Research On Computer-aided Diagnosis Of Congenital Abnormalities Of The Kidney And Urinary Tract Based On Deep Learning

Congenital anomalies of the kidney and urinary tract(CAKUT)accounts for about 20%-30% of all prenatal malformations,which is also the leading cause of chronic kidney disease in children.Early screening diagnosis in children's CAKUT is important for early treatment intervention to improve the survival rate of children.Ultrasound imaging is the most practical,effective and safe method to diagnose CAKUT,and it is also the most important means to monitor and evaluate the prognosis.Automatic,fast and accurate CAKUT computer-aided diagnosis based on ultrasonic images has become a research direction in the field of intelligent medicine in the past ten years.Computer-aided diagnosis of CAKUT based on ultrasound images includes key technologies such as accurate,rapid segmentation and robust classification of the kidney region.Existing related algorithms usually pre-process kidney ultrasound image for representation features extraction by special manual design methods,these manual representation features are generally more complex,the extraction process is cumbersome.The deep learning algorithms learn features directly from the original input,and in recent years has made very remarkable research progress,providing a new way of rethinking for the diagnosis of kidney ultrasound images.On the other hand,because of the inherent characteristics of renal ultrasound imaging,the existing deep learning framework can not achieve good performance in the tasks of renal ultrasound image segmentation and classification.The research content of this paper focuses on the key technologies in the computer-aided diagnosis of CAKUT in ultrasound image based on deep learning,including the kidney ultrasound image segmentation and CAKUT classification diagnostic algorithm based on convolutional neural network.The work includes the following three key points:(1)The existing kidney ultrasound image segmentation method based on deep learning is generally realized by the pixel classification network,which is more suitable for the image with relatively consistent pixels in the segmentation area.However,due to the different tissues inside the kidneys and the influence of different types of CAKUT lesions,the pixels in the kidney tissue region show heterogeneity.In order to realize the accurate segmentation of kidney ultrasound image,based on the relatively consistent characteristics of the kidney edge in the appearance of ultrasound image,this paper proposes a method based on edge detection network to accurately segment the kidney ultrasound image,and a data augmentation method based on kidney shape registration is proposed to further improve the accuracy of segmentation.The experimental results show that this method can effectively improve the performance of automatic segmentation of the kidney region,which is obviously better than existing deep learning segmentation networks based on pixel classification,and effectively overcomes the problem of inaccurate segmentation caused by pixel heterogeneity in the region of kidney tissue.(2)To overcome the the problem of missing or blurring of kidney edges in some renal ultrasound images,a multi-task cascade end-to-end segmentation network framework is proposed.First,the edge detection network is used to obtain the kidney edge distance map,the rough positioning of the kidney region is obtained,and then the fine kidney segmentation region is further obtained from the kidney edge distance map by an area segmentation network based on pixel-wise classification.In addition,this paper proposes a multi-task loss function for endto-end training of the cascade network.The experimental results show that the automatic segmentation performance of kidney ultrasound images with missing or blurred edges can be significantly improved.(3)Because a single scan can only provide partial anatomical information of the kidneys,the appearance difference is large and the label is unknown,the classification diagnostic model established on a single 2D ultrasound image is not robust for the different ultrasound scanns of the same kidney.In order to make full use of the different scans of the same kidney,the CAKUT classification diagnostic model based on multi-instance deep learning is proposed,and the multiple ultrasound scans of each subject are modeled as multiple instances of a bag,so as to realize the robust classification diagnosis of kidney ultrasound images.The experimental results show that the proposed methods can diagnose CAKUT accurately based on ultrasound imaging,and its performance is better than that of single instance learning method or other existing deep learning models,which improves the efficiency and reliability of CAKUT diagnosis in children.