阅读说明：本技术 一种基于dmsa图像的膀胱输尿管反流预测方法 (Vesicoureteral reflux prediction method based on DMSA image ) 是由 徐虹 吴明妍 沈茜 毕允力 龚一女 许丽楠 吴哈 阮彤 曾健骏 于 2019-10-23 设计创作，主要内容包括：本发明公开了一种基于DMSA图像的膀胱输尿管反流预测方法,该方法通过预先筛选病人数据来得到相对平衡的数据集,在该数据集上利用三层CNN网络作为图像解码器,将图像解码器的输出拼接从文本得到的DMSA检查特征,接入两层密集连接层,从DMSA图像预测膀胱输尿管反流的存在与否与级别轻重。本发明提前预测患儿膀胱输尿管反流的存在与否与级别轻重,避免患儿进行不必要的逆行性膀胱尿路造影检查。(The invention discloses a bladder ureter reflux prediction method based on a DMSA image, which obtains a relatively balanced data set by pre-screening patient data, uses a three-layer CNN network as an image decoder on the data set, splices the output of the image decoder with DMSA inspection characteristics obtained from texts, accesses two dense connecting layers, and predicts the existence and grade of bladder ureter reflux from the DMSA image. The method predicts the existence and grade of the reflux of the vesicoureteral tract of the infant in advance, and avoids unnecessary retrograde vesicoureterography examination of the infant.)

1. A vesicoureteral reflux prediction method based on DMSA images is characterized by comprising the following steps:

g1: obtaining a plurality of groups of scanned images of double kidneys of a patient through DMSA kidney static scanning, dividing a DMSA front inspection image into two images of a left kidney and a right kidney, horizontally turning the left kidney image, and taking the left kidney image and the right kidney image together as two sub-kidney images;

g2: selecting a group of DMSA images, and taking the DMSA grades of the upper part, the middle part and the lower part of the side kidney, the absolute value/100 of the DMSA function of the kidney, the absolute function ratio of the kidney/the kidney, and the difference of the DMSA functions as additional characteristics;

g3: converting the two kidney-divided images obtained in G2 into gray-scale images and scaling to 128 × 128 pixels;

g4: inputting the gray level image in G3 to a prediction model of an image decoder based on three CNNs, wherein the number of output channels of a 1 st CONV layer (convolution layer) of the prediction model is 32, the convolution kernel size is 3 x 3, an activation function adopts relu, and the activation function is input to a 2 nd CONV layer after being subjected to maximum pooling of a layer of kernel size of 2 x 2; the number of output channels of the 2 nd CONV layer is 64, the convolution kernel size is 3 x 3, the activation function adopts relu, and the activation function is input to the 3 rd CONV layer after being subjected to maximum pooling with the kernel size of 2 x 2; the number of output channels of the 3 rd layer of CONV layer is 128, the convolution kernel size is 3 x 3, the activation function adopts relu, the two-dimensional tensor is flattened into a one-dimensional tensor after one layer of maximum pooling with the kernel size of 2 x 2, then additional features are spliced into the one-dimensional tensor, the obtained one-dimensional tensor inputs 512 dense connection layers of the number of neurons, the activation function adopts relu, then 1 dense connection base layer of the number of neurons is input, and the activation function adopts sigmoid;

g5: continuously selecting DMSA images of other groups, repeating the steps G2-G4, and repeatedly training a prediction model;

g6: and D, sequentially processing the DMSA images of the double kidneys of the patient to be evaluated through the steps G2-G4 to obtain the final shunt grade.

2. The DMSA image-based vesicoureteral reflux prediction method according to claim 1, wherein the shunt grade of a patient undergoing reflux examination is classified into 0-5 grades, wherein 0 is no reflux, and 5 grades are the highest reflux grade, and in the training stage of the prediction model, the shunt grade of the side kidney is 0-2 and is recorded as 0, and the grade is 3-5 and is recorded as 1 according to the severity of ureteral reflux shown in the kidney-separating image.

Technical Field

The invention relates to an image-combined recognition algorithm, which is particularly applied to a prediction model for predicting whether vesicoureteral reflux exists or not in children patients who clinically perform DMSA examination.

Background

Vesicoureteral reflux (VUR) is one of the congenital renal and urinary tract malformations common to children. The chronic renal failure diagnosis and treatment method is a common cause of repeated urinary infection and fever of children, and if the children are diagnosed not early and are not subjected to timely intervention treatment, the chronic renal failure diagnosis and treatment method can progress into the reflux nephropathy, seriously affect the kidney function and even progress into the end-stage nephropathy.

Currently, the clinical diagnosis VUR is usually performed by performing DMSA renal static scan on those children with febrile urinary tract infection, to determine the degree of renal injury, for example, if the kidney has been damaged (for example, renal scar appears, difference > =10 on both sides of renal function), then reverse bladder urography (MCU) is considered to determine the existence and grade of VUR. However, MCU examination is invasive examination, which can cause discomfort to children patients, and meanwhile MCU has certain radioactivity, so that children patients can avoid multiple examinations in a short period.

There is currently no other test available to replace the clinical use of MCU diagnostics VUR.

A Convolutional Neural Network (CNN) is a feedforward Neural Network, and an artificial neuron can respond to peripheral units and perform large-scale image processing. The convolutional neural network includes convolutional layers and pooling layers.

Disclosure of Invention

The invention aims to predict the existence and the grade of the infant VUR in advance by analyzing the DMSA image and adopting an image recognition technology, thereby avoiding unnecessary MCU examination of the infant.

The invention is realized according to the following technical scheme.

A vesicoureteral reflux prediction method based on DMSA images comprises the following steps:

g1: obtaining a plurality of groups of scanned images of double kidneys of a patient through DMSA kidney static scanning, dividing a DMSA front inspection image into two images of a left kidney and a right kidney, horizontally turning the left kidney image, and taking the left kidney image and the right kidney image together as two sub-kidney images;

g2: selecting a group of DMSA images, and taking the DMSA grades of the upper part, the middle part and the lower part of the side kidney, the absolute value/100 of the DMSA function of the kidney, the absolute function ratio of the kidney/the kidney, and the difference of the DMSA function of the kidney as additional characteristics.

G3: converting the two kidney-divided images obtained in G2 into gray-scale images and scaling to 128 × 128 pixels;

g4: inputting the gray level image in G3 to a prediction model of an image decoder based on three CNNs, wherein the number of output channels of a 1 st CONV layer (convolution layer) of the prediction model is 32, the convolution kernel size is 3 x 3, an activation function adopts relu, and the activation function is input to a 2 nd CONV layer after being subjected to maximum pooling of a layer of kernel size of 2 x 2; the number of output channels of the 2 nd CONV layer is 64, the convolution kernel size is 3 x 3, the activation function adopts relu, and the activation function is input to the 3 rd CONV layer after being subjected to maximum pooling with the kernel size of 2 x 2; the number of output channels of the 3 rd layer of CONV layer is 128, the convolution kernel size is 3 x 3, the activation function adopts relu, the two-dimensional tensor is flattened into a one-dimensional tensor after one layer of maximum pooling with the kernel size of 2 x 2, then additional features are spliced into the one-dimensional tensor, the obtained one-dimensional tensor inputs 512 dense connection layers of the number of neurons, the activation function adopts relu, then 1 dense connection base layer of the number of neurons is input, and the activation function adopts sigmoid;

g5: continuously selecting DMSA images of other groups, repeating the steps G2-G4, and repeatedly training a prediction model;

g6: and D, sequentially processing the DMSA images of the double kidneys of the patient to be evaluated through the steps G2-G4 to obtain the final shunt grade.

Furthermore, the shunt grade of the patient undergoing the reflux examination is divided into 0-5 grades, wherein the 0 grade is no reflux, and the highest grade of the 5 grades is reflux, and in the training stage of the prediction model, according to the reflux severity of the ureter reflected in the kidney-separating picture, the shunt grade of the side kidney is 0-2 and is recorded as 0, and the grade of the side kidney is 3-5 and is recorded as 1.

The present invention obtains the following advantageous effects.

By analyzing the DMSA image and adopting the image recognition technology, the invention predicts the existence and the grade of the reflux of the uroureter vesicae of the infant in advance and avoids unnecessary retrograde urocystography examination of the infant, thereby preventing the infant from causing discomfort and avoiding the injury to a patient caused by the radioactivity of the retrograde urocystography examination.

Drawings

FIG. 1 is a flow chart of the prediction method of the present invention.

Detailed Description

The present invention will be further illustrated with reference to the following examples.

A vesicoureteral reflux prediction method based on DMSA images comprises the following steps:

g1: obtaining a plurality of groups of scanned images of double kidneys of a patient through DMSA kidney static scanning, dividing a DMSA front inspection image into two images of a left kidney and a right kidney, horizontally turning the left kidney image, and taking the left kidney image and the right kidney image together as two sub-kidney images;

g2: selecting a group of DMSA images, and taking the DMSA grades of the upper part, the middle part and the lower part of the side kidney, the absolute value/100 of the DMSA function of the kidney, the absolute function ratio of the kidney/the kidney, and the difference of the DMSA function of the kidney as additional characteristics.

G3: converting the two kidney-divided images obtained in G2 into gray-scale images and scaling to 128 × 128 pixels;

g4: inputting the gray level image in G3 to a prediction model of an image decoder based on three CNNs, wherein the number of output channels of a 1 st CONV layer (convolution layer) of the prediction model is 32, the convolution kernel size is 3 x 3, an activation function adopts relu, and the activation function is input to a 2 nd CONV layer after being subjected to maximum pooling of a layer of kernel size of 2 x 2; the number of output channels of the 2 nd CONV layer is 64, the convolution kernel size is 3 x 3, the activation function adopts relu, and the activation function is input to the 3 rd CONV layer after being subjected to maximum pooling with the kernel size of 2 x 2; the number of output channels of the 3 rd layer of CONV layer is 128, the convolution kernel size is 3 x 3, the activation function adopts relu, the two-dimensional tensor is flattened into a one-dimensional tensor after one layer of maximum pooling with the kernel size of 2 x 2, then additional features are spliced into the one-dimensional tensor, the obtained one-dimensional tensor inputs 512 dense connection layers of the number of neurons, the activation function adopts relu, then 1 dense connection base layer of the number of neurons is input, and the activation function adopts sigmoid;

g5: continuously selecting DMSA images of other groups, repeating the steps G2-G4, and repeatedly training a prediction model;

g6: and D, sequentially processing the DMSA images of the double kidneys of the patient to be evaluated through the steps G2-G4 to obtain the final shunt grade.

Furthermore, the shunt grade of the patient undergoing the reflux examination is divided into 0-5 grades, wherein the 0 grade is no reflux, and the highest grade of the 5 grades is reflux, and in the training stage of the prediction model, according to the reflux severity of the ureter reflected in the kidney-separating picture, the shunt grade of the side kidney is 0-2 and is recorded as 0, and the grade of the side kidney is 3-5 and is recorded as 1.

The DMSA image report shows the renal function dividing function, which is obtained by drawing the outline of professional nuclear medicine software for the nuclear medicine doctor and calculating the outline by a system.

Image scoring was performed by nuclear medicine physicians and trained medical staff according to the american society for nuclear medicine evaluation criteria, the criteria for pyelonephritis: (ii) 1 or more than 1 focal radioactivity reduction or defect in renal cortex; or diffuse reduced radioactivity; the outline of the kidney is normal. Criteria for renal scarring: reduced focal radioactivity or defects in 1 or more than 1 of the renal cortex, associated with decreased contraction or volume of the affected cortex, or wedge-shaped defects.

The grade of reflux was assessed by urologic retrograde imaging medical staff according to the vesicoureteral reflux five-stage classification developed by the international reflux research council in 1981.