Method for full-automatic blood vessel segmentation and stenosis detection from angiography image

Abstract

Coronary artery disease is the most common disease in the cardiovascular field and is also one of important factors causing death in the global range, and by conducting blood vessel segmentation and stenosis detection on coronary angiography images, doctors can diagnose the coronary heart disease more accurately. The invention provides a brand new deep learning architecture SAM-VMNet, and the model combines the powerful feature extraction capability of MedSAM and the linear complexity advantage of a visual state space model of VM-UNet, so that the model has higher reasoning speed and stronger data processing capability than Transformer, and higher segmentation accuracy and stability of coronary angiography images are realized. Moreover, a'dynamic queue 'is used for capturing the narrow part of the blood vessel, and the narrow part is compared with the normal blood vessel diameter before and after the narrow part, so that quantitative calculation of the blood vessel stenosis degree is realized. The algorithm is high in operation speed, and full-automatic blood vessel contour extraction and stenosis degree detection can be achieved. The kit can be applied to the clinical diagnosis process of coronary heart disease.

Description

Technical Field

[0001] The invention relates to a method for blood vessel segmentation and stenosis detection in angiography images, in particular to a fully automatic blood vessel segmentation and stenosis detection method, and belongs to the field of medical image processing. Background Art

[0002] Coronary artery disease is the most common cardiovascular disease and one of the most common causes of death in the world. According to the latest data released by the World Health Organization (WHO), cardiovascular disease is one of the four major non-communicable chronic diseases and the disease with the largest increase in deaths. Since coronary angiography images are relatively complex, it is difficult to obtain clear subtraction images like brain and limb angiography images. Therefore, it is crucial to extract a clear and complete vascular tree structure from coronary angiography images.

[0003] At present, there are many classic methods for arterial angiography image segmentation, such as threshold segmentation, edge detection, region growing, and tracking-based methods. However, since arterial angiography images usually have low contrast and complex background, traditional algorithms may be difficult to meet clinical needs in practical applications.

[0004] Deep learning methods have improved the accuracy and robustness of arterial angiography image segmentation, such as convolutional neural networks, Transformer models, and the emerging Mamba model. Among them, VM-Unet, which is based on the Mamba model, is more efficient in processing large-scale medical image data. The model uses a state space model to capture a wide range of contextual information and shows strong competitiveness in medical image segmentation tasks.

[0005] Furthermore, in terms of vascular stenosis detection, the centerline extraction is first required. First, the centerline is obtained through the vascular segmentation map through erosion operation. Then the vascular diameter measurement is required. Commonly used vascular diameter estimation methods include edge intersection method, polynomial fitting method, Gaussian curve fitting method, MOM method, etc. Finally, the degree of vascular stenosis is determined by comparing the diameter of the current point with the diameter of the previous and next vascular segments to calculate the specific degree of stenosis. Summary of the invention

[0006] Coronary angiography images are regarded as the authoritative standard for the diagnosis of coronary heart disease. By performing vascular segmentation and stenosis detection on images, doctors can diagnose coronary heart disease more accurately. The present invention establishes a deep learning framework for accurately extracting vascular trees and detecting the degree of vascular stenosis. The model combines the advantages of MedSAM and VM-UNet to solve the problems of slow reasoning speed and low segmentation accuracy of existing deep learning methods. In terms of stenosis detection, the first detection method using dynamic queues to capture stenotic blood vessels has higher accuracy in determining the degree of vascular stenosis.

[0007] The technical solution of the present invention is: extracting a blood vessel tree based on a VM-UNet model, comprising: step 1, denoising an image using ROF and CLAHE methods.

[0008] Step 2: First, train a simple VM-UNet to perform rough segmentation on the image.

[0009] Step 3: Select 10 points at equal intervals on the mask image as prompts and feed them back to MedSAM, so as to automatically obtain the feature vector output by the MedSAM encoder.

[0010] In step 4, the feature vectors of the two encoders are concatenated and then sent to the decoder of VM-UNet. The weights are continuously updated by back propagation to finally obtain the output mask image.

[0011] Detection of vascular stenosis based on a dynamic cohort, including:

[0012] Step 1: Use the BFS algorithm to divide the blood vessels into multiple paths

[0013] Step 2: Create an empty queue, search for a sequence of points with decreasing radius on each path, and record the radius of the head of the queue.

[0014] Step 3: When the radius of the point sequence starts to increase, record the minimum value of the radius.

[0015] Step 4: Subsequent point sequences are continuously queued until the radius of the last point in the queue is larger than the next point to be queued. The process ends and the degree of narrowness is calculated.

[0016] Step 5, repeat the above steps and continue traversing the blood vessel path until all blood vessels are traversed.

[0017] Compared with the prior art, the advantages of the present invention are mainly reflected in the following aspects:

[0018] First of all, the present invention is more perfect in the process design of blood vessel extraction and stenosis detection. Through the optimization of deep learning algorithms, we can identify and extract the vascular contours of coronary arteries with higher accuracy, thus providing doctors with more accurate and reliable diagnostic indicators.

[0019] Secondly, the present invention innovatively combines the feature extraction capability of MedSAM with the visual state space model of VM-UNet. This combination not only takes advantage of the powerful feature extraction capability of MedSAM, but also takes advantage of the linear complexity of the VM-UNet model. Compared with the traditional Vision Transformer, the present invention has faster reasoning speed and stronger data processing capability, thus achieving significant improvements in the segmentation accuracy and stability of vascular images. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The workflow of the blood vessel segmentation and stenosis detection method proposed in the present invention is shown in the following figure (abstract figure). (a) Artery segmentation is performed using a deep learning-based method; (b) Stenosis detection includes extracting the arterial centerline, calculating the arterial diameter, and determining the degree of stenosis.

[0021] Figure 2 This is the SAM-VMUNet network framework diagram of the present invention.

[0022] Figure 3 is a structural diagram of the VSSblock used in the present invention.

[0023] Figure 4 It is the image segmentation detection result of the specific blood vessel of the present invention.

[0024] Figure 5 This is a flow chart of the "dynamic queue" vascular stenosis detection method proposed by the present invention.

[0025] Figure 6 This is an example diagram of the vascular stenosis detection results obtained by the present invention. DETAILED DESCRIPTION

[0026] The network framework of SAM-VMNet is shown in the figure Figure 1 shown.

[0027] Step S101, reading a coronary artery angiography image, the data is a single two-dimensional angiography image.

[0028] Step S102: Input the contrast image into the VM-Unet image segmentation model for initial rough segmentation. First, the Patch Embedding layer transforms the input image into The image is divided into non-overlapping patches of size 4×4 and then mapped to C dimensions. This process will produce an image embedding

[0029] Step S103, before the input image is embedded into the encoder for feature extraction, Layer Normalization is used to normalize it.

[0030] Step S104, the encoder consists of four stages. Patch merging operations are applied at the end of the first three stages to reduce the height and width of the input features while increasing the number of channels. The four stages use [2, 2, 2, 2] VSS blocks respectively, and the channel count of each stage is [C, 2C, 4C, 8C].

[0031] Step S105, the decoder also consists of four stages. In the last three stages, the Patch Expanding operation is applied to reduce the number of feature channels and increase the height and width of the feature channels. In the four stages, [2, 2, 2, 1] VSS blocks (asymmetric architecture) are used, and the channel count of each stage is [8C, 4C, 2C, C].

[0032] Step S106, after the decoder, a Final Projection layer is used to restore the size of the feature to match the segmentation target. The height and width of the feature are restored by upsampling 4 times through patch expansion, and then the number of channels is restored through the projection layer.

[0033] Step S107, the loss function of VM-UNet is used for two-class and multi-class segmentation tasks, and the calculation formula is as follows:

[0034] L BceDice =λ1L Bce +λ2L Dice

[0035] L CeDice =λ1L Ce +λ2L Dice

[0036]

[0037] It is stipulated that N is the total number of samples, C is the total number of categories; y i and Represent the true label and prediction respectively; y i,c is an indicator, which is equal to 1 if sample i belongs to class c, otherwise it is equal to 0; is the probability that the model predicts that sample i belongs to category c; |X| and |Y| represent the true value and the predicted value respectively; λ1 and λ2 are the weights of the loss function, and the default values ​​are both 1.

[0038] Step S108: The image input is passed through the VM-UNet of branch 1 to obtain a coarse segmentation image. Ten points are selected at equal intervals on this image as the prompt of MedSAM to obtain the feature vector The image input passes through the VM-UNet encoder of branch 2 to obtain the feature vector

[0039] In step S109, x1 is downsampled using a 1×1 convolution and an average pooling layer to make its dimension consistent with that of x2, and linear fusion is used to perform feature fusion on x1 and x2.

[0040] In step S110, the fused vector is then upsampled by the decoder of VM-UNet, and the prediction result is finally output.

[0041] Step S111, the vascular image obtained after segmentation is subjected to an erosion operation to obtain the vascular centerline.

[0042] Step S112, determine the bifurcation point: if there are 3 neighbors around the point on the current center line, it is a bifurcation point.

[0043] Step S113, find the points on each blood vessel path by breadth-first search, and then calculate the diameter of each blood vessel.

[0044] Step S114 defines parameters: the coordinates of the center of the circle (x0, y0), the maximum radius range of the search rr, and image(x, y) represents the pixel value of the image position (x, y).

[0045] Step S115: Parameterize the circle where r is the radius of the circle and α is in the range [0,2π].

[0046] Step S116, starting from radius r=1, gradually increasing to r=rr.

[0047] Step S117, for each radius, calculate the pixel position corresponding to each angle α on the circle, and determine whether it meets a specific condition (egimage(x,y)~=255).

[0048] Step S118, when the first pixel position that meets the conditions is found, the current radius (r) is recorded as the radius (R) of the circular segment.

[0049] Step S119: If no pixel position that meets the conditions is found in the entire range, (R) is output as the radius.

[0050] After calculating the blood vessel diameter, it is necessary to detect the stenosis points on the path and quantitatively estimate the degree of stenosis.

[0051] Step S120, defining the starting point, the ending point of the initialization decreasing queue, the starting point, the ending point and the minimum radius of the increasing sequence.

[0052] Step S121, for each pair of adjacent points in the path, obtain their radius values. If the radius of the current point is greater than the radius of the next point, a narrow point sequence may begin to form. Record the starting point of the decreasing sequence, and update the current minimum radius and end point; if the radius of the current point is less than or equal to the radius of the next point, and there is a decreasing sequence that has started, it means that the narrow point sequence ends, and record the end point of the increasing sequence.

[0053] Step S122, calculate the degree of stenosis, and use the radius of the starting point and the end point of the decreasing sequence to calculate the average radius startRadius and endRadius. Then the formula for calculating the degree of stenosis is

[0054] Step S123, traverse the stored stenosis point coordinates and stenosis degrees, and output the position of each stenosis point and the corresponding stenosis degree. If 0.25<1-stenosisRatio<0.5, it is mild stenosis, 0.5≤1-stenosisRatio<0.75, it is moderate stenosis, and 0.75≤1-stenosisRatio, it is severe stenosis.

Claims

1. A fully automatic blood vessel extraction and stenosis detection method from angiography images, consisting of four key modules. These include: (1) VM-UNet feature extraction module: freeze the decoder, extract features through the core dual-branch structure VSS block in the encoder, combined with linear layers, depth-separable convolutions and 2D-Selective-Scan modules to obtain the feature vector of branch 1. (2) SAM-VMNet feature extraction module: unfreeze the decoder, input the image into VM-UNet, select the annotation points from the output image as the MedSAM prompt, obtain the feature vector, downsample through 1×1 convolution and average pooling, and obtain the feature vector of branch 2. (3) Dual-branch feature fusion module: linearly fuse the feature vectors of the two branches, upsample the fused vectors through the VM-UNet decoder, and output the final prediction result. (4) Dynamic queue stenosis detection module: dynamically capture the minimum value point of the vascular diameter and automatically mark the location of vascular stenosis.

2. The method for fully automatic blood vessel extraction and stenosis detection as described in claim 1, wherein the VM-UNet feature extraction module can maintain linear complexity while establishing long-range dependencies and capture a wide range of contextual information.

3. The method for fully automatic blood vessel extraction and stenosis detection as described in claim 1, wherein the SAM-VMNet feature extraction module can enhance the overall feature extraction capability of the model and accelerate the reasoning speed.

4. The method for fully automatic blood vessel extraction and stenosis detection as described in claim 1, wherein the dual-branch feature fusion module can combine the advantages of MedSAM and VM-UNet to obtain more comprehensive image feature information and improve the accuracy and speed of blood vessel extraction.

5. The method for fully automatic blood vessel extraction and stenosis detection as claimed in claim 1, wherein the dynamic queue stenosis detection module can detect the specific stenosis position of the blood vessel and visualize the detection process.

Images