WO2020093563A1

WO2020093563A1 - Medical image processing method, system, device, and storage medium

Info

Publication number: WO2020093563A1
Application number: PCT/CN2018/124660
Authority: WO
Inventors: 徐勇; 罗笑玲; 蒲祖辉; 牟丽莎; 胡吉英
Original assignee: 哈尔滨工业大学（深圳）; 深圳市第二人民医院
Priority date: 2018-11-07
Filing date: 2018-12-28
Publication date: 2020-05-14
Also published as: CN109671049A; CN109671049B

Abstract

Disclosed in the present invention are a medical image processing method, a system, a device, and a storage medium. In one aspect, a plurality of first abnormality probabilities are acquired by means of a lesion image, and in another aspect, a plurality of second abnormality probabilities are acquired by means of a medical image and the lesion image. Then, on the basis of the plurality of first abnormality probabilities, the plurality of second abnormality probabilities, and different weight coefficients, and according to different image abnormality degrees, a final probability that the medical image belongs to the different image abnormality degrees is acquired, and the image abnormality degree having the greatest final probability is set as the image abnormality degree of the medical image, thus achieving analysis of the abnormality degree of the medical image, and overcoming the problems in the prior art of reliance on the naked eye and low efficiency in pathological image processing and analysis.

Description

Medical image processing method, system, equipment and storage medium

Technical field

The invention relates to the field of image processing, in particular to a medical image processing method, system, equipment and storage medium.

Background technique

The rapid growth of diabetes during the past few decades has attracted attention from all walks of life. The exponential growth of diseases caused by diabetes has become a huge challenge facing the current healthcare industry. And unfortunately, the number of patients with diseases caused by diabetes continues to grow at an alarming rate. What is more worrying is that only about 70% of patients realize that they have this disease. From a medical point of view, diabetes is considered to be the basis of many health problems and late-stage disorders, that is, diabetes can cause a series of diseases and complications, including: leading to serious heart disease, diabetic retinopathy (DR) and kidney problems. DR is one of the most common complications of diabetes, and it is considered to be one of the most important causes of blindness. Studies have shown that regardless of population size or socioeconomic background, diabetes problems in all regions of the world are increasing at an alarming rate. In addition, a study showed that due to defects in living conditions and treatment facilities in developing countries, nearly 75% of DR patients belong to developing countries, and the probability of blindness of DR patients is almost 25 times higher than those without DR . However, there is currently a lack of medical experts for diabetic retinopathy. Most DR patients in the world cannot get effective lesion detection and treatment in a timely manner. Most patients only develop when retinopathy has developed to a point where the treatment becomes highly complex and sometimes almost impossible. To be aware of seeking treatment. However, the cure rate of DR in the initial stage can reach 90%, so early detection of DR and effective treatment can greatly reduce the risk of DR causing blindness. Therefore, DR detection technology is particularly important, and medical image analysis is one of the research fields that have attracted great interest from scientists and physicians.

DR can be divided into two major categories, namely non-proliferative diabetic retinopathy (NPDR) and proliferative retinopathy (PDR). There are three subcategories of NPDR: mild NPDR, moderate NPDR, and severe NPDR. The non-proliferative type is the early stage of the lesion, and the lesion is limited to the retina, manifested as microhemangioma, hemorrhage, hard and soft exudate, retinal artery and venous lesions. Proliferative lesions at least partially extend beyond the inner limiting membrane, and the appearance of neovascularization is a sign of proliferative. DR lesions are described in detail as follows:

(1) Microaneurysm

Microaneurysms represent the most primitive perceptible signs of retinal damage. The abnormal permeability of retinal blood vessels leads to the formation of microaneurysms. On medical images, it is small, round, and has dark red spots that are usually in the macula cycle. The microaneurysm can be regarded as a red dot with sharp edges and a size between 20 μm and 200 μm, which is approximately 8.25% of the size of the disc.

(2) Hard exudates

Unlike microaneurysms, hard exudates are formed by the leakage of lipoproteins and other proteins from the blood vessels of the retina. Visually, it looks like small white or yellowish-white deposits with distinct edges. Rigid exudates are usually organized in the form of a ring, usually appearing in the outer layer of the retina. Rigid exudates are usually irregular and shiny, and are found close to the edges of microaneurysms or retinal edema.

(3) Soft exudate

In general, soft exudates are formed due to occlusion of small arteries. Reduced blood flow to the retina leads to ischemia of the retinal nerve fiber layer (RNFL), which ultimately affects axoplasmic flow, thereby accumulating axoplasmic debris on the axons of retinal ganglion cells. This accumulation can be visualized like a fluffy white lesion in RNFL, which is often referred to as soft exudate.

(4) Bleeding

Bleeding occurs due to leakage of weak blood vessels. The focus of bleeding is in the form of red dots with different densities and uneven edges, and it is found in the range of 125 μm. In a broad sense, hemorrhage is divided into two categories: flame and dot blotting. The first type originates from the anterior capillary artery and appears on nerve fibers. The second type of dot blot bleeding is round and smaller than a microaneurysm. Dot blot bleeding can appear on the retina at different levels, however, in most cases it will appear at the end of the capillaries' meridians.

(5) New blood vessel formation

Neovascularization usually indicates the atypical appearance of new blood vessels that appear on the inner surface of the retina. The new blood vessels are small and repeatedly penetrate into the vitreous cavity, which reduces visual ability and makes it significantly blurry, which eventually leads to blindness.

(6) Macular edema

Macular edema is identified as a swollen part of the retina, which usually occurs due to the permeability of abnormal retinal capillaries. Macular edema causes leakage of fluid or other solutes around the macula and severely affects vision.

Therefore, in the prior art, for patients with lesions including diabetic retinopathy (DR), only relying on the doctor ’s eye to determine whether the lesion occurred and the degree of the lesion by observing the pathological image, relying on the naked eye to observe and analyze the image takes a lot of time and effort And, the efficiency is low, and the judgments of different doctors are still different. This problem needs to be solved urgently.

Summary of the invention

The present invention aims to solve one of the technical problems in the related art at least to a certain extent. To this end, an object of the present invention is to provide a medical image processing method, system, device, and storage medium for improving the efficiency of processing and analyzing medical images.

The technical scheme adopted by the present invention is:

In a first aspect, the present invention provides a medical image processing method, including the following steps:

In the lesion image acquisition step, the lesion image is extracted based on the medical image;

In the score generation step, multiple lesion scores are obtained according to the lesion image, and the lesion score is a first abnormal probability that the lesion image belongs to different image abnormalities;

A second abnormal probability acquisition step, acquiring multiple second abnormal probabilities of the medical image according to the medical image and the lesion image, the second abnormal probability being the probability that the medical image belongs to different image abnormalities;

In the classification step, according to a plurality of the first abnormal probability, a plurality of the second abnormal probability and different weighting coefficients, the final probability that the medical image belongs to different image abnormalities is obtained according to different image abnormalities, and the largest The degree of image abnormality of the final probability is taken as the degree of image abnormality of the medical image.

Further, the medical image includes a fundus photograph.

Further, the step of acquiring a lesion image includes:

Acquire a variety of focus mask images based on the medical image and the segmentation neural network;

Acquire multiple lesion images based on the lesion mask image and the medical image.

Further, the step of generating scores includes:

Extracting lesion features of multiple lesion images separately, the lesion features including color or shape;

Splicing multiple lesion features and inputting them into the first machine learning classification algorithm to obtain multiple lesion scores.

Further, the step of acquiring the second abnormal probability includes:

Acquire a plurality of second abnormal probabilities of the medical image according to the medical image, a plurality of the lesion images, and a second machine learning classification algorithm.

Further, the degree of image abnormality includes normal, mild first abnormality, moderate first abnormality, severe first abnormality, and second abnormality.

Further, the weight coefficient of the first abnormal probability is 0.2, and the weight coefficient of the second abnormal probability is 0.8.

In a second aspect, the present invention provides a medical image processing system, including:

The lesion image acquisition unit is used to extract the lesion image according to the medical image;

A score generating unit, configured to obtain a plurality of lesion scores based on the lesion image, where the lesion score is a first abnormal probability that the lesion image belongs to different image abnormalities;

A second abnormal probability obtaining unit, configured to obtain a plurality of second abnormal probabilities of the medical image according to the medical image and the lesion image, the second abnormal probability being the probability that the medical image belongs to different image abnormalities ;

The classification unit is used to obtain the final probability that the medical image belongs to different image abnormalities according to different image abnormalities according to multiple first abnormal probabilities, multiple second abnormal probabilities and different weighting coefficients The image abnormality of the final probability is taken as the image abnormality of the medical image.

In a third aspect, the present invention provides a medical image processing device, including:

At least one processor; and,

A memory communicatively connected to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the medical image processing method.

In a fourth aspect, the present invention provides a computer-readable storage medium that stores computer-executable instructions that are used to cause a computer to execute the medical image processing method.

The beneficial effects of the present invention are:

On the one hand, the present invention obtains multiple first abnormal probabilities through the focus image, on the other hand obtains multiple second abnormal probabilities through the medical image and focus image, and then according to multiple first abnormal probabilities, multiple second abnormal probabilities and different The weight coefficients are used to obtain the final probability of medical images belonging to different image abnormalities according to different image abnormalities, and the image abnormality of the maximum final probability is taken as the image abnormality of medical images to realize the analysis of medical image abnormality and overcome the existing There is a technical problem that the processing and analysis of pathological images rely on the naked eye and the efficiency is low.

In addition, the present invention also obtains multiple lesion mask images of medical images by segmenting a neural network to further obtain multiple lesion images, ensuring accurate extraction of lesion images.

BRIEF DESCRIPTION

1 is a flowchart of a specific embodiment of a medical image processing method in the present invention;

2 is a schematic diagram of a specific embodiment of a segmented neural network of a medical image processing method in the present invention.

detailed description

It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other if there is no conflict.

Example 1

A medical image processing method includes the following steps:

In the score generation step, multiple lesion scores are obtained according to the lesion image. The lesion score is the first abnormal probability that the lesion image belongs to different image abnormalities. For example, if the image abnormality includes both normal and abnormal, the score generation step is to obtain the focus image. Normal or abnormal probability, get two first abnormal probabilities, and so on;

In the second abnormal probability obtaining step, multiple second abnormal probabilities of the medical image are obtained according to the medical image and the lesion image, and the second abnormal probability is the probability that the medical image belongs to different image abnormalities;

The classification step is to obtain the final probabilities of medical images belonging to different image abnormalities according to different image abnormalities according to multiple first abnormal probabilities, multiple second abnormal probabilities and different weighting coefficients, and take the maximum final probability image abnormality as the The degree of abnormality of medical images. That is, the first abnormal probability and the second abnormal probability of the same image abnormal degree are added according to different weight coefficients to obtain the final probability of the image abnormal degree. Since there are multiple different image abnormal degrees, multiple final probabilities of the medical image can be obtained .

A medical image processing method. On the one hand, multiple first abnormal probabilities are obtained from focus images, on the other hand, multiple second abnormal probabilities are obtained from medical images and focus images, and then multiple first abnormal probabilities and multiple second Abnormal probability and different weight coefficients are used to obtain the final probability that medical images belong to different image abnormalities according to different image abnormalities, and the image abnormality of the maximum final probability is taken as the image abnormality of medical images to realize the abnormality of medical images. The analysis improves the processing and analysis efficiency of medical images, and overcomes the existing technical problems of low efficiency in relying on the naked eye for processing and analyzing pathological images.

Further, medical images include, but are not limited to, fundus photos. In this embodiment, the fundus photos are used as examples to describe the medical image processing method. For other types of medical image processing methods, refer to the following specific methods for processing fundus photos The description is similar to the image processing method and will not be repeated here. The degree of image abnormality includes but is not limited to normal, mild first abnormality, moderate first abnormality, severe first abnormality, and second abnormality. Taking the fundus photo as an example, the first abnormality is a non-proliferative lesion. The second abnormality is a proliferative lesion. The abnormality of the fundus image includes, but is not limited to, normal, mild non-proliferative lesion, moderate non-proliferative lesion, severe non-proliferative lesion, and proliferative lesion. Referring to FIG. 1, FIG. 1 is a flowchart of a specific embodiment of a medical image processing method according to the present invention; the medical image processing method will be specifically described below:

Step 1: Pre-process the data. The data used in the present invention is a two-dimensional color fundus photograph. The original picture size is 3000 × 3000. The picture needs to be pre-processed, and the background and other unnecessary parts of the picture are cropped to adjust the picture size to 128 × 128.

Step 2: The step of acquiring lesion images, including:

Acquire multiple lesion mask images based on medical images and segmented neural networks;

Acquire multiple lesion images based on the lesion mask image and medical image.

Among them, the main task of segmenting the neural network is to realize the segmentation and extraction of multiple lesions. In this embodiment, the types of lesions include microaneurysms, rigid exudates, and neonatal retinal blood vessels. Lesions, new retinal blood vessels are proliferative lesions. Input the image to be tested into the trained segmentation neural network, and predict whether each pixel on the image belongs to the background, microaneurysm, hard exudate or retinal blood vessel through the segmentation neural network, and output the predicted result into three The three masks are the micro-aneurysm mask, the hard exudate mask, and the retinal vascular mask. The mask is specifically composed of 0 and 1, the background is 0, and the target is 1. For example, when generating a micro-aneurysm mask, predict whether each pixel belongs to the micro-aneurysm area by segmenting the neural network. If the pixel position is marked as 1, Otherwise, the pixel position is marked as 0, and finally a micro-aneurysm mask image with the same size as the original image of 128 × 128 is obtained. The mask can accurately describe the location of the target. Corresponding the mask to the original image, you can extract the image of each lesion, which is the lesion image.

Referring to Table 1 and FIG. 2, FIG. 2 is a schematic diagram of a specific embodiment of a segmentation neural network of a medical image processing method in the present invention; the implementation of the segmentation neural network is mainly based on a multi-task learning framework, and there are three segmentation tasks of the segmentation neural network , Respectively, segmentation of microaneurysm, segmentation of hard exudate, and segmentation of retinal vessels. The first part of the network is called the "shared layer", and the weights of these layers are common to all three tasks. Next, there is a specific layer for each task, called "task-specific layer". The parameters between these layers are calculated independently and do not participate in cross-layer sharing. In these task-specific layers, the network learns task-specific information. Each of these independent task layers will generate a separate output, specifically a segmentation mask that outputs microaneurysms, hard exudates, and retinal blood vessels.

Table 1 Segmented neural network hierarchy

The above segmentation neural network can segment multiple lesion images simultaneously, and is not limited to the three lesion images in this embodiment. Before using the segmented neural network, the network needs to be trained. Specifically, the model is first initialized, and then the sample with the marked mask data is input for training. The entire picture is used for training without patch sampling. The experiment proves that it is more effective to use the whole picture. After inputting the picture, perform three times convolution and downsampling to get the feature small image (5 × 5 × 128), and then perform another convolution to get the feature (1 × 1 × 1024), and then perform three sets of full connection to achieve three tasks The features of the three tasks are deconvolution and upsampling, respectively, to get a large picture of the same size as the original picture, these pictures are the masks that are segmented out. The segmentation neural network model is continuously optimized during the training process, and the parameters are adjusted until convergence. The goal of optimization is to continuously reduce the difference between the segmentation map (lesion image) predicted by the network and the sample annotation data, that is, to minimize the loss function L1:

In the above formula, N is the number of training samples, L _MA is the pixel-level loss of micro-aneurysm segmentation, L _HES is the segmentation loss of hard exudate, and L _NV is the segmentation loss of retinal vessels; f (x, θ) is the segmentation The segmentation result predicted by the neural network, where x is a pixel of the sample and θ is the learning rate; S ₁ , S ₂ , and S ₃ are the labeled positions of the samples of microaneurysm, hard exudate, and retinal blood vessel; Φ (θ) Is a regular item.

Then use the trained segmentation neural network to obtain a variety of lesion mask images of the fundus photos, including micro-aneurysm mask images, hard exudate mask images, and retinal vascular mask images; according to the lesion mask images and medicine The image can be segmented to obtain a variety of lesion images, including micro-aneurysm lesion images, rigid exudate lesion images, and retinal vascular lesion images.

Step 3: The score generation step, including:

Extract the lesion features of multiple lesion images separately, and the lesion features include color or shape;

The features of multiple lesions are spliced and input into the first machine learning classification algorithm to obtain multiple lesion scores.

Specifically, after obtaining the images of each lesion, the features such as the color and shape of each lesion are extracted according to the lesion image as shown in Table 2, 86-dimensional features can be extracted from each lesion image, and the lesion features of the 3 lesion images are The sequence of micro-aneurysm lesion images, rigid exudate lesion images, and retinal vascular lesion images are stitched together into a 258-dimensional feature vector, and the feature vector is input into the score generator (that is, the first machine learning classification algorithm) to obtain multiple The score of each lesion is

i is the category of the abnormality of the image, X is the number of the medical image, that is, the acquired lesion image belongs to different image abnormality (including normal, mild non-proliferative lesions, moderate non-proliferative lesions, severe non-proliferative lesions and proliferative Disease), that is, the score of 5 lesions can be obtained.

Table 2 List of lesion feature extraction

特征feature	特征描述Feature description
f ₁ f ₁	RGB图像上的均值Mean on RGB image
f ₂ f ₂	RGB图像上的方差Variance on RGB images
f ₃-f ₁₂ f ₃ -f ₁₂	RGB图像上的颜色矩Color moments on RGB images
f ₁₃ f ₁₃	目标区域的周长和The perimeter of the target area and
f ₁₄ f ₁₄	目标区域的面积和The area of the target area and
f ₁₅-f ₇₄ f ₁₅ -f ₇₄	目标区域的LBP特征LBP features of the target area
f ₇₅-f ₈₆ f ₇₅ -f ₈₆	HSV特征HSV characteristics

In this embodiment, the score generator uses a random forest algorithm. The score generator needs to be trained before it is used. Specifically, the 258-dimensional feature vector of the lesion feature of the training sample and the degree of lesion marked by the sample are input into the score generator for training. The trained score generator can calculate the probability that the input lesion belongs to 5 image abnormalities (that is, normal, mild non-proliferative lesions, moderate non-proliferative lesions, severe non-proliferative lesions, and proliferative lesions) , Take this probability value as the lesion score of the sample.

Step 4: The second abnormal probability acquisition step, including:

Acquire multiple second abnormal probabilities of the medical image according to the medical image, multiple lesion images, and the second machine learning classification algorithm.

In this embodiment, the image abnormality includes normal, mild non-proliferative lesions, moderate non-proliferative lesions, severe non-proliferative lesions, and proliferative lesions; the second machine learning classification algorithm is designed based on 3D convolutional neural network The 3D convolutional neural network can simultaneously input the original medical images (ie, fundus photos), microaneurysm lesion images, rigid exudate lesion images, and retinal vascular lesion images, which can extract features from multiple dimensions and then perform 3D convolution To capture feature information from multiple images. Specifically, referring to Table 3, the 3D convolutional neural network includes a solid-wired layer (ie, hard-connected layer), 4 convolutional layers, 3 down-sampled layers, and a fully connected layer. The convolved cubes of each convolution kernel are 4 pictures including the original image, the image of the microhemangioma lesion, the image of the retinal vascular lesion, and the image of the hard exudate lesion. The size of each picture is 128 × 128. At the first layer, a fixed solid-line kernel is applied to process the input graph to generate multiple channels of information, then multiple channels are processed separately, and finally the information of all channels are combined to obtain the final Feature description. This solid layer actually encodes a priori knowledge of the lesion, which is better than random initialization performance. Each picture extracts the information of five channels, namely: R, G, B color channels, and gradients in the x and y directions. So there are 4 × 5 = 20 lesion maps. Then, a 9 × 9 × 2 3D convolution kernel is used to perform convolution on each of the five channels. In order to increase the number of feature maps to extract different features, two different convolution kernels are used at each location, so that the L ₁ layer contains two feature map groups, each group contains (4-2 + 1 ) × 5 = 15 feature maps. Immediately following is the down-sampling layer L ₃ layer, which is down-sampled with a 3 × 3 window in the feature map of the L ₂ layer, so that the same number of lesion maps with reduced spatial resolution will be obtained. L ₄ is calculated by using 9 × 9 × 2 3D convolution kernels in 5 channels. In order to increase the number of feature maps, three different convolution kernels are used at each position, so that 6 different feature maps can be obtained, each group has (4-2 + 1-2 + 1) × 5 = 10 Feature maps. The L ₅ layer uses a 2 × 2 downsampling window, followed by two 2D convolutions and downsampling to obtain a 128 × 1 × 1 feature vector, and the 128 dimensions are determined based on past experience. After inputting medical images, after multiple layers of convolution and downsampling, the input multiple images are converted into a 128-dimensional feature vector, which captures the feature information of the retinal fundus photo. Then the 128-dimensional feature vector obtained is input to the softmax layer. The number of nodes in the softmax layer is consistent with the number of image abnormality categories, and each node is fully connected to the 128 nodes in L ₉ . Finally, the softmax layer will obtain a series of prediction values, that is, the probability value of the medical image belonging to the five image abnormalities, that is, the second abnormal probability, assuming this value

Among them, i (i = 0,1,2,3,4) is the category of image abnormality (ie normal, mild non-proliferative lesions, moderate non-proliferative lesions, severe non-proliferative lesions and proliferative lesions), X is the input medical image number.

Table 3 3D Convolutional Neural Network Hierarchy

层数Number of layers	类型Types of	神经元/特征尺寸Neuron / feature size	卷积核/窗口尺寸Convolution kernel / window size
00	输入层Input layer	4×128×1284 × 128 × 128	---
11	硬连接层Hard link layer	20×128×12820 × 128 × 128	---
22	卷积层Convolutional layer	15×2×120×12015 × 2 × 120 × 120	9×9×29 × 9 × 2
33	下采样层Downsampling layer	15×2×40×4015 × 2 × 40 × 40	3×33 × 3
44	卷积层Convolutional layer	10×6×32×3210 × 6 × 32 × 32	9×9×29 × 9 × 2
55	下采样层Downsampling layer	10×6×16×1610 × 6 × 16 × 16	2×22 × 2
66	卷积层Convolutional layer	10×6×10×1010 × 6 × 10 × 10	7×77 × 7
77	下采样层Downsampling layer	10×6×5×510 × 6 × 5 × 5	2×22 × 2
88	卷积层Convolutional layer	10×6×1×110 × 6 × 1 × 1	5×55 × 5

99	全连接层Fully connected layer	128×1×1128 × 1 × 1	---
1010	softmax层softmax layer	A	A

Before using the 3D convolutional neural network, you need to train it, first initialize the model, and then input the sample marked with abnormal degree for training. Input the original image of the sample, the image of the microhemangioma lesion, and the image of the hard exudate lesion during training. 4 pictures of retinal vascular lesion images for training. The 3D convolutional neural network model is continuously optimized during the training process, and the parameters are adjusted until convergence. The goal of optimization is to continuously reduce the difference between the classification results predicted by the network and the abnormal degree of sample labeling, that is, to minimize the loss function L2:

In the above formula, N is the number of training samples, L _softmax represents the network classification loss, and annotated data; f (x, θ) is the classification result predicted by the network, where x is a sample, θ is the learning rate; C is the label category; Φ (θ) is the regular term.

Step 5: Score fusion classification, specifically, in obtaining the second anomaly probability

(There are 5 second abnormal probabilities in this embodiment) and the first abnormal probability

(There are 5 first abnormal probabilities in this embodiment, that is, the lesion score), according to the setting of different weight coefficients to obtain the final abnormal probability, in this embodiment, the weight coefficient of the first abnormal probability is 0.2, the second abnormal probability The weight coefficient of is 0.8, then the final abnormal probability, that is, the score fusion value is

(There are 5 score fusion values in this embodiment). In the calculation result, which image abnormality category has the highest score fusion value, the category is the image abnormality category of the medical image, that is, C = argmax _i F ⁱ (X).

A medical image processing method, a score fusion mechanism is proposed, and the lesion scores obtained from each lesion image are integrated into a classification network, which improves the sensitivity and classification accuracy of each lesion. Moreover, the score fusion classification framework of the present invention is not limited to the three lesions detected in the present invention, and can be expanded when more lesions are detected, and the system has strong adaptability.

Example 2

A medical image processing system, including:

The score generating unit is used to obtain multiple lesion scores according to the lesion image, and the lesion score is the first abnormal probability that the lesion image belongs to different image abnormalities;

A second abnormal probability acquisition unit, configured to acquire multiple second abnormal probabilities of the medical image according to the medical image and the lesion image, the second abnormal probability is the probability that the medical image belongs to different image abnormalities;

The classification unit is used to obtain the final probabilities that medical images belong to different image abnormalities according to different image abnormalities according to multiple first abnormal probabilities, multiple second abnormal probabilities and different weight coefficients, and to abnormalize the images with the largest final probability The degree is regarded as the abnormal degree of the medical image.

For the specific working process and beneficial effects of the medical image processing system, please refer to the specific description of the medical image processing method in Embodiment 1, which will not be repeated here.

Example 3

A medical image processing device, including:

At least one processor; and,

A memory communicatively connected to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the medical image processing method. For a detailed description of the medical image processing method, refer to the description in Embodiment 1, and no further description is required.

Example 4

A computer-readable storage medium stores computer-executable instructions for causing a computer to execute the medical image processing method. For a detailed description of the medical image processing method, refer to the description in Embodiment 1, and no further description is required.

The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments, and those skilled in the art can make various equivalent modifications or replacements without violating the spirit of the present invention These equivalent variations or replacements are included in the scope defined by the claims of this application.

Claims

A medical image processing method, characterized in that it includes the following steps:

In the lesion image acquisition step, the lesion image is extracted based on the medical image;

In the score generation step, multiple lesion scores are obtained according to the lesion image, and the lesion score is a first abnormal probability that the lesion image belongs to different image abnormalities;

A second abnormal probability acquisition step, acquiring multiple second abnormal probabilities of the medical image according to the medical image and the lesion image, the second abnormal probability being the probability that the medical image belongs to different image abnormalities;

In the classification step, according to a plurality of the first abnormal probability, a plurality of the second abnormal probability and different weighting coefficients, the final probability that the medical image belongs to different image abnormalities is obtained according to different image abnormalities, and the largest The degree of image abnormality of the final probability is taken as the degree of image abnormality of the medical image.
The medical image processing method according to claim 1, wherein the medical image includes a fundus photograph.
The medical image processing method according to claim 1, wherein the lesion image acquisition step includes:

Acquire a variety of focus mask images based on the medical image and the segmentation neural network;

Acquire multiple lesion images based on the lesion mask image and the medical image.
The medical image processing method according to claim 3, wherein the score generation step includes:

Extracting lesion features of multiple lesion images separately, the lesion features including color or shape;

Splicing multiple lesion features and inputting them into the first machine learning classification algorithm to obtain multiple lesion scores.
The medical image processing method according to claim 3, wherein the step of acquiring the second abnormal probability includes:

Acquire a plurality of second abnormal probabilities of the medical image according to the medical image, a plurality of the lesion images, and a second machine learning classification algorithm.
The medical image processing method according to any one of claims 1 to 5, wherein the degree of image abnormality includes normal, mild first abnormality, moderate first abnormality, severe first abnormality, and second abnormality.
The medical image processing method according to any one of claims 1 to 5, wherein the weight coefficient of the first abnormal probability is 0.2, and the weight coefficient of the second abnormal probability is 0.8.
A medical image processing system, characterized in that it includes:

The lesion image acquisition unit is used to extract the lesion image according to the medical image;

A score generating unit, configured to obtain multiple lesion scores based on the lesion image, where the lesion score is a first abnormal probability that the lesion image belongs to different image abnormalities;

A second abnormal probability obtaining unit, configured to obtain a plurality of second abnormal probabilities of the medical image according to the medical image and the lesion image, the second abnormal probability being the probability that the medical image belongs to different image abnormalities ;

The classification unit is used to obtain the final probability that the medical image belongs to different image abnormalities according to different image abnormalities according to multiple first abnormal probabilities, multiple second abnormal probabilities and different weighting coefficients The image abnormality of the final probability is taken as the image abnormality of the medical image.
A medical image processing device, characterized in that it includes:

At least one processor; and,

A memory communicatively connected to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute any one of claims 1 to 7. Medical image processing method.
A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-executable instructions for causing a computer to perform the medicine according to any one of claims 1 to 7. Image processing method.