Kidney Detection In Ultrasound Images Based On Deep Neural Networks

Chronic kidney disease(CKD)is a disease with high morbidity and mortality,in which kidney function gradually loses over a period of months to years.It has become a worldwide public health problem of great concern.Early detection and early intervention can significantly improve the survival rate of CKD patients.Currently the most commonly used detection methods is a doctor,however,the use of ultrasound image data for diagnosis of reading,this way take a doctor a lot of time and energy.On the one hand,doctors work under great pressure in the face of huge outpatient volume every day.On the other hand,the doctor's diagnosis of judgment are susceptible to subjective factors such as their knowledge,experience and cognitive level,making the diagnosis of chronic kidney disease is difficult to achieve accurate analysis,and the subtle features some imperceptible to the naked eye.In view of the existing problems,the deep learning method is used to help the computer aided diagnosis(CAD).In order to facilitate the computer-aided diagnosis of chronic kidney disease,this paper proposes two methods based on deep neural network to realize the automatic detection of kidney in ultrasonic images.The significance of kidney detection is that the kidney ultrasonic parameter data obtained from the detection box can be provided to physicians for the auxiliary diagnosis of CKD.At the same time,in view of the academic circles of kidney ultrasound image data set is less,in order to facilitate the study of the subject of kidney detection of ultrasound image,this paper made a kidney detection data set of ultrasound image,which can be used by scientific and technological workers in related fields.The kidney detection methods proposed in this paper are as follows:(1)A kidney detection method based on DYOLO ultrasound image was proposed.Yolov3 has good applicability in the industry,with fast detection speed and high detection accuracy,but its performance is not superior in the face of renal ultrasound images with specific artifacts,noise and irregular deformation.In this paper,a heuristic algorithm with local spatial mutability,Deformable Convolutional Network(DCN),was integrated into YOLOV3 to achieve better adaptability and detection effect.Experimental results show that DYOLO algorithm can achieve higher detection accuracy than Yolov3 on kidney detection data set,up to 90.5%.(2)This paper presents a kidney detection method based on SYOLO ultrasound image.YOLO series detection algorithms have won wide praise in the industry.In 2020,YOLOV4,which is the collection of experts from many countries,was proposed to further improve the detection effect and bring more good news to the academia and related scientific researchers.However,its detection effect on ultrasonic images is still not perfect.This paper attempts to integrate the heuristic algorithm with global spatial variability--Spatial Transformation Network(STN)into YOLOV4,and the obtained SYOLO has better adaptability and detection effect for irregular deformation renal ultrasound images.Experimental results show that SYOLO algorithm can achieve the highest detection accuracy of 97.2% on kidney detection data set.The validity of the algorithm is proved.The two methods proposed in this paper is aimed at a spatial invariance of convolution neural network adaptive adjustment.So that the adjusted overall network performs better in the kidney detection task in ultrasonic images,greatly improves the detection accuracy,and provides practical significance for the clinical application in the early diagnosis of kidney diseases and the detection of other medical targets.