CN113570576A

CN113570576A - Method for evaluating detection of cirrhosis index by platelet counting method

Info

Publication number: CN113570576A
Application number: CN202110859107.1A
Authority: CN
Inventors: 付苗苗; 郭劲宏
Original assignee: Chengdu Yunxin Medical Technology Co Ltd
Current assignee: Chengdu Yunxin Medical Technology Co Ltd
Priority date: 2021-07-28
Filing date: 2021-07-28
Publication date: 2021-10-29
Anticipated expiration: 2041-07-28
Also published as: CN113570576B

Abstract

The invention discloses a method for evaluating cirrhosis index detection by a platelet counting method, which comprises the steps of establishing a yolo _ v3 network structure, training a network model, outputting network prediction, and analyzing an experimental result; the invention can classify the cells with three different sizes in blood by using the yolo _ v3 network structure, and the used main network darknet _53 has excellent characteristic extraction capability, and the algorithm can be more accurately detected without gradient disappearance caused by too large depth by introducing the residual block into the deep network, thereby avoiding manual counting, reducing the waste of manpower and material resources, realizing real-time detection of the target, returning to the position and counting, and saving time for clinical research and improving the efficiency of liver cirrhosis detection.

Description

Method for evaluating detection of cirrhosis index by platelet counting method

Technical Field

The invention relates to the technical field of medical diagnosis, in particular to a method for evaluating cirrhosis index detection by a platelet counting method.

Background

The liver is an organ in the body that is mainly responsible for metabolic functions and has the actions of detoxification, hematopoiesis and coagulation. And alcoholism, nutritional disorder, virus and other reasons easily cause people to suffer from various liver diseases, the early symptoms of the liver diseases are not obvious, and a plurality of tissue systems and even liver cirrhosis can be affected along with the development of the disease condition. The cirrhosis detection model is complex, has poor real-time effect, and may miss the best time for treating the patient, so a separate formula for detecting cirrhosis is generated, namely the APRI index: (AST (aspartate aminotransferase) to platelet ratio, which is an index that can efficiently and accurately identify patients with cirrhosis, and that concerns the number of platelets is of great importance for cirrhosis detection and prognosis;

the existing platelet counting algorithm is sensitive to the resolution of an image, has large influence on interference of illumination and scattering, needs a plurality of manual processing steps, can be influenced and judged by various factors, and is easy to generate the problems of gradient disappearance and gradient dispersion by using a mask-rcnn-based feature extraction network, has small target classification loss and large main position loss, so the invention provides a method for evaluating the liver cirrhosis index detection by using a platelet counting method to solve the problems in the prior art.

Disclosure of Invention

In view of the above problems, an object of the present invention is to provide a method for evaluating detection of liver cirrhosis index by a platelet counting method, which can classify three cells of different sizes in blood using yolo _ v3 network structure, and uses a backbone network darknet _53 with superior feature extraction capability, and introduces a residual block in a deep network to enable an algorithm to detect more accurately without gradient disappearance due to too large depth, and simultaneously avoids manpower and material resources waste due to manual counting, so as to achieve real-time detection of target, return to position and count, and save time for clinical research and improvement of liver cirrhosis detection efficiency.

In order to realize the purpose of the invention, the invention is realized by the following technical scheme: a method of assessing cirrhosis index detection by a platelet enumeration method comprising the steps of:

step one, building a Yolo _ v3 network structure, and building a Yolo _ v3 network consisting of a backbone network Darknet _53 and a feature prediction Yolo _ head;

training a network model, namely pre-training the network model by using a COCO data set in a transfer learning mode, and then performing small sample training by using blood cells to obtain a prediction frame;

thirdly, network prediction output, namely, operating the network by using the parameters after network training, performing prediction output, inputting the original image into the network, predicting the position, the type and the probability of each cell on the original image in the network, outputting the position type probability of each cell and marking;

and step four, analyzing an experimental result, using the BCCD blood cell data set and the blood cell picture with data enhancement as a training set and a test set test network model, and obtaining a result.

The further improvement lies in that: in the first step, the trunk network Darknet _53 is a feature extraction trunk network, and is composed of five residual blocks, each of which is composed of 1,2,8,8,4 residual units, namely a DBL convolutional layer, a batch normalization layer and a linear unit with leakage correction.

The further improvement lies in that: in the first step, the feature prediction Yolo _ head is a prediction output network, and after the feature maps of three scales are input into the trunk network Darknet _53, the diversity of the receptive fields is realized, and the prediction of the three scales is output through the Yolo _ head.

The further improvement lies in that: three prior frames are provided for each grid of the feature map for each scale in the three-scale predictions, for a total of 13 × 3+26 × 3+52 × 3 — 10647 predictions.

The further improvement lies in that: the prediction process in the third step comprises the following specific steps

1) Selecting a blood cell color picture input of a BCCD data set, performing relevant processing, converting the blood cell color picture input into a format BRG which can be processed by a pytorch, and adjusting the BRG into 416 × 416;

2) sending the data to a neural network to obtain a prediction result, dividing the picture into S × S grids by the Yolo _ head network, then enabling each unit grid to be responsible for predicting a target with a central point in the grid, wherein the final prediction value of each layer is a tensor with the size of S × S3 (4+1+3), the first 3 is a dimension number which represents the number of anchors in each layer, 4 is an adjustment parameter (x, y, w, h) of the size and the central position of a boundary frame, 1 is a confidence coefficient, and 3 is a cell category number;

3) decoding the prediction result, adding corresponding x _ offset and y _ offset to each grid point to obtain a result, namely the center of the prediction frame, and calculating the position of the whole prediction frame by utilizing the combination of the prior frame, h and w;

4) finally, performing score sorting and non-maximum inhibition screening on the prediction result, filtering the anchor frame through confidence, and then removing a plurality of detection frames by using non-maximum inhibition;

5) the decoded output is the position of the actual detection target, which is marked on the original image.

The further improvement lies in that: the key for realizing non-maximum inhibition in the step 4) is to select a frame with the highest score, calculate the coincidence degree of the frame with other frames, remove the frame with the coincidence degree exceeding the IOU threshold value, return to the previous step of iteration and stop the iteration until the frame which is not lower than the currently selected frame.

The further improvement lies in that: the prediction frames obtained by training the model in the second step are divided into three conditions of a positive case, a negative case and an ignored case, the positive case is any one ground route, IOU thresholds are all calculated with all the prediction frames output by the network model, the prediction frame with the largest IOU threshold is the positive case, and the positive case generates confidence coefficient loss, detection frame loss and category loss; the negative example is that the IOU threshold values of all the group entries are smaller than the set threshold value after the positive example is removed, namely the negative example is obtained, and the negative example only has confidence coefficient loss; except for the normal case, the case that the IOU threshold value of any one group route is larger than the set threshold value is the neglected case, and the neglected case does not generate loss.

The further improvement lies in that: and in the second step, the learning attenuation strategy selected by training is fixed step attenuation, the learning rate of the fixed step attenuation at the initial stage of training is high, so that the network model is converged quickly, the learning rate at the later stage of training is low, and the network model is converged to the optimal solution better.

The invention has the beneficial effects that: the invention can classify the cells with three different sizes in blood by using the yolo _ v3 network structure, and the used main network darknet _53 has excellent characteristic extraction capability, and the algorithm can be more accurately detected without gradient disappearance caused by too large depth by introducing the residual block into the deep network, thereby avoiding manual counting, reducing the waste of manpower and material resources, realizing real-time detection of the target, returning to the position and counting, and saving time for clinical research and improving the efficiency of liver cirrhosis detection.

Drawings

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

Fig. 2 is a diagram of a network structure implementation architecture of yolo _ v3 according to an embodiment of the present invention.

FIG. 3 is a structural diagram of Darknet _53 according to the present invention.

Fig. 4 is a diagram of the residual block of the present invention.

Detailed Description

In order to further understand the present invention, the following detailed description will be made with reference to the following examples, which are only used for explaining the present invention and are not to be construed as limiting the scope of the present invention.

Example 1

According to fig. 1,2, 3 and 4, the present embodiment provides a method for evaluating the detection of cirrhosis index by the platelet counting method, comprising the following steps:

step one, building a Yolo _ v3 network structure, building a Yolo _ v3 network composed of a backbone network Darknet _53 and a feature prediction Yolo _ head, wherein the backbone network Darknet _53 is a feature extraction backbone network and composed of five residual blocks, each residual block is composed of 1,2,8,8 and 4 residual units, the residual units are DBL convolution layers, batch normalization layers and leakage correction linear units, the feature prediction Yolo _ head is a prediction output network, after three-dimension feature maps are input into the backbone network Darknet _53, the diversification of sense fields is realized, the prediction of three dimensions is output through the Yolo _ head, three priori frames are arranged in each grid of the feature maps of each dimension, and the total number of 13, 3, 26, 3, 52, 3 is 10647;

training a network model, wherein a learning attenuation strategy selected by training is fixed step attenuation, the learning rate of the fixed step attenuation at the initial stage of training is high, the network model is converged quickly, the learning rate at the later stage of training is low, the network model is converged to an optimal solution better, the COCO data set is used for pre-training the network model in a transfer learning mode, and then blood cells are used for carrying out small sample training to obtain a prediction frame, wherein the prediction frame is divided into three conditions of a positive case, a negative case and an neglected case:

the positive case is that any one ground route is selected, the IOU value is calculated with all the prediction boxes output by the network model, the prediction box with the largest IOU value is the positive case, and the positive case generates confidence loss, detection box loss and category loss;

the negative example is that the IOU values of all the group entries are smaller than the set threshold after the positive example is removed, and the negative example only has confidence loss;

except for the normal case, the case that the IOU value of any one group route is larger than the set threshold value is the neglected case, and the neglected case does not generate loss.

And thirdly, network prediction output, namely, using the parameters after the network training to operate the network to perform prediction output. Inputting an original image into a network, predicting the position, the type and the probability of each cell on the original image in the network, outputting the position type probability of each cell and marking, wherein the predicting process specifically comprises the following steps:

2) sending the data to a neural network to obtain a prediction result, dividing the network bar picture into S-S grids, then enabling each unit grid to be responsible for predicting a target with a central point in the grid, wherein the final prediction value of each layer is a tensor with the size of S-3 (4+1+3), the first 3 is a dimension number which represents the number of anchors in each layer, 4 is an adjusting parameter (x, y, w, h) of the size and the central position of a boundary frame, 1 is a confidence coefficient, and 3 is a cell type number;

4) the score sorting and non-maximum inhibition screening of the final prediction result are implemented by filtering an anchor frame through confidence degrees, and then removing a plurality of detection frames by using non-maximum inhibition, wherein the key for realizing the non-maximum inhibition is to select a frame with the highest score, calculate the coincidence degree of the frame with other frames, remove the frame with the coincidence degree exceeding the IOU threshold value, return to the previous step for iteration, and stop iteration until the frame which is not lower than the currently selected frame;

5) the decoded output is the position of the actual detection target, which is marked on the original image;

Example 2

step one, building a Yolo _ v3 network structure, building a Yolo _ v3 network composed of a backbone network Darknet _53 and a feature prediction Yolo _ head, wherein the backbone network Darknet _53 is a feature extraction backbone network and composed of five residual blocks, each residual block is composed of 1,2,8,8 and 4 residual units, the residual units are a DBL convolution layer, a batch standardization layer and a band leakage correction linear unit, the batch standardization layer is used as a regularization method and can accelerate convergence and avoid overfitting, the difficulty of learning features is increased along with the deepening of the network, the learning process of the residual blocks is changed from directly learning features to adding certain features on the basis of the previously learned features to obtain better features, therefore, the learning target is changed from learning complete information to learning residual, so that the difficulty of learning high-quality features is greatly reduced, wherein DBL is an abbreviation for Darknetconv2D _ BN _ Leaky.

The feature prediction Yolo _ head is a prediction output network, the backbone network Darknet _53 realizes the diversity of receptive fields after inputting feature maps of three scales, the feature maps of the three scales are output through the Yolo _ head, each grid of the feature maps of each scale is provided with three priori frames, 13 x 3+26 x 3+52 x3 10647 predictions are total, each prediction is an 8-dimensional vector (4+1+3), the 8-dimensional vector contains frame coordinates (x, y, w, h), frame confidence (namely the probability that the priori frame has an object, which is a numerical value), and the probability of an object class; the smallest signature 13 x13, with the largest receptive field, may employ the larger three prior boxes, (116x90), (156x198), (373x326) suitable for detecting larger subjects. The 26 x 26 signature has a moderate receptive field, suitable for detecting medium sized objects, assigned a priori boxes of sizes (30x61), (62x45), (59x119), and the 52 x 52 signature has a priori boxes of sizes (10x13), (16x30), (33x23), having a minimal receptive field, suitable for small target object detection;

the implementation mode of the yolo _ v3 network structure is as follows: on the basis of a feature map obtained by Darknet-53, obtaining a first feature map through six DBL structures and a last convolutional layer, performing first prediction on the feature map, connecting up-sampled features with a convolutional feature tensor output by a Res8 structure from the output of a last-to-last-3 convolutional layer from back to front through one DBL structure and one up-sampling on a Y1 branch, obtaining a second feature map through the six DBL structures and the last convolutional layer, and performing second prediction on the feature map. And on the Y2 branch, the output of the 3 rd convolutional layer from back to front is subjected to one DBL structure and one upsampling, the upsampling characteristics are connected with the convolutional characteristic tensor output by the 1 st Res8 structure, a third characteristic map is obtained through six DBL structures and the last convolutional layer, and third prediction is carried out on the characteristic map.

the positive case is that any one group route is selected, the IOU value is calculated with all 10647 boxes, the prediction box with the largest IOU value is the positive case, the positive case generates confidence loss, detection box loss and category loss, and one prediction box can only be allocated to one group route;

the negative example is that the IOU threshold value of all the group entries is smaller than the set threshold value (set to 0.5) after the positive example is removed, and the negative example only has confidence loss;

except for the normal case, the case that the IOU threshold value of any one group route is greater than the set threshold value (set to 0.5) is the neglected case, and the neglected case does not generate loss;

in the final output of the network, there are three types of parameters for the output of the bounding box corresponding to each grid cell: one is the box parameter of the object, which is a total of four values, namely the coordinates of the center point of the box (x, y) and the width and height of the box (w, h); one is confidence, which is the value of the interval between [0,1 ]; the last is a set of conditional class probabilities, all values in the interval between [0,1], representing the probability.

4) screening the final prediction result according to score sorting and a non-maximum inhibition algorithm, filtering an anchor frame through confidence, and removing a plurality of detection frames by using non-maximum inhibition, wherein the key for realizing non-maximum inhibition is to select a frame with the highest score, calculate the contact ratio of the frame and other frames, remove the frame with the contact ratio exceeding the IOU value, return to the previous step for iteration, and stop iteration until the frame which is not lower than the currently selected frame;

And step four, analyzing the experimental result, namely extracting 328 blood cells as a training set and extracting 59 blood cells as a testing set by using the BCCD blood cell data set and 387 blood cell pictures with enhanced data.

Each iteration is performed 100 times, a model parameter is stored, total loss is an index of an error rate, the lower the total loss is, the better the total loss is, the larger the total loss is, the training set loss is 3642, the verification set loss reaches 553, but the network convergence is fast, the loss is fast, even the minimum loss is 44 of the training set, the verification set loss is 35, and finally, the parameter when epoach is 100 is selected, namely the training set loss is 44, and the verification set loss is 36.

The final experimental results were: the recall rate was 90.1% (56/62), the accuracy was 85.9% (56/65), and the platelet count was completely correct in 59 of the 45 pictures.

According to the method for evaluating the detection of the cirrhosis index by the platelet counting method, the yolo _ v3 network structure is used for classifying cells with three different sizes in blood, the used backbone network darknet _53 has excellent feature extraction capability, a residual block is introduced into a deep network, so that the algorithm is more accurately detected, gradient disappearance caused by too large depth is avoided, meanwhile, manual counting is avoided, waste of manpower and material resources is reduced, real-time detection target is achieved, positions are regressed and counted, clinical research is achieved, and the time is saved for improving the efficiency of cirrhosis detection.

The foregoing illustrates and describes the principles, general features, and advantages of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims

1. A method for assessing cirrhosis index detection by a platelet enumeration method, comprising the steps of:

2. A method for assessing the detection of cirrhosis index by the platelet enumeration method according to claim 1, wherein: in the first step, the trunk network Darknet _53 is a feature extraction trunk network, and is composed of five residual blocks, each of which is composed of 1,2,8,8,4 residual units, namely a DBL convolutional layer, a batch normalization layer and a linear unit with leakage correction.

3. A method for assessing the detection of cirrhosis index by the platelet enumeration method according to claim 1, wherein: in the first step, the feature prediction Yolo _ head is a prediction output network, and after the feature maps of three scales are input into the trunk network Darknet _53, the diversity of the receptive fields is realized, and the prediction of the three scales is output through the Yolo _ head.

4. A method for assessing the detection of cirrhosis index by the platelet enumeration method according to claim 3, wherein: three prior frames are provided for each grid of the feature map for each scale in the three-scale predictions, for a total of 13 × 3+26 × 3+52 × 3 — 10647 predictions.

5. The method for evaluating the detection of cirrhosis index by the platelet counting method according to claim 1, wherein the prediction process in the third step is specifically as follows:

6. The method of claim 5 for assessing the index of cirrhosis by a platelet enumeration method, wherein: the key for realizing non-maximum inhibition in the step 4) is to select a frame with the highest score, calculate the coincidence degree of the frame with other frames, remove the frame with the coincidence degree exceeding the IOU threshold value, return to the previous step of iteration and stop the iteration until the frame which is not lower than the currently selected frame.

7. A method for assessing the detection of cirrhosis index by the platelet enumeration method according to claim 1, wherein: the prediction frames obtained by training the model in the second step are divided into three conditions of a positive case, a negative case and an ignored case, the positive case is any one ground route, IOU thresholds are all calculated with all the prediction frames output by the network model, the prediction frame with the largest IOU threshold is the positive case, and the positive case generates confidence coefficient loss, detection frame loss and category loss; the negative example is that the IOU threshold values of all the group entries are smaller than the set threshold value after the positive example is removed, namely the negative example is obtained, and the negative example only has confidence coefficient loss; except for the normal case, the case that the IOU threshold value of any one group route is larger than the set threshold value is the neglected case, and the neglected case does not generate loss.

8. A method for assessing the detection of cirrhosis index by the platelet enumeration method according to claim 1, wherein: and in the second step, the learning attenuation strategy selected by training is fixed step attenuation, the learning rate of the fixed step attenuation at the initial stage of training is high, so that the network model is converged quickly, the learning rate at the later stage of training is low, and the network model is converged to the optimal solution better.