CN114359288B - Medical image cerebral aneurysm detection and positioning method based on artificial intelligence - Google Patents

Abstract

The invention relates to a medical image cerebral aneurysm detection method based on artificial intelligence, which obtains a first confidence level map of the medical image to be detected; extracts the region of interest of the medical image to be detected, and obtains its eight neighborhoods according to the region of interest image, and then obtain the reference image; replace the pixels of the reference image with the pixels of the region of interest, obtain the reference medical image, and obtain the second confidence map; make the difference between the first part of the confidence map and the second part of the confidence map, Obtain the eight-neighborhood confidence change map; perform binarization processing on it, perform curve fitting on the pixel points with the pixel value of 1, and obtain N straight lines; judge whether the N straight lines have intersections in the region of interest, if If there is an intersection point and the confidence level of the intersection point is greater than the set threshold, the intersection point is a lesion point, and the medical image to be tested is a pathological image. That is, the present invention additionally analyzes the change of confidence in the neighborhood of the target area, thereby improving the accuracy of judgment and reducing the probability of misjudgment.

Description

Artificial intelligence-based method for detection and localization of cerebral aneurysm in medical imaging

technical field

The invention relates to the field of medical image processing, in particular to a method for detecting and locating cerebral aneurysm in medical images based on artificial intelligence.

Background technique

There are existing methods for cerebral aneurysm detection and localization based on medical image processing, in order to use neural network, combined with network optimization methods such as self-attention mechanism to realize the detection and localization of cerebral aneurysm. However, it needs enough lesion images as training samples, and it is difficult to obtain a large number of lesion images in practice, and due to the different attributes such as the degree of lesion and the location of the lesion, the trained detection network can be improved when the number of training samples is small. Poor generalization ability when identifying lesions.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an artificial intelligence-based medical imaging cerebral aneurysm detection and localization method, which is used to solve the problem that the trained detection network has poor generalization ability when recognizing lesions when the training samples are few. .

The technical solution of an artificial intelligence-based medical imaging cerebral aneurysm detection and localization method provided by the present invention includes the following steps:

acquiring a medical image to be detected, and preprocessing the medical image to be detected to obtain a first confidence map;

Extract the region of interest of the medical image to be detected, and obtain eight neighborhood images of the region of interest according to the region of interest; combine the eight neighborhoods to obtain an eight-channel neighborhood image, and combine the eight-channel neighborhood image Input the network prediction model to obtain the corresponding reference image; replace the pixels of the reference image with the pixels of the region of interest in the medical image to be detected to obtain the reference medical image, and perform the reference medical image. Preprocessing to obtain a second confidence map;

Extract the first part of the confidence map corresponding to the eight-neighborhood image of the medical image to be tested and the second part of the confidence map corresponding to the eight-neighborhood image of the reference medical image, and combine the first part of the confidence map and the second part of the confidence map Make the difference, and obtain the eight-neighborhood confidence change map; binarize the eight-neighborhood confidence change map, set the pixel points greater than 0 to 1, and perform curve fitting on the pixel points with the pixel value of 1 , get the fitted N straight lines, where N is greater than or equal to 1;

Determine whether the N straight lines have an intersection in the region of interest. If there is an intersection and the confidence of the intersection is greater than the set threshold, then the intersection is a lesion, and the medical image to be tested is a pathological image; otherwise, there is no lesions, the medical image to be tested is a normal image.

Further, the method for extracting the region of interest of the medical image to be detected is:

Selecting the first k pixels with high confidence in the first confidence map to form a first set of pixels;

Calculate the Euclidean distance between any two pixels in the first set of pixels, and when the Euclidean distance is less than a preset threshold, combine the two pixels into a single pixel, and the pixel of the single pixel The value is the pixel value of any one of the two pixel points; and the pixels that do not constitute a set are supplemented with the first pixel set, and the merging judgment is continued until the number of elements in the first pixel set is k, and then obtain a new first pixel point set;

大小的滑窗，采用所述滑窗提取新的第一像素点集合的每个像素点周围信息，并根据提取区域的置信度均值作为评价指标

，若评价指标

，则以

为最优模板尺寸，否则，令

，

，直至获取到最优滑窗尺寸；Construct

The size of the sliding window, the sliding window is used to extract the surrounding information of each pixel point of the new first pixel point set, and the confidence average value of the extracted area is used as the evaluation index

, if the evaluation index

, then with

is the optimal template size, otherwise, let

,

, until the optimal sliding window size is obtained;

The region of the medical image to be detected is extracted according to the optimal sliding window size, so as to obtain a region of interest with the same size as the optimal sliding window.

Further, the training process of the network prediction model is:

Obtain multiple real diseased medical images and non-lesioned medical images, and randomly select the center point of each medical image, obtain the center image and eight neighborhood images according to the optimal sliding window size, take the center image as the labeling data, and the eight neighborhoods The images are used as input training samples for the network prediction model, and the network prediction model is trained.

Further, the acquisition process of the described fitted N straight lines is:

1) Perform straight line fitting on the pixel points with a pixel value of 1 in the eight-neighborhood confidence change graph to obtain the first straight line;

2) Calculate the distance from each pixel to the first straight line, and obtain the average distance of all pixels. When the average distance is greater than or equal to the fit threshold, extract the pixel with the largest distance in turn, until the remaining The fitting degree of the pixel points is less than the fitting degree threshold, and the number of the extracted pixel points is counted. When the number is greater than the set number, a straight line is performed on the extracted pixel points to obtain a second straight line. , and calculate the corresponding distance mean, determine whether it is greater than or equal to the fit threshold, and so on, until all the pixels of the fitted lines meet the fit threshold, then N straight lines are obtained.

Further, the lesion point is screened according to the gradient of the confidence change value of the pixel points included in the fitted straight line. If the gradient from the intersection point in the region of interest to the outside shows a downward trend, the intersection point is retained, otherwise, the intersection point is deleted. Further, it also includes modifying the first confidence level corresponding to each pixel in the region of interest according to the intersection point:

，

表示第

个交点，

表示第

个像素点；1) Calculate the Euclidean distance between the intersection and each pixel as

,

means the first

a point of intersection,

means the first

pixels;

为：2) Then the confidence correction value of each intersection point

for:

，其中

为第一置信度图中各像素点的第一置信度，

为交点总个数。Then the corrected first confidence level of each pixel is

,in

is the first confidence level of each pixel in the first confidence level map,

is the total number of intersections.

Further, before comparing the confidence level of the intersection point with the set threshold, it also includes normalizing the confidence level of the pixel point.

Further, a category image is obtained according to the determined lesion point and the corrected first confidence map, and a connected domain is obtained by performing a connected domain analysis on the category graph, and the connected domain with the number of pixels less than the set number is filtered out. , take the center point of the bounding box of the remaining connected domains as the anchor point.

The present invention has the following beneficial effects: compared with the method of directly judging by the confidence, the present application obtains the confidence map before and after the replacement through the replacement operation, additionally analyzes the confidence change of the neighborhood of the target area, and based on the neighborhood The trend of confidence change assists the pixel confidence of the target area to be corrected to a certain extent, which improves the accuracy of judgment and reduces the probability of misjudgment and misjudgment.

Description of drawings

In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or in the prior art, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description The drawings are only some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

FIG. 1 is a method flowchart of a method for detecting a brain aneurysm in medical images based on artificial intelligence according to the present invention.

Detailed ways

In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following describes the solution according to the present invention, its specific implementation, structure, features and effects in detail with reference to the accompanying drawings and preferred embodiments. as follows. In the following description, different "one embodiment" or "another embodiment" are not necessarily referring to the same embodiment. Furthermore, the particular features, structures, or characteristics in one or more embodiments may be combined in any suitable form.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

The specific scheme of an artificial intelligence-based medical imaging cerebral aneurysm detection method provided by the present invention will be specifically described below with reference to the accompanying drawings.

Specifically, please refer to FIG. 1, which shows a method flowchart of an artificial intelligence-based medical imaging cerebral aneurysm detection method provided by an embodiment of the present invention, and the method includes the following steps:

Step 1: Acquire a medical image to be detected, and preprocess the medical image to be detected to obtain a first confidence map.

The medical image to be tested in this embodiment is a CT image.

Among them, the process of preprocessing the CT image is as follows:

1) Build a semantic segmentation network, and the semantic segmentation network structure is an encoder-decoder structure; use the label data to train the semantic segmentation network to obtain a trained semantic segmentation network;

The training label data is: several CT images, which are composed of multiple CT images with cerebral aneurysm and multiple CT images without cerebral aneurysm, and the label data is pixel-level annotation.

The loss function of the semantic segmentation network adopts binary cross entropy as the loss function.

Since the semantic segmentation network is a well-known network, its training process will not be described in detail here.

2) Take the image to be processed as the input, input the trained semantic segmentation network, output the semantic segmentation image, and then obtain the first confidence map;

。It should be noted that when acquiring the semantic segmentation image, each pixel has a corresponding category, and the acquisition process of the category is that the network outputs the category confidence of each pixel. Since the semantic segmentation described in this application is divided into two categories, then A pixel has two categories of confidence (tumor category and non-tumor category), and then processed by the Softmax function, the category corresponding to the maximum confidence is output as the category of the pixel; and the corresponding tumor category of the pixel is extracted. Confidence degree, the confidence degree corresponding to the tumor type of all pixel points is formed into a first confidence degree map, wherein the pixel value range of each pixel in the first confidence degree map is

.

Step 2: Extract the region of interest of the medical image to be detected, and obtain eight neighborhood images of the region of interest according to the region of interest; combine the eight neighborhoods to obtain an eight-channel neighborhood image, and combine the eight-channel neighborhood images. The neighborhood image is input into the network prediction model, and the corresponding reference image is obtained; the pixels of the reference image are replaced by the pixels of the region of interest in the medical image to be detected, and the reference medical image is obtained, and the reference medical image is compared. Perform preprocessing to obtain a second confidence map.

In this embodiment, the region of interest can be extracted by using the connected domain analysis method; as another implementation manner, the following steps can also be used to extract the region of interest, specifically:

1) Select the first k pixels with high confidence in the first confidence map to form the first set of pixels;

2) Calculate the Euclidean distance between any two pixels in the first set of pixels, and when the Euclidean distance is less than a preset threshold, combine the two pixels into a single pixel, the single pixel The pixel value of the pixel value is the pixel value of any pixel point in the two pixel points; and the pixel points that do not constitute a set are supplemented with the first pixel point set, and the merging judgment is continued until the number of elements in the first pixel point set is k, and then obtain a new first set of pixels;

可基于实施者对检测的精度自适应调整，所需检测精度越高，则

值越大，在本申请中设置为10；预设阈值

用于将第一像素点集合内距离较近的像素点合并，进而可以避免同一动脉瘤区域像素点被多次处理，在本实施例中预设阈值

设置为

，其中，

为第一置信度图的宽和高。It should be noted that here

It can be adaptively adjusted based on the detection accuracy of the implementer. The higher the required detection accuracy, the

The larger the value, it is set to 10 in this application; the preset threshold

It is used to combine the pixels with short distances in the first pixel point set, so as to avoid the pixels in the same aneurysm area being processed multiple times. In this embodiment, the threshold is preset.

Set as

,in,

are the width and height of the first confidence map.

大小的滑窗，采用所述滑窗提取新的第一像素点集合的每个像素点周围信息，并根据提取区域的置信度均值作为评价指标

，若评价指标

，则以

为最优模板尺寸，否则，令

，

，直至获取到最优滑窗尺寸；3) Build

The size of the sliding window, the sliding window is used to extract the surrounding information of each pixel point of the new first pixel point set, and the confidence average value of the extracted area is used as the evaluation index

, if the evaluation index

, then with

is the optimal template size, otherwise, let

,

, until the optimal sliding window size is obtained;

初始值设置为5，

表示以

尺寸模板提取信息后获取的评价指标，

为评价指标阈值，在本申请中设置为0.2。In the above steps,

The initial value is set to 5,

means with

The evaluation index obtained after the information is extracted from the size template,

For the evaluation index threshold, it is set to 0.2 in this application.

4) Extracting a region of the medical image to be detected according to the optimal sliding window size to obtain a region of interest with the same size as the optimal sliding window.

Specifically, the extraction of the region of interest of the present invention can use a neural network model to extract features. Since the neural network model extraction of features is in the prior art, it will not be described here. It should be noted that, by determining the optimal sliding window size, the present invention can limit the size of the region of interest when extracting the region of interest.

，则以相同尺寸模板基于像素点坐标

，

，

，

，

，

，

，

为中心，在CT图像中提取八邻域图像；同时对八邻域图像进行Concat后，得到八通道邻域图像。Based on the extracted region of interest, the eight neighborhood images of the region of interest are obtained. Specifically, the coordinates of the center pixel of the extracted region of interest are:

, then use the same size template based on pixel coordinates

,

,

,

,

,

,

,

As the center, the eight neighborhood images are extracted from the CT image; at the same time, after the eight neighborhood images are Concat, the eight channel neighborhood images are obtained.

The acquisition of the reference image in this embodiment is as follows: the reference image is output by taking the eight-channel neighborhood image as the input of the network prediction model.

The training of the network prediction model in the above is as follows: obtaining multiple real diseased medical images and non-lesioned medical images, randomly selecting the center point of each medical image, and obtaining the center image and eight neighborhood images according to the optimal sliding window size, The central image is used as the labeled data, and the eight-neighborhood image is used as the input training sample of the network prediction model, and the network prediction model is trained.

The method for acquiring the second confidence map corresponding to the reference medical image in this embodiment is the same as the method for preprocessing in step 1, and details are not repeated here.

Step 3, extract the first part of the confidence map corresponding to the eight-neighborhood image of the medical image to be tested and the second part of the confidence map corresponding to the eight-neighborhood image of the reference medical image, and combine the first part of the confidence map and the second part. The confidence degree map is made difference, and the eight-neighborhood confidence degree change map is obtained; the eight-neighborhood confidence degree change map is binarized, the pixels greater than 0 are set to 1, and the pixel points with the pixel value of 1 are processed. Curve fitting, and N straight lines are obtained, where N is greater than or equal to 1;

The partial confidence map for extracting the corresponding image in the above steps can be directly determined according to the pixel points in step 2 and the optimal sliding window size, that is, through the comparison between the image and the confidence map, the corresponding position of the pixel point is extracted. Confidence can be formed by forming a partial confidence map.

The difference between the first part of the confidence map and the second part of the confidence map in this embodiment is the difference of the confidence of the corresponding pixel points in the two images, and then the confidence change value is obtained.

It should be noted that the above steps use the eight-neighborhood confidence change map of the eight-neighborhood image as a reference factor, because the region of interest may be a tumor body or a non-tumor body. When it is a non-tumor body, its neighborhood image The information is usually the same as the confidence level of the region of interest, but if it is a tumor, the region of interest affects its neighborhood image information, that is, the neighborhood image information changes. Whether the region of interest is a tumor type.

In this embodiment, the acquisition process of the fitted N straight lines is:

1) Perform straight line fitting on the pixel points with a pixel value of 1 in the eight-neighborhood confidence change graph to obtain the first straight line;

2) Calculate the distance from each pixel to the first straight line, and obtain the average distance of all pixels. When the average distance is greater than or equal to the fit threshold, extract the pixel with the largest distance in turn, until the remaining The fitting degree of the pixel points is less than the fitting degree threshold, and the number of the extracted pixel points is counted. When the number is greater than the set number, a straight line is performed on the extracted pixel points to obtain a second straight line. , and calculate the corresponding distance mean, determine whether it is greater than or equal to the fit threshold, and so on, until all the pixels of the fitted lines meet the fit threshold, then N straight lines are obtained.

设置为5，当然其可以根据实际情况进行确定。The set number in the above is 3; the fit threshold is

Set to 5, of course, it can be determined according to the actual situation.

Step 4: Determine whether the N straight lines have an intersection in the region of interest. If there is an intersection and the confidence of the intersection is greater than the set threshold, then the intersection is a lesion point, and the medical image to be tested is a pathological image; otherwise , then there is no lesion, and the medical image to be tested is a normal image.

In this embodiment, the lesion points are screened according to the gradient of the confidence change value of the pixel points included in the fitted straight line. If the gradient from the intersection point in the region of interest shows a downward trend, the intersection point is retained, otherwise, the intersection point is deleted. The downward trend of the gradient from the intersection point in the region of interest to the outside is from the region of interest to the outside, and it is judged whether the gradient of the confidence change value of the formed intersection point has a downward trend.

Further, in order to be more accurate, this embodiment also includes modifying the first confidence level corresponding to each pixel point in the region of interest according to the intersection point:

，

表示第

个交点，

表示第

个像素点；1) Calculate the Euclidean distance between the intersection and each pixel as

,

means the first

a point of intersection,

means the first

pixels;

为：2) Then the confidence correction value of each intersection point

for:

，其中

为第一置信度图中各像素点的第一置信度，

为交点总个数。Then the corrected first confidence level of each pixel is

,in

is the first confidence level of each pixel in the first confidence level map,

is the total number of intersections.

Then, the confidence level of the corrected pixel point is compared with the set threshold to determine whether the intersection point is a lesion point. At the same time, in the above steps, before comparing the confidence level of the intersection with the set threshold, it also includes normalizing the confidence level of the pixel points, that is, normalizing the corrected pixel values of all pixel points in the region of interest. After normalization, pixels with a normalized value greater than or equal to 0.5 are extracted as lesion pixels.

Based on the above determined lesion pixel points, the present invention can also locate the tumor position. Specifically, a category image is obtained according to the determined lesion point and the corrected first confidence level map, and a connected domain is performed on the category graphic. A number of connected domains are obtained by analysis, and the connected domains whose number of pixels is less than the set number are filtered out, and the center point of the bounding box of the remaining connected domains is used as the positioning point.

The set data in this embodiment is 5; the category image in the above is obtained by assigning the pixel point corresponding to the lesion point as the lesion category to obtain the category image.

The scheme of the present invention obtains the confidence map before and after replacement through the replacement operation, additionally analyzes the confidence change of the neighborhood of the target area, and assists the pixel point confidence in the target area based on the trend of the neighborhood confidence change to a certain degree. The correction improves the accuracy of judgment, reduces the probability of misjudgment and misjudgment, and can accurately detect and locate cerebral aneurysm CT images.

It should be noted that: the above-mentioned order of the embodiments of the present invention is only for description, and does not represent the advantages or disadvantages of the embodiments. And the foregoing describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve desirable results. Additionally, the processes depicted in the figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from other embodiments.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (8)

1. A medical image cerebral aneurysm detection method based on artificial intelligence is characterized by comprising the following steps:

acquiring a medical image to be detected, and preprocessing the medical image to be detected to obtain a first confidence map;

extracting an interested area of the medical image to be detected, and obtaining an eight-neighborhood image of the interested area according to the interested area; combining the eight neighborhoods to obtain eight-channel neighborhood images, and inputting the eight-channel neighborhood images into a network prediction model to obtain corresponding reference images; replacing the pixel points of the interested region in the medical image to be detected with the pixel points of the reference image to obtain a reference medical image, and preprocessing the reference medical image to obtain a second confidence map;

extracting a first partial confidence map corresponding to an eight-neighborhood image of a medical image to be detected and a second partial confidence map corresponding to an eight-neighborhood image of a reference medical image, and subtracting the first partial confidence map and the second partial confidence map to obtain an eight-neighborhood confidence change map; carrying out binarization processing on the eight-neighborhood confidence coefficient change map, setting the pixel point greater than 0 to be 1, and carrying out curve fitting on the pixel point with the pixel value of 1 to obtain N fitted straight lines, wherein N is greater than or equal to 1;

judging whether the N straight lines have intersection points in the region of interest, if the intersection points exist and the confidence coefficient of the intersection points is greater than a set threshold value, the intersection points are pathological change points, and the medical image to be detected is a pathological image; otherwise, if no lesion point exists, the medical image to be detected is a normal image.

2. The method for detecting cerebral aneurysm based on artificial intelligence medical image according to claim 1, wherein the method for extracting the region of interest of the medical image to be detected is as follows:

selecting the first k pixels with high confidence coefficients in the first confidence coefficient image to form a first pixel point set;

calculating the Euclidean distance between any two pixels in the first pixel set, and when the Euclidean distance is smaller than a preset threshold value, combining the two pixels into a single pixel, wherein the pixel value of the single pixel is the pixel value of any one of the two pixels; supplementing the pixels which do not form the set with the first pixel set, and continuing to perform merging judgment until the number of elements in the first pixel set is k, so as to obtain a new first pixel set;

construction of

And a sliding window with the size, extracting the surrounding information of each pixel point of the new first pixel point set by adopting the sliding window, and taking the confidence coefficient mean value of the extracted area as an evaluation index

If, if

Then to

For optimum sliding window size, otherwise, order

To do so by

Update by self-adding 2 mode

，

Is an initial value until an optimal sliding window size is obtained, wherein

Is shown in

The evaluation index obtained after the information is extracted by the size template,

is an evaluation index threshold value;

and extracting the region of the medical image to be detected according to the optimal sliding window size to obtain the region of interest with the same size as the optimal sliding window.

3. The method for detecting cerebral aneurysm based on artificial intelligence medical image of claim 2, wherein the training process of the network prediction model is as follows:

acquiring a plurality of real pathological medical images and non-pathological medical images, randomly selecting the central point of each medical image, acquiring a central image and eight neighborhood images according to the optimal sliding window size, taking the central image as annotation data, and taking the eight neighborhood images as input training samples of a network prediction model to train the network prediction model.

4. The method for detecting cerebral aneurysm based on artificial intelligence medical image of claim 1, wherein the obtaining procedure of the fitted N straight lines is as follows:

respectively performing straight line fitting on pixel points with pixel values of 1 in the eight-neighborhood confidence coefficient change image to obtain a first straight line;

respectively calculating the distance from each pixel point to the first straight line, obtaining the distance mean value of all the pixel points, when the distance mean value is larger than or equal to the fitting degree threshold value, sequentially extracting the pixel points with the largest distance until the fitting degree of the remaining pixel points is smaller than the fitting degree threshold value, counting the number of the extracted pixel points, when the number is larger than the set number, performing straight line fitting on the extracted pixel points to obtain a second straight line, calculating to obtain the corresponding distance mean value, judging whether the fitting degree threshold value is larger than or equal to the fitting degree threshold value, and repeating until the pixel points of all the fitted straight lines meet the fitting degree threshold value, and obtaining N straight lines.

5. The method of claim 1, wherein the lesion points are selected according to a gradient of confidence level change values of pixel points included in the fitted straight line, and if the gradient of the intersection point in the region of interest is decreasing, the intersection point is retained, otherwise, the intersection point is deleted.

6. The method according to claim 1 or 5, further comprising modifying the first confidence level of each pixel point in the region of interest according to the intersection:

1) calculating Euclidean distance between the intersection point and each pixel point to be set as

，

Is shown as

The number of the intersection points is equal to or greater than the number of the intersection points,

is shown as

Each pixel point;

2) confidence correction value of each intersection

Comprises the following steps:

the corrected first confidence of each pixel point is

Wherein

Is the first confidence of each pixel point in the first confidence map,

the total number of the intersection points.

7. The method of claim 6, further comprising normalizing the confidence levels of the pixels before comparing the confidence level of the intersection with a predetermined threshold.

8. The method for detecting cerebral aneurysm based on artificial intelligence medical imaging of claim 7, wherein a category image is obtained according to the determined lesion point and the corrected first confidence map, connected domain analysis is performed on the category image to obtain a plurality of connected domains, the connected domains with the number of pixels smaller than the set number are screened out, and the central point of the bounding box of each remaining connected domain is used as a positioning point.

Images