A method for obtaining coronary microvascular disease blood flow reserve based on improved Unet

Abstract

The application discloses a kind of coronary microvascular disease blood flow reserve acquisition methods based on improved Unet, belong to medical technical field, including the following steps: step S1: obtaining coronary angiography sequence data;Step S2: select the angiography main blood vessel section that needs to be analyzed, utilize the segmented model of depth learning trained to main blood vessel, obtain four frame segmentation images of angiographic agent start filling blood vessel section and full blood vessel section under resting and hyperemia state;Step S3: the image after segmentation is extracted blood vessel center line and based on topological structure limit correction breakpoint, generate continuous and accurate blood vessel center line;Step S4: calculate main blood vessel center line length, according to the frame number used by angiographic agent full blood vessel in sequence calculate blood flow velocity;Step S5: the ratio between blood flow velocity under hyperemia and resting state is obtained by calculation coronary blood flow reserve.The application can be quickly and accurately obtained coronary blood flow reserve based on improved UNet segmentation network only through coronary angiography sequence, can effectively simplify detection step and improve detection efficiency.

Description

Technical Field

[0001] This invention belongs to the field of medical technology, specifically relating to a method for obtaining blood flow reserve in coronary microvascular diseases based on an improved Unet. Background Technology

[0002] Coronary microvascular disease (CMVD) refers to a clinical syndrome caused by abnormalities in the small coronary arteries. Patients may not show signs of coronary artery obstruction on coronary angiography, but they still experience symptoms of myocardial ischemia. Diagnosis of CMVD often relies on a dual assessment of coronary angiography and coronary flow reserve (CFR). However, current clinical testing procedures are cumbersome and difficult; therefore, finding a rapid and convenient way to obtain CFR is a pressing issue.

[0003] Currently, the main methods for obtaining coronary flow reserve include: 1. Invasive detection methods using Doppler guidewires as the gold standard, which are expensive and involve invasive procedures; 2. Non-invasive methods using PET / CT as the gold standard, which require sophisticated equipment and are also expensive; 3. Transthoracic ultrasound for coronary flow reserve, which is safe, non-invasive, and has good accuracy and repeatability, but requires a high level of skill from the ultrasound physician and specific ultrasound equipment. Furthermore, it can currently only measure blood flow in the mid-segment and distal segment of the left anterior descending artery. Due to anatomical location, it is difficult to measure blood flow in the left main coronary artery, circumflex artery, and right coronary artery. In summary, obtaining coronary flow reserve requires prior coronary angiography to rule out coronary artery obstruction, making the process cumbersome and hindering the rapid identification of coronary microvascular diseases.

[0004] Therefore, there is a need for a simplified detection process based on an improved Unet method for obtaining blood flow reserves in coronary microvascular disease. Summary of the Invention

[0005] The purpose of this invention is to provide a method for obtaining coronary microvascular blood flow reserve based on an improved Unet, which can quickly and accurately obtain coronary blood flow reserve using only coronary angiography sequences, effectively simplifying the detection steps while ensuring detection accuracy.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for obtaining blood flow reserve in coronary microvascular disease based on an improved Unet includes the following steps:

[0008] Step S1: Acquire coronary angiography sequence data, wherein the coronary angiography sequence includes images of the same body position at rest and under congested state;

[0009] Step S2: Select the angiographic main vessel segment to be analyzed, and use the trained deep learning segmentation model to segment the main vessel to obtain four segmented images of the vessel segment at rest and when the contrast agent begins to fill the vessel segment and when the vessel segment is full in the hyperemia state.

[0010] Step S3: Use a thinning algorithm to extract the vessel centerline of the segmented image and correct the breakpoints based on topological constraints to generate continuous and accurate vessel centerlines;

[0011] Step S4: Calculate the length of the main vessel centerline and calculate the bleeding flow velocity based on the number of frames required for the contrast agent to fill the vessel in the sequence;

[0012] Step S5: Obtain coronary blood flow reserve by calculating the ratio between blood flow velocity under congested and resting conditions.

[0013] A further improvement of the technical solution of the present invention is that: in step S2, the selectable vascular segment includes any vascular segment of the anterior descending branch, the circumflex branch, and the right coronary artery.

[0014] A further improvement to the technical solution of this invention lies in the fact that the segmentation model adopts an improved UNet3Plus network structure, and the steps for segmenting the main blood vessel using the improved UNet3Plus network structure are as follows:

[0015] Step S201: Perform preliminary feature extraction and contextual feature association on the encoder composed of five improved HANC convolutional blocks;

[0016] Step S202: After multi-scale feature fusion, the extracted deep feature map is passed to the subsequent decoder to generate high-resolution prediction;

[0017] Step S203: Use the Sigmoid layer to segment the main blood vessel.

[0018] A further improvement of the technical solution of the present invention is as follows: In step S201, the encoder is implemented based on the context feature enhancement module, which includes four HANC modules. Each pair of modules contains a MaxPooling layer. In the HANC module, feature channels are first added through a 1x1 Conv+BN+LeakyReLU structure, and then a 3×3 depth Conv+BN+LeakyReLU structure is used for compensation to reduce the complexity of the model. Then, the mean and maximum values ​​of the features of adjacent pixels are added to provide an approximate concept of neighborhood comparison. The Maxpooling layer and the Average pooling layer are used to simulate context information and the 1x1 Conv+BN+LeakyReLU structure is used to capture the long-distance dependency of context information. Finally, it is added to the input feature map, and adaptive feature combination is achieved through the 1x1 Conv+BN+LeakyReLU structure and the channel attention module.

[0019] A further improvement of the technical solution of this invention lies in the following: In step S202, the multi-scale feature fusion module adjusts the size of the convolutional feature maps obtained from different encoder levels. First, it uses Maxpooling or Upsample to shrink or enlarge the encoder features, and then adjusts them to the required number of feature channels using a 3X3Conv+BN+ReLU structure, making them equal in size and connecting them before sending them into the decoder. This aggregates information from multiple encoder levels and enriches the feature map of a single encoder. The module is shown below:

[0020]

[0021] A further improvement to the technical solution of the present invention is that, in step S203, the segmentation model is composed of cross-entropy loss L. ce Similarity coefficient loss L with Dice Dice Combination loss function L union For deep supervised training, the joint loss function L used in the segmentation model is... union As shown below:

[0022]

[0023] In the formula, y i and These represent the true value and the predicted result, respectively, and M represents the number of training samples.

[0024] A further improvement to the technical solution of the present invention is as follows: In step S3, a skeleton extraction algorithm is used to extract the blood vessel centerline with a width of only 1 pixel from the segmented image, and the breakpoints of the discontinuous blood vessel centerlines are corrected. The steps for correcting the breakpoints of the blood vessel centerlines are as follows:

[0025] Step S301: Use morphological operations to calculate and analyze the connectivity of the vessel centerline image to obtain breakpoints;

[0026] Step S302: Reconnect the breakpoints based on the distance of the breakpoints and the angle of the blood vessel travel, thereby generating a continuous and accurate blood vessel centerline.

[0027] A further improvement to the technical solution of the present invention is that, in step S302, the distance calculation formula for the breakpoint is as follows:

[0028]

[0029] In the formula, (a i1 ,a i2 ) and (a j1 ,a j2 ) represent the coordinates of the breakpoints obtained by analyzing the connectivity of the blood vessel centerline image;

[0030] The physiologically reasonable angle for the vascular centerline is set between 135° and 180°, and the formula for limiting the vascular angle is shown below:

[0031]

[0032] In the formula, Q and A1 represent the nodes of the central line of the main blood vessel and the central line of the disconnected blood vessel, respectively.

[0033] Fit a quadratic polynomial curve y = ax to four nodes 2 +bx+c is used for reconnection to ensure that the center line of the connected blood vessels exhibits a smooth curve trend under natural morphology.

[0034] A further improvement to the technical solution of the present invention is that the calculation steps for blood flow velocity in step S4 are as follows:

[0035] Step S401: Calculate the length of the vessel segment in the image when the contrast agent begins to be injected and the length of the vessel segment in the image when the contrast agent fills the vessel, respectively, based on the TIMI frame count method;

[0036] Step S402: Calculate the blood flow velocity of the current blood vessel segment based on the frame difference between the two images. The formula for calculating the blood flow velocity is as follows:

[0037]

[0038] In the formula, L end L start F represents the length at the end of the vessel segment frame and the length at the beginning of the vessel segment frame, respectively. end F start These represent the end frame number and the start frame number of the blood vessel segment, respectively.

[0039] A further improvement of the technical solution of the present invention is that the coronary blood flow reserve obtained in step S5 maintains a certain correlation with the value of coronary blood flow reserve detected by transthoracic ultrasound in actual clinical practice, wherein the calculation formula for verifying the correlation index is as follows:

[0040]

[0041] In the formula, σ X They represent X respectively i Sample standard score, sample mean, and sample standard deviation, X i With Y i These represent the real sample and the predicted sample, respectively.

[0042] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0043] This invention provides a method for obtaining coronary microvascular blood flow reserve based on an improved UNet segmentation network. This method can quickly and accurately obtain coronary blood flow reserve using only a single coronary angiography sequence, without the need for coronary angiography to screen for signs of coronary artery obstruction. This effectively simplifies the detection process and improves detection efficiency, thereby assisting doctors in quickly and accurately identifying coronary microvascular diseases.

[0044] This invention improves the UNet model by replacing the basic double convolutional block with the Context Long Distance Modeling (HANC) module, which enhances the model's ability to capture long-distance dependencies. Compared to conventional deep learning segmentation methods that can only capture local features, this method captures long and narrow tubular structures such as blood vessels more effectively and comprehensively.

[0045] The multi-scale fusion module used in this invention can fuse features of different spatial scales and send them to the decoder to collect contextual information of different scales. This can effectively distinguish between the main blood vessel and other small blood vessels, making the model focus more on the main blood vessel that needs to be segmented.

[0046] The centerline extraction and correction module used in this invention ensures the connectivity of the extracted main blood vessel through topological constraints.

[0047] The vascular segment used in this invention includes any segment of the anterior descending artery, the circumflex artery, and the right coronary artery. Compared with the whole vessel calculation, it effectively avoids the problem of inaccurate average velocity calculation caused by the influence of contrast agent injection speed at the beginning of the vessel.

[0048] Based on the morphological characteristics of blood vessels, this invention sets the physiologically reasonable angle of the blood vessel centerline between 135° and 180°, so that the broken part is more in line with the physiological structure of blood vessels and is closer to the real situation. Attached Figure Description

[0049] Figure 1 This is the overall framework roadmap of the present invention;

[0050] Figure 2 This is a schematic diagram of the network structure of the segmentation model in this invention;

[0051] Figure 3 This is a schematic diagram of the HANC module of the segmentation model in this invention;

[0052] Figure 4 This is a schematic diagram of the multi-scale fusion module of the segmentation model in this invention, taking HANC Block6 as an example;

[0053] Figure 5 These are the original image and the result image before the segmentation model prediction in this invention;

[0054] Figure 6 This is a diagram showing the centerline extraction result in this invention;

[0055] Figure 7 This is a diagram showing the centerline correction result in this invention;

[0056] Figure 8 This is a schematic diagram illustrating the correlation between the predicted and actual blood flow reserve values ​​in this invention. Detailed Implementation

[0057] The present invention will be further described in detail below with reference to embodiments:

[0058] like Figure 1 As shown, a method for obtaining blood flow reserve in coronary microvascular disease based on an improved Unet includes the following steps:

[0059] Step S1: Acquire the patient's coronary angiography sequence data, which includes images of the patient in the same position at rest and under congested conditions;

[0060] Step S2: Select the angiographic main vessel segment to be analyzed. Selectable vessel segments include any segment of the left anterior descending artery, circumflex artery, and right coronary artery. Compared with whole vessel calculation, this effectively avoids the problem of inaccurate average velocity calculation caused by the injection speed of contrast agent at the beginning of the vessel. The trained deep learning segmentation model is used to segment the main vessel, and four segmented images are obtained in the resting state, the congested state, the segment where the contrast agent begins to fill the vessel, and the segment that is full of the vessel.

[0061] Step S3: Use a thinning algorithm to extract the vessel centerline of the segmented image and correct the breakpoints based on topological constraints to generate continuous and accurate vessel centerlines;

[0062] Step S4: Calculate the length of the main vessel centerline and calculate the bleeding flow velocity based on the number of frames required for the contrast agent to fill the vessel in the sequence;

[0063] Step S5: Obtain coronary blood flow reserve by calculating the ratio between blood flow velocity under congested and resting conditions.

[0064] Specifically:

[0065] In step S2, the network structure of the entire segmentation model is as follows: Figure 2 As shown, this segmentation model uses an improved UNet3Plus network structure. The steps for segmenting the main blood vessel using the improved UNet3Plus network structure are as follows:

[0066] Step S201: Perform preliminary feature extraction and contextual feature association on the encoder composed of five improved HANC convolutional blocks;

[0067] Step S202: After multi-scale feature fusion, the extracted deep feature map is passed to the subsequent decoder to generate high-resolution prediction;

[0068] Step S203: Segment the main blood vessel using the Sigmoid layer;

[0069] The encoder is implemented based on the context feature enhancement module, such as... Figure 3 As shown, the context feature enhancement module contains four HANC modules, with a MaxPooling layer between each pair of modules. Within the HANC modules, a 1x1 Conv+BN+LeakyReLU structure is first used to increase feature channels, followed by a 3×3 deep Conv+BN+LeakyReLU structure for compensation, reducing model complexity. Then, the mean and maximum values ​​of neighboring pixel features are appended to provide an approximate concept of neighborhood comparison. Maxpooling and Average pooling layers are used to simulate contextual information, and a 1x1 Conv+BN+LeakyReLU structure is employed to capture long-range dependencies of contextual information. Finally, the feature maps are added to the input feature map, and adaptive feature combination is achieved through a 1x1 Conv+BN+LeakyReLU structure and a channel attention module.

[0070] Before the encoder features enter the decoder, they are processed by a multi-scale feature fusion module, such as... Figure 4As shown, taking HANCBLOCK6 as an example, this module implements multi-level feature combination. It adjusts the size of convolutional feature maps obtained from different encoder levels. First, it uses Maxpooling or Upsampling to shrink or enlarge the encoder features, then adjusts them to the required number of feature channels using a 3x3 Conv+BN+ReLU structure, making them equal in size and concatenating them before feeding them into the decoder. This aggregates information from multiple encoder levels and enriches the feature map of a single encoder. The module is shown below:

[0071]

[0072] Finally, the decoder structure is symmetrical to the encoder structure. After generating high-resolution predictions, the main blood vessel is segmented through a sigmoid layer. The segmented image is shown below. Figure 5 As shown;

[0073] During the training of this model, the segmentation model is determined by cross-entropy loss L. ce Similarity coefficient loss L with Dice Dice Combination loss function L union For deep supervised training, the joint loss function L used in the segmentation model is... union As shown below:

[0074]

[0075] In the formula, y i and These represent the true value and the predicted result, respectively, and M represents the number of training samples.

[0076] In step S3, the Zhang-Suen skeleton thinning algorithm is used to extract the vessel centerlines with a single pixel width of only 1 pixel from the segmented image, and the breakpoints of discontinuous vessel centerlines are corrected based on topological constraints. The centerline extraction results are as follows: Figure 6 As shown, the steps for correcting the breakpoints of the blood vessel centerline are as follows:

[0077] Step S301: Use morphological operations to calculate and analyze the connectivity of the centerline image to obtain breakpoints;

[0078] Step S302: Correct the breakpoints of the vessel centerline based on multiple factors such as the distance between breakpoints and the vessel angle, thereby generating a continuous and accurate vessel centerline. The formula for the distance between breakpoints is as follows:

[0079]

[0080] In the formula, (a i1 ,a i2 ) and (a j1 ,a j2) represent the coordinates of the breakpoints obtained by analyzing the connectivity of the blood vessel centerline image;

[0081] Based on vascular morphological characteristics, the physiologically reasonable angle of the vascular centerline is set between 135° and 180° to make the broken sections more consistent with the physiological structure of blood vessels. The formula for limiting the vascular angle is shown below:

[0082]

[0083] In the formula, Q and A1 represent the nodes of the central line of the main blood vessel and the central line of the disconnected blood vessel, respectively.

[0084] Fit a quadratic polynomial curve y = ax to four nodes 2 +bx+c is reconnected to ensure that the center line of the connected blood vessels exhibits a smooth curve trend under natural morphology.

[0085] The corrected image is as follows Figure 7 As shown.

[0086] In step S4, the blood flow velocity is calculated as follows:

[0087] Step S401: Calculate the length of the vessel segment in the image when the contrast agent begins to be injected and the length of the vessel segment in the image when the contrast agent fills the vessel, respectively, based on the TIMI frame count method;

[0088] Step S402: Calculate the blood flow velocity of the current blood vessel segment based on the frame difference between the two images. The specific formula for blood flow velocity is as follows:

[0089]

[0090] In the formula, L end L start F represents the length at the end of the vessel segment frame and the length at the beginning of the vessel segment frame, respectively. end F start These represent the end frame number and the start frame number of the blood vessel segment, respectively.

[0091] In step S5, the specific formula for obtaining coronary flow reserve is as follows:

[0092]

[0093] In the formula, V 充血 V 静息 These represent the blood flow velocities calculated under congested and resting states, respectively.

[0094] The obtained coronary flow reserve maintains a certain correlation with the values ​​measured by transthoracic ultrasound in actual clinical practice. The formula for verifying the correlation index is shown below:

[0095]

[0096] In the formula, and σ X They represent X respectively i Sample standard score, sample mean, and sample standard deviation, X i With Y i These represent the real sample and the predicted sample, respectively.

[0097] Example:

[0098] This invention used 447 coronary angiography images as the training set and 47 images as the validation set to train and validate the main vessel segmentation model. The coronary blood flow reserve was then obtained from the segmented images and compared with the actual values ​​for validation. Specifically, 100 coronary angiography image sequences were tested, and the obtained coronary blood flow reserve was compared with the actual values ​​measured by doctors after surgery. The correlation is shown below. Figure 8 As shown, by Figure 8 It can be seen that the correlation between the predicted value and the actual value is 0.65, which can effectively assist doctors in quickly and accurately identifying coronary microvascular diseases.

[0099] In summary, this invention, based on an improved UNet segmentation network, can quickly and accurately obtain coronary blood flow reserve using only a single coronary angiography sequence, without the need for coronary angiography to screen for signs of coronary artery obstruction. This effectively simplifies the detection process and improves detection efficiency, thereby assisting doctors in quickly and accurately identifying coronary microvascular diseases.

[0100] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for obtaining blood flow reserve in coronary microvascular disease based on an improved Unet, characterized in that: Includes the following steps: Step S1: Acquire coronary angiography sequence data, wherein the coronary angiography sequence includes images of the same body position at rest and under congested state; Step S2: Select the angiographic main vessel segment to be analyzed, and use the trained deep learning segmentation model to segment the main vessel, obtaining four segmented images: the segment where the contrast agent begins to fill the vessel in the resting state, the segment where the vessel is full, and the segment where the contrast agent begins to fill the vessel in the congested state. The segmentation model uses an improved UNet3Plus network structure. The steps for segmenting the main vessel using the improved UNet3Plus network structure are as follows: Step S201: Perform preliminary feature extraction and association with context features on the encoder composed of five improved HANC convolutional blocks. The encoder is implemented based on the context feature enhancement module, which contains four HANC modules. Each pair of HANC modules contains a MaxPooling layer. In the HANC module, feature channels are first added through a 1x1 Conv+BN+LeakyReLU structure, and then a 3×3 depth Conv+BN+LeakyReLU structure is used for compensation to reduce the complexity of the model. Then, the mean and maximum values ​​of the features of adjacent pixels are added to provide an approximate concept of neighborhood comparison. The Maxpooling layer and Average pooling layer are used to simulate context information and the 1x1 Conv+BN+LeakyReLU structure is used to capture the long-distance dependencies of context information. Finally, it is added to the input feature map. Adaptive feature combination is achieved through the 1x1 Conv+BN+LeakyReLU structure and the channel attention module. Step S202: After multi-scale feature fusion, the extracted deep feature map is passed to the subsequent decoder to generate high-resolution prediction; Step S203: Use the Sigmoid layer to segment the main blood vessel; Step S3: Use a thinning algorithm to extract the blood vessel centerline from the segmented image and correct the breakpoints based on topological constraints to generate a continuous and accurate blood vessel centerline. Step S4: Calculate the length of the main vessel centerline and calculate the bleeding flow velocity based on the number of frames required for the contrast agent to fill the vessel in the sequence; Step S5: Obtain coronary blood flow reserve by calculating the ratio between blood flow velocity under congested and resting conditions.

2. The method for obtaining coronary microvascular blood flow reserve based on an improved Unet according to claim 1, characterized in that: In step S2, the vessel segment includes any segment of the anterior descending branch, the circumflex branch, and the right coronary artery.

3. The method for obtaining coronary microvascular blood flow reserve based on an improved Unet according to claim 2, characterized in that: In step S202, the multi-scale feature fusion module adjusts the size of the convolutional feature maps obtained from different encoder levels. First, it uses Maxpooling or Upsampling to shrink or enlarge the encoder features. Then, it uses a 3x3 Conv+BN+ReLU structure to adjust the number of feature channels to the required number, making them equal in size and concatenating them before sending them to the decoder. This aggregates information from multiple encoder levels and enriches the feature map of a single encoder. The feature transformation process in the multi-scale feature fusion module is as follows: 。 4. The method for obtaining coronary microvascular blood flow reserve based on an improved Unet according to claim 3, characterized in that: In step S203, the segmentation model is determined by cross-entropy loss. Similarity coefficient loss with Dice Combination loss function The joint loss function used in the segmentation model for deep supervised training. As shown below: In the formula, and These represent the true value and the predicted result, respectively, and M represents the number of training samples.

5. The method for obtaining coronary microvascular flow reserve based on an improved Unet according to claim 1, characterized in that: In step S3, a skeleton extraction algorithm is used to extract the vessel centerline with a width of only 1 pixel from the segmented image, and the breakpoints of the discontinuous vessel centerline are corrected. The steps for correcting the breakpoints of the vessel centerline are as follows: Step S301: Use morphological operations to calculate and analyze the connectivity of the vessel centerline image to obtain breakpoints; Step S302: Reconnect the breakpoints based on the distance of the breakpoints and the angle of the blood vessel travel, thereby generating a continuous and accurate blood vessel centerline.

6. The method for obtaining coronary microvascular flow reserve based on an improved Unet according to claim 5, characterized in that: In step S302, the distance to the breakpoint is calculated using the following formula: In the formula, and These represent the coordinates of the breakpoints obtained from analyzing the connectivity of the blood vessel centerline image; The physiologically reasonable angle for the vascular centerline is set between 135° and 180°, and the formula for limiting the vascular angle is shown below: In the formula, Q, and P and P represent the nodes of the central line of the main blood vessel and the central line of the disconnected blood vessel, respectively. Fit a quadratic polynomial curve to four nodes Reconnection is performed to ensure that the centerline of the connected blood vessels exhibits a smooth curve trend as it naturally appears.

7. The method for obtaining coronary microvascular blood flow reserve based on an improved Unet according to claim 1, characterized in that: In step S4, the blood flow velocity is calculated as follows: Step S401: Calculate the length of the vessel segment in the image when the contrast agent begins to be injected and the length of the vessel segment in the image when the contrast agent fills the vessel, respectively, based on the TIMI frame count method; Step S402: Calculate the blood flow velocity of the current blood vessel segment based on the frame difference between the two images. The formula for calculating the blood flow velocity is as follows: In the formula, , These represent the length at the end of the frame and the length at the beginning of the frame, respectively. , These represent the end frame number and the start frame number of the blood vessel segment, respectively.

8. The method for obtaining coronary microvascular blood flow reserve based on an improved Unet according to claim 1, characterized in that: The coronary flow reserve obtained in step S5 maintains a certain correlation with the values ​​measured by transthoracic ultrasound in actual clinical practice. The formula for calculating the correlation index is shown below: In the formula, , , They represent respectively to The standard score, sample mean, and sample standard deviation of the sample. and These represent the real sample and the predicted sample, respectively.

Images