combines the advantages of these two methods and obtains a better classification performance on nephropathy classification through joint training. Accuracy of PASnet increases by 5.89% compared to a single network.