HIFU micro-lesion detection method and system based on optical flow multi-parameter time segment analysis

Abstract

The present invention discloses a HIFU micro-lesion detection method and system based on optical flow multi-parameter time segment analysis, the method comprising: obtaining a HIFU ultrasound monitoring image sequence; preprocessing the HIFU ultrasound monitoring image sequence; segmenting and pairing the preprocessed image sequence according to a preset segmentation rule to generate multiple image pairs, each image pair containing two adjacent time-series images; performing multi-resolution optical flow calculation on the image pairs in the multiple image pairs of a multi-resolution pyramid optical flow model; extracting multi-parameter components from the final optical flow results to obtain light intensity components, grayscale change components and optical flow direction components; performing multi-parameter time segment analysis on the multi-parameter components, thereby detecting HIFU micro-lesions. The method of the present invention significantly improves the accuracy of HIFU micro-lesion detection.

Description

Technical Field

[0001] The present invention belongs to the field of HIFU micro-lesion detection, and in particular relates to a HIFU micro-lesion detection method and system based on optical flow multi-parameter time segment analysis. Background Art

[0002] As a non-invasive physical therapy, high-intensity focused ultrasound (HIFU) technology has been increasingly valued by doctors due to its advantages of non-invasiveness, non-ionization, and few complications after treatment. It has been used in the clinic to treat various types of tumors. As a minimally invasive treatment technology from in vitro to in vivo, HIFU surgery must avoid damage to normal tissues or unnecessary energy accumulation in the body caused by excessive irradiation, and must also avoid problems such as incomplete tumor ablation caused by insufficient irradiation. In current clinical practice, the most common imaging methods for monitoring HIFU treatment are magnetic resonance imaging (MRI) and ultrasound imaging. Compared with ultrasound imaging, MRI is expensive and has high maintenance costs. Ultrasound imaging has the characteristics of no radiation, real-time, and portability, and can perform dynamic imaging of the HIFU treatment process. However, the random noise during ultrasound scanning will reduce the quality of ultrasound imaging, thereby affecting the judgment of clinicians, resulting in missed detection and misjudgment of minor injuries. Therefore, it is crucial to develop a HIFU micro-damage detection method to achieve accurate monitoring during HIFU surgery and understand the dynamic changes in the biological activity of target biological tissues under high-intensity energy, in order to improve the safety and success rate of HIFU surgery.

[0003] Given the strong real-time characteristics of ultrasound imaging, there are many studies that use ultrasound images to detect tissue damage. However, due to the uniqueness of the HIFU irradiation device in clinical treatment, the irradiation probe and the B-ultrasound probe are coaxial, which results in the researchers being unable to fully record the entire process of HIFU damage when conducting damage localization research, resulting in the lack of detailed information of many intermediate processes in the later data analysis. The thermal effect during continuous HIFU irradiation can cause tissue damage. Due to the change in the acoustic channel caused by the heterogeneity of the tissue, many scattered micro-lesions will be produced; at the same time, the bubbles produced by the cavitation effect in the target area change the acoustic impedance, thereby causing the reflection of the sound wave and increasing the heat deposition in the nearby area, resulting in the forward movement and deformation of the thermal damage area. This makes it difficult for doctors to accurately observe the location of the damage point during HIFU surgery. Due to the many irregular salt and pepper noises in B-mode imaging and the irregular shape of the HIFU damage point, it is more difficult for HIFU doctors to locate the damage during treatment. Summary of the invention

[0004] In view of this, the object of the present invention is to provide a HIFU micro-lesion detection method and system based on optical flow multi-parameter time segment analysis.

[0005] One aspect of the present invention provides a HIFU micro-lesion detection method based on optical flow multi-parameter time segment analysis, the method comprising:

[0006] Acquire HIFU ultrasound monitoring image sequences;

[0007] Preprocessing the HIFU ultrasound monitoring image sequence to generate a preprocessed image sequence;

[0008] The preprocessed image sequence is segmented and paired according to a preset segmentation rule to generate multiple image pairs, each of which contains two images in adjacent time series;

[0009] Among the multiple image pairs, use the multi-resolution pyramid optical flow model to perform multi-resolution optical flow calculation on a first image pair among the multiple image pairs to obtain a first optical flow result corresponding to the first image pair;

[0010] Performing multi-resolution optical flow calculation on the second image in the first image pair and the first image in the next image pair of the second image pair to generate a first optical flow supplement increment;

[0011] Using the first optical flow result corresponding to the first image pair and the first optical flow supplement increment, the initial optical flow of the second image pair is obtained;

[0012] For each image pair except the first image pair and the last image pair among the multiple image pairs, the image pair is taken as the current image pair, and S1-S5 are executed:

[0013] S1: Calculate the multi-resolution optical flow of the current image pair using the multi-resolution pyramid optical flow model, and generate a first optical flow result corresponding to the current image pair according to the initial optical flow of the current image pair;

[0014] S2: performing multi-resolution optical flow calculation on the second image in the current image pair and the first image in the next image pair of the current image pair to generate a second optical flow supplement increment;

[0015] S3: summing the first optical flow result corresponding to the current image pair and the second optical flow supplement increment, and using the summed result as the initial optical flow corresponding to the next image pair of the current image pair;

[0016] S4: taking the next image pair of the current image pair as the current image pair;

[0017] S5: Repeat S1-S4 until the current image is the last image pair;

[0018] Perform multi-resolution optical flow calculation on the last image pair, and obtain the final optical flow result based on the initial optical flow of the last image pair;

[0019] Perform multi-parameter component extraction on the final optical flow result corresponding to the last image pair among the multiple image pairs to obtain a light intensity component, a grayscale change component and an optical flow direction component;

[0020] Multi-parameter time-segment analysis is performed based on the light intensity component, grayscale change component, and optical flow direction component to detect HIFU micro-lesions.

[0021] The present invention also provides a HIFU micro-lesion detection system based on optical flow multi-parameter time segment analysis, the system comprising:

[0022] An image acquisition module, used for acquiring a HIFU ultrasound monitoring image sequence;

[0023] A preprocessing module, used for preprocessing the HIFU ultrasound monitoring image sequence to generate a preprocessed image sequence;

[0024] A segmentation pairing module is used to segment and pair the preprocessed image sequence according to a preset segmentation rule to generate multiple image pairs, each of which contains two images in adjacent time series;

[0025] The optical flow estimation module is used to perform the following processes:

[0026] Among the multiple image pairs, use the multi-resolution pyramid optical flow model to perform multi-resolution optical flow calculation on a first image pair among the multiple image pairs to obtain a first optical flow result corresponding to the first image pair;

[0027] Performing multi-resolution optical flow calculation on the second image in the first image pair and the first image in the next image pair of the second image pair to generate a first optical flow supplement increment;

[0028] Using the first optical flow result corresponding to the first image pair and the first optical flow supplement increment, the initial optical flow of the second image pair is obtained;

[0029] For each image pair except the first image pair and the last image pair among the multiple image pairs, the image pair is taken as the current image pair, and S1-S5 are executed:

[0030] S1: Calculate the multi-resolution optical flow of the current image pair using the multi-resolution pyramid optical flow model, and generate a first optical flow result corresponding to the current image pair according to the initial optical flow of the current image pair;

[0031] S2: performing multi-resolution optical flow calculation on the second image in the current image pair and the first image in the next image pair of the current image pair to generate a second optical flow supplement increment;

[0032] S3: summing the first optical flow result corresponding to the current image pair and the second optical flow supplement increment, and using the summed result as the initial optical flow corresponding to the next image pair of the current image pair;

[0033] S4: taking the next image pair of the current image pair as the current image pair;

[0034] S5: Repeat S1-S4 until the current image is the last image pair;

[0035] Perform multi-resolution optical flow calculation on the last image pair, and obtain the final optical flow result based on the initial optical flow of the last image pair;

[0036] A component extraction module, used for performing multi-parameter component extraction on a final optical flow result corresponding to a last image pair among the multiple image pairs to obtain a light intensity component, a grayscale change component and an optical flow direction component;

[0037] The damage detection module is used to perform multi-parameter time-segment analysis based on the light intensity component, grayscale change component and optical flow direction component, so as to detect HIFU micro-damage.

[0038] Furthermore, the multi-resolution optical flow calculation specifically includes:

[0039] For the two images in the input image pair, a pyramid is established. For each layer of the pyramid, the two images in the image pair are downsampled according to the number of layers of the current pyramid, so as to generate an image group with a resolution corresponding to the number of layers of the pyramid. In the obtained pyramid, the lowest resolution image is at the top layer and the original image is at the bottom layer.

[0040] Performing optical flow calculation on the image group obtained at each layer to generate an optical flow calculation result corresponding to the layer; multiplying the optical flow calculation result with the grayscale transformation matrix to generate an incremental optical flow corresponding to the layer;

[0041] Add the incremental optical flow corresponding to each layer in the pyramid model to obtain the total incremental optical flow;

[0042] The initial optical flow is added to the total incremental optical flow to obtain the total optical flow corresponding to the image pair.

[0043] Furthermore, the multi-resolution optical flow calculation is expressed as:

[0044] ,

[0045] in,

[0046] ,

[0047] represents the initial optical flow; Indicates The scaling index of the layer pyramid, represents the total number of layers of the pyramid, It represents the topmost layer of the pyramid; Represents the image pair for which multi-resolution optical flow calculation is to be performed conduct Multiple downsampling operations, , and The image pairs The previous and next images in ; Indicates The grayscale transformation matrix after normalization of the layer pyramid, and They are respectively The previous and next images in the layer pyramid, Calculate kernel function for optical flow; Expressing the The grayscale difference between the two images in the layer pyramid is calculated.

[0048] Furthermore, the detection of HIFU micro-damage includes:

[0049] Get the optical flow direction vector matrix corresponding to the optical flow direction component , Size ;

[0050] For each pixel , using the template matrix , by sliding the window from the optical flow direction vector matrix Extract submatrix from ,Right now ,in, , , The size is , The size is smaller than The size, is a complex matrix consisting of unit vectors;

[0051] For each sub-matrix , calculate the cosine similarity element by element, and get ,in,

[0052] ,

[0053] in, is the matrix L2 norm calculation, represents the complex conjugate operation;

[0054] By calculating the mean, we get the submatrix With the template matrix The similarity between them is recorded as pixel points Similarity ,

[0055] ;

[0056] The quantized light intensity component, grayscale change component and optical flow direction component are normalized and aggregated to obtain the damage judgment matrix ,Right now

[0057] ;

[0058] in, is the weight coefficient, Represents the normalized result of the matrix composed of the similarities of all pixels in the image. Represents the normalized image grayscale difference matrix, which represents the normalized grayscale difference between the last image and the first image in the image sequence. represents the normalized optical flow intensity matrix;

[0059] According to the damage determination matrix Detection of HIFU micro-damage.

[0060] Furthermore, the detection of HIFU micro-damage also includes:

[0061] Set the grayscale transformation filter;

[0062] Using grayscale transform filter to determine the damage matrix Perform filtering and obtain the filtering result ,in,

[0063] ,

[0064] Represents grayscale transformation filter, grayscale transformation filter Represents a matrix composed of 0s and 1s. Specifically, a convolution kernel is used to perform a convolution operation on the grayscale subtraction of the first image and the last image in the image sequence. The result after the convolution is filtered by setting a grayscale threshold. The pixels whose convolution results exceed the grayscale threshold are marked as 1, and the pixels whose convolution results exceed the grayscale threshold are marked as 0.

[0065] The filtering results Perform threshold filtering to generate damage location results.

[0066] The beneficial effects of the present invention are:

[0067] The present invention uses dynamic ultrasound data to conduct HIFU injury monitoring research. The dynamic data provides richer injury transformation information. Moreover, the present invention applies the optical flow method to HIFU injury detection and designs a HIFU micro-injury detection model with multi-parameter time segment analysis. The model makes full use of the timing information of ultrasound images and uses multiple parameters to jointly perform injury judgment and detection, further enhancing the reliability of injury detection.

[0068] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0070] Figure 1 This is the overall framework diagram of micro-damage detection using the HIFU micro-damage detection model;

[0071] Figure 2 It is a multi-parameter time-division analysis network structure diagram;

[0072] Figure 3 It is the experimental system and tissue section diagram;

[0073] Figure 4 This is the HIFU damage detection result diagram;

[0074] Figure 5 This is a comparison chart of the PR curves of various methods;

[0075] Figure 6 It is a comparison diagram of the detection results of the method of the present invention and the optical flow intensity part. DETAILED DESCRIPTION

[0076] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments are only for illustrating the present invention, rather than for limiting the protection scope of the present invention.

[0077] The present invention proposes a HIFU micro-lesion detection method based on optical flow multi-parameter time segment analysis. The method is based on the optical flow model and constructs a multi-parameter time segment HIFU micro-lesion detection algorithm. Figure 1As shown in the figure, it mainly consists of image preprocessing, motion information acquisition, multi-parameter time segment analysis, damage location and other processes. First, the collected HIFU dynamic damage data is preprocessed by cropping the region of interest, denoising, and size normalization. Then the processed data is input into the pairing script, and the dynamic ultrasound image is segmented and paired according to the preset segmentation rules (for example, acquiring images at a certain time interval), and the two adjacent images are divided into an image pair (i.e., data pair). The first pair of data pairs is subjected to initial optical flow calculation, and the original multi-parameter components are extracted from them. Then, the optical flow iterative calculation of the next pair of image pairs is performed in sequence. The calculated results are used as multi-parameter increments and added to the original multi-parameter components according to a certain weight coefficient. The confidence optimization process of the time series multi-parameter components is repeated in sequence. When the image file traversal is completed, the multiple final iterative results obtained are quantified and combined to form a damage judgment matrix, and the damage of the input HIFU image is located.

[0078] Specifically, a HIFU micro-lesion detection method based on optical flow multi-parameter time segment analysis may include:

[0079] Acquire a HIFU ultrasound monitoring image sequence, wherein the HIFU ultrasound monitoring image sequence may be extracted from the HIFU dynamic injury data in chronological order;

[0080] The HIFU ultrasound monitoring image sequence is preprocessed to generate a preprocessed image sequence; the preprocessing may include cropping of a region of interest, image denoising, size normalization, etc.

[0081] Next, the preprocessed image sequence is segmented and paired according to the preset segmentation rules to generate multiple image pairs (i.e. Figure 1 "Data Segmentation and Pairing" in the ), each image pair contains two images in adjacent time series;

[0082] Then, the optical flow estimation is performed in different time periods according to the multiple image pairs. In the prior art, usually only the first image and the last image in the image sequence are selected, while in the present invention, the HIFU damage of a period is quantified in time into multiple segments with time series correlation for optical flow calculation, and then the results are superimposed. The optical flow estimation result obtained by this method is more accurate, so that the accuracy of subsequent damage detection is also higher.

[0083] Specifically, the optical flow calculation of multiple time periods can refer to Figure 2 .

[0084] like Figure 2 As shown, among multiple image pairs, a multi-resolution pyramid optical flow model is used to perform multi-resolution optical flow calculation on a first image pair among the multiple image pairs to obtain a first optical flow result corresponding to the first image pair; it should be noted that: Figure 2 The image contents are for illustration only. The ultrasound images in the left column represent the data pairs input after image pairing. The images in the boxes of multi-parameter increments and multi-parameter variables on the right are, from left to right, schematic diagrams of light intensity changes, grayscale transformations, and motion estimation optical flow directions.

[0085] Performing multi-resolution optical flow calculation on the second image in the first image pair and the first image in the next image pair of the second image pair to generate a first optical flow supplement increment;

[0086] Using the first optical flow result corresponding to the first image pair and the first optical flow supplement increment, the initial optical flow of the second image pair is obtained;

[0087] For each image pair except the first image pair and the last image pair among the multiple image pairs, the image pair is taken as the current image pair, and S1-S5 are executed:

[0088] S1: Calculate the multi-resolution optical flow of the current image pair using the multi-resolution pyramid optical flow model, and generate a first optical flow result corresponding to the current image pair according to the initial optical flow of the current image pair;

[0089] S2: performing multi-resolution optical flow calculation on the second image in the current image pair and the first image in the next image pair of the current image pair to generate a second optical flow supplement increment;

[0090] S3: summing the first optical flow result corresponding to the current image pair and the second optical flow supplement increment, and using the summed result as the initial optical flow corresponding to the next image pair of the current image pair;

[0091] S4: taking the next image pair of the current image pair as the current image pair;

[0092] S5: Repeat S1-S4 until the current image is the last image pair;

[0093] Multi-resolution optical flow calculation is performed on the last image pair, and the final optical flow result is obtained based on the initial optical flow of the last image pair.

[0094] Then, a final optical flow result corresponding to the last image pair among the multiple image pairs is subjected to multi-parameter component extraction to obtain a light intensity component, a grayscale change component and an optical flow direction component;

[0095] Multi-parameter time-segment analysis is performed based on the light intensity component, grayscale change component, and optical flow direction component to detect HIFU micro-lesions.

[0096] The present invention also provides a HIFU micro-lesion detection system based on optical flow multi-parameter time segment analysis, the system comprising:

[0097] An image acquisition module, used for acquiring a HIFU ultrasound monitoring image sequence;

[0098] A preprocessing module, used for preprocessing the HIFU ultrasound monitoring image sequence to generate a preprocessed image sequence;

[0099] A segmentation pairing module is used to segment and pair the preprocessed image sequence according to a preset segmentation rule to generate multiple image pairs, each of which contains two images in adjacent time series;

[0100] The optical flow estimation module is used to perform the following processes:

[0101] Among the multiple image pairs, use the multi-resolution pyramid optical flow model to perform multi-resolution optical flow calculation on a first image pair among the multiple image pairs to obtain a first optical flow result corresponding to the first image pair;

[0102] Performing multi-resolution optical flow calculation on the second image in the first image pair and the first image in the next image pair of the second image pair to generate a first optical flow supplement increment;

[0103] Using the first optical flow result corresponding to the first image pair and the first optical flow supplement increment, the initial optical flow of the second image pair is obtained;

[0104] For each image pair except the first image pair and the last image pair among the multiple image pairs, the image pair is taken as the current image pair, and S1-S5 are executed:

[0105] S1: Calculate the multi-resolution optical flow of the current image pair using the multi-resolution pyramid optical flow model, and generate a first optical flow result corresponding to the current image pair according to the initial optical flow of the current image pair;

[0106] S2: performing multi-resolution optical flow calculation on the second image in the current image pair and the first image in the next image pair of the current image pair to generate a second optical flow supplement increment;

[0107] S3: summing the first optical flow result corresponding to the current image pair and the second optical flow supplement increment, and using the summed result as the initial optical flow corresponding to the next image pair of the current image pair;

[0108] S4: taking the next image pair of the current image pair as the current image pair;

[0109] S5: Repeat S1-S4 until the current image is the last image pair;

[0110] Perform multi-resolution optical flow calculation on the last image pair, and obtain the final optical flow result based on the initial optical flow of the last image pair;

[0111] A component extraction module, used for performing multi-parameter component extraction on a final optical flow result corresponding to a last image pair among the multiple image pairs to obtain a light intensity component, a grayscale change component and an optical flow direction component;

[0112] The damage detection module is used to perform multi-parameter time-segment analysis based on the light intensity component, grayscale change component and optical flow direction component, so as to detect HIFU micro-damage.

[0113] The above process is described in detail below.

[0114] As an optical flow method widely used in the field of target detection, the motion vector of each pixel in the image is calculated by comparing the changes in pixel values ​​between adjacent frame images, thereby obtaining the movement information of the object in the image, which is very suitable for deformation detection of small targets. Classic optical flow calculation methods include Lucas-Kanade (LK) optical flow, Lucas-Kanade (HS) optical flow, Brox, etc. Among them, HS optical flow estimates the optical flow by globally minimizing the energy function, which can handle a large range of motion, but may fall into a local minimum. LK optical flow is a local optical flow calculation method, which is simple and fast to calculate, but is sensitive to image noise and is not suitable for the characteristics of HIFU ultrasound imaging data with many random noises. Compared with the first two methods, the Brox method can provide higher accuracy by combining the consistency of optical flow and the characteristics of image data, and has better robustness to noise and occlusion, and can be applied to more complex scenes. Therefore, the present invention selects the Brox optical flow calculation structure as the basic model for optical flow calculation.

[0115] The Brox optical flow method constructs an energy equation based on assumptions such as constant grayscale, constant gradient, and discontinuous spatiotemporal smoothness constraints. Its function expression is as follows:

[0116] , (1)

[0117] , (2)

[0118] , (3)

[0119] Formulas (1), (2), and (3) are grayscale terms, gradient terms, and smoothing terms, respectively, where: and Respectively represent two adjacent frames of images; is the pixel coordinate in the image; Respectively represent the serial numbers of pixel coordinates in the image; is the optical flow component, which is used to represent the displacement of pixels in the image; is the gradient operator, Used to indicate the direction and rate of change of image brightness; It is the L2 norm operation; is the weight parameter used to balance the data term and the gradient term and can be set to 80.

[0120] The total energy function is constructed by combining the data term and the smoothing term. By minimizing it and establishing a numerical solution method, the motion optical flow field of the two frames of images can be obtained. The total energy function is shown in formula (4):

[0121] , (4)

[0122] in: , is a balancing parameter used to adjust the weights of the data term and the smooth term. For example, You can take 7.5.

[0123] In order to make full use of the information of ultrasound images, this paper designs a multi-resolution pyramid optical flow model based on the Brox optical flow model. A rough optical flow estimate is made at a low resolution, and then the estimate is gradually refined at a higher resolution to finally obtain a more accurate optical flow. First, a multi-resolution pyramid optical flow model is constructed. The input ultrasound image is downsampled according to the current number of pyramid layers to generate target image groups with different resolutions. Then, the optical flow calculation is performed on the scale group corresponding to each layer, and the optical flow calculation result of the layer is multiplied by the grayscale transformation matrix to generate the brightened optical flow of the initial optical flow calculation, and the total optical flow result is updated. Repeat the above steps until all pyramids are traversed.

[0124] Specifically, multi-resolution optical flow calculation may include:

[0125] For the two images in the input image pair, a pyramid is established. For each layer of the pyramid, the two images in the image pair are downsampled according to the number of layers of the current pyramid, so as to generate an image group with a resolution corresponding to the number of layers of the pyramid. In the obtained pyramid, the lowest resolution image is at the top layer and the original image is at the bottom layer.

[0126] Performing optical flow calculation on the image group obtained at each layer to generate an optical flow calculation result corresponding to the layer; multiplying the optical flow calculation result with the grayscale transformation matrix to generate an incremental optical flow corresponding to the layer;

[0127] Add the incremental optical flow corresponding to each layer in the pyramid model to obtain the total incremental optical flow;

[0128] The initial optical flow corresponding to the image pair (ie, the optical flow corresponding to the previous image pair of the image pair) is added to the total incremental optical flow to obtain the total optical flow corresponding to the image pair.

[0129] The multi-resolution optical flow calculation can be expressed by formula (5):

[0130] , (5)

[0131] in,

[0132] ,

[0133] represents the initial optical flow; Indicates The scaling index of the layer pyramid can be a number between 0 and 1. Indicates the total number of layers of the pyramid, which can be any suitable value such as 3, 4, 5, etc. Represents the topmost layer of the pyramid, which is the layer with the lowest image resolution; Represents the image pair for which multi-resolution optical flow calculation is to be performed conduct Multiple downsampling operations, , and The image pairs The previous and next images in ; Indicates The grayscale transformation matrix of the layer pyramid normalization, and They are respectively The previous and next images in the layer pyramid, Calculate kernel function for optical flow; Indicates that the grayscale difference between the two images before and after the k-th layer of the pyramid is calculated.

[0134] HIFU damage is a slowly formed process that gradually appears over time, but during this process, ultrasound images often have noise and artifacts. When using optical flow calculation to track damage changes, if only a single data pair (i.e., image pair) is relied upon, the true edge of the damage may not be distinguished due to edge noise. At the same time, the detection will also be inaccurate because noise and artifacts will cause deviations in the calculated damage position. In order to solve this problem, the present invention designs a multi-parameter, multi-period analysis network, and the data image pair performs multi-resolution optical flow calculation on the data input into the timing analysis network, and then accumulates the results of each component calculated in each group according to a certain weight. For each time-series image group (i.e., image pair), in addition to performing optical flow calculation within the current time period, multi-resolution optical flow calculation is also performed on the next image of the previous time period and the previous image of the next time period, and the result is added as a supplementary increment to the image calculation result of the next time period to achieve jump connection between adjacent time-series image pairs. Finally, the bidirectional calculation results (i.e., the first image of an image pair can calculate an optical flow with the second image in the previous image pair, and can also calculate an optical flow with the second image in the image pair) are comprehensively superimposed, and damage detection is performed based on multi-modal quantitative results such as light intensity, optical flow direction matching, image grayscale difference, etc. This process can be referred to Figure 2 .

[0135] Combination Figure 2 ,The optical flow estimation in the multi-parameter multi-period analysis network can include the following processes:

[0136] For a plurality of input image pairs, a multi-resolution pyramid optical flow model is used to perform multi-resolution optical flow calculation on a first image pair among the plurality of image pairs to obtain a first optical flow result corresponding to the first image pair;

[0137] Performing multi-resolution optical flow calculation on the second image in the first image pair and the first image in the next image pair of the second image pair to generate a first optical flow supplement increment;

[0138] Using the first optical flow result corresponding to the first image pair and the first optical flow supplement increment, the initial optical flow of the second image pair is obtained;

[0139] For each image pair except the first image pair and the last image pair among the multiple image pairs, the image pair is taken as the current image pair, and S1-S5 are executed:

[0140] S1: Calculate the multi-resolution optical flow of the current image pair using the multi-resolution pyramid optical flow model, and generate a first optical flow result corresponding to the current image pair according to the initial optical flow of the current image pair;

[0141] S2: performing multi-resolution optical flow calculation on the second image in the current image pair and the first image in the next image pair of the current image pair to generate a second optical flow supplement increment;

[0142] S3: summing the first optical flow result corresponding to the current image pair and the second optical flow supplement increment, and using the summed result as the initial optical flow corresponding to the next image pair of the current image pair;

[0143] S4: taking the next image pair of the current image pair as the current image pair;

[0144] S5: Repeat S1-S4 until the current image is the last image pair;

[0145] Multi-resolution optical flow calculation is performed on the last image pair, and the final optical flow result is obtained based on the initial optical flow of the last image pair.

[0146] After obtaining the optical flow calculation results of the ultrasound image, it is necessary to perform damage determination. When clinicians determine whether HIFU damage occurs in B-ultrasound imaging, they usually observe whether there are new bright spots in the imaging results as a basis for detection. However, due to the influence of the coherent imaging system, the ultrasound image is always covered by large granular spots. It is not reliable to judge HIFU micro-damage only by grayscale transformation. These undoubtedly increase the difficulty for doctors to judge the location of micro-damage. To this end, the present invention proposes a multi-parameter HIFU micro-damage determination method.

[0147] First, the precise optical flow obtained by multi-resolution optical flow calculation is extracted to obtain the optical flow intensity matrix And the optical flow direction vector matrix ,in, , , and are the horizontal and vertical components of the optical flow, respectively. is an imaginary unit. The intensity of optical flow reflects the motion state of the object in the image. The intensity matrix composed of this parameter can qualitatively determine the intensity of motion at each point in the ultrasound image. And through the observation of the HIFU damage from scratch in a large number of ultrasound images, it is found that the bright spot at the damage point in the image has a process of expanding from inside to outside. Therefore, the present invention designs an overall outward diffusion matrix for the matrix Perform template matching and use the result as the second basis for damage judgment. The diffusion matrix is ​​the template matrix mentioned below. The diffusion matrix is ​​a shape that diffuses outward. It is a complex matrix composed of unit vectors in 8 directions (i.e., up, down, left, right, upper left, lower left, upper right, and lower right). By comparing the similarity between the template matrix and the calculated optical flow direction matrix, the damage positions with the same motion trend can be captured.

[0148] Specifically, the detection of HIFU micro-damage includes:

[0149] Get the optical flow direction vector matrix corresponding to the optical flow direction component , Size ;

[0150] For each pixel , using the template matrix , by sliding the window from the optical flow direction vector matrix Extract submatrix from , as shown in formula (6):

[0151] , (6)

[0152] in, , , The size is , The size is smaller than The size, is a complex matrix composed of unit vectors. The submatrix with the currently calculated pixel as the vertex can be extracted through formula (6), which is convenient for the subsequent local similarity calculation.

[0153] Then, for each submatrix , calculate the cosine similarity element by element, and get ,in,

[0154] , (7)

[0155] in, is the matrix L2 norm calculation, represents the complex conjugate operation;

[0156] After calculating the cosine similarity of each pixel, similarity aggregation is performed and the submatrix is ​​obtained by performing mean calculation. With the template matrix The similarity between them is recorded as pixel points Similarity :

[0157] , (8)

[0158] Then, the quantized components (i.e., light intensity component, optical flow direction matching component, and image grayscale difference component) are normalized and aggregated to obtain the damage judgment matrix ,Right now

[0159] , (9)

[0160] in, is the weight coefficient, Represents the matrix composed of the similarities of all pixels in the image, that is, the optical flow direction matching matrix. Represents the image grayscale difference matrix, which represents the normalized grayscale difference (i.e., grayscale subtraction) between the last image and the first image in the image sequence. For example, after image pairing, the image sequence I1-I8 generates four time-related image pairs (I1, I2), (I3, I4), (I5, I6), and (I7, I8). Then, here Represents the normalized result of the grayscale difference between I8 and I1, represents the normalized optical flow intensity matrix;

[0161] According to the damage determination matrix HIFU micro-damage can be detected.

[0162] In order to further ensure the accuracy of the damage detection results, the present invention also adds a grayscale transformation filter. Specifically, the detection of HIFU micro-damage also includes:

[0163] Set the grayscale transformation filter;

[0164] Using grayscale transform filter to determine the damage matrix Perform filtering and obtain the filtering result ,in,

[0165] , (10)

[0166] represents the grayscale transformation filter, is the weight coefficient, which can be 0.5, 0.2 and 0.3 respectively as an example;

[0167] Grayscale transformation filter Represents a matrix composed of 0s and 1s. Specifically, a convolution kernel (for example, a 3*3 matrix of all 1s) is used to perform a convolution operation on the grayscale subtraction of the first image and the last image in the image sequence (the SAME mode convolution is used so that the size of the image before and after the convolution does not change). The result after the convolution is filtered by setting a grayscale threshold, and the pixels whose convolution results exceed the grayscale threshold are marked as 1, and the pixels whose convolution results exceed the grayscale threshold are marked as 0.

[0168] Through convolution, the correlation of elements near the variable can be reflected to a certain extent. The threshold filter is set by counting the total calculated value of the surrounding elements rather than the value of each element itself. In this way, some weak damage near the damaged area will not be filtered out.

[0169] Finally, the filtering results Perform threshold filtering to generate damage location results. Compare with the final judgment threshold. If If the value of a certain pixel in the image is greater than the judgment threshold, it means that there is damage at the position corresponding to the pixel.

[0170] In order to measure the performance of the multi-parameter HIFU damage determination method proposed in this paper, this paper introduces precision, recall and F1-score as evaluation indicators.

[0171] The precision is defined as the ratio of the area correctly identified as damaged to all areas identified as damaged in the damage localization results. This indicator reflects the accuracy of the method in identifying damaged areas. As shown in formula (11), TP is the correctly detected damage point, and FP is the falsely detected point.

[0172] , (11)

[0173] The recall rate refers to the ratio of the number of correctly detected damaged areas to the number of actual damaged areas in the damage localization results. This indicator measures the performance of the method in detecting all damaged areas. As shown in formula (12), FN is the undetected damaged area.

[0174] , (12)

[0175] In order to prevent the contradiction caused by too many missed detections, resulting in too high precision P and low recall, the present invention also introduces the F1 score, as shown in formula (13). This evaluation index is the harmonic mean of precision and recall, and is used to comprehensively evaluate the performance of the method. The higher the F1 score, the better the balance between precision and recall.

[0176] , (13)

[0177] Next, the advantages of the present invention are explained through specific experiments.

[0178] The HIFU device used to produce micro-injury in this experiment is a small HIFU concentrator. The experimental system mainly consists of a HIFU amplifier, an ultrasonic imager, and a computer control system. Figure 3 As shown in Figure A. The HIFU probe is a Siansonic TQ20-2020 transducer with a center frequency of 2.0Mhz, a geometric focal length of 2cm, and a maximum output sound power of 25w. The distance between the focus center of the ex vivo biological tissue and the HIFU transducer is maintained at 2cm by a fixing device to ensure that the focusing energy received by the central area of ​​the tissue is the strongest. At the same time, the high-frequency superficial ultrasound probe is aligned with the biological tissue at the same horizontal height for imaging, and the data collected by the probe is transmitted to the ultrasound imaging device for video acquisition, with a frame rate of 30fps. The input voltage of the power amplifier is set to 280mv, and the corresponding transducer output power is 10w.

[0179] According to the above set irradiation parameters, the fresh ex vivo pork tenderloin was irradiated for 20 seconds, and a 1-minute video was recorded using an ultrasound imaging machine for doctors to determine the location of micro-injury points. After irradiation, the pork slices were sliced ​​to find the injury points, and photos were taken for evaluation. Figure 3 As shown in Figure B. Figure 3 In Figure B, it can be observed that an elliptical micro-damage of about 3 mm in size was produced in the tissue (i.e., the white elliptical area in the figure).

[0180] The process of making segmented data is as follows: crop the collected data video, find the time segment of HIFU irradiation, and capture the entire irradiation process (20s). The frame interval is set to 3, that is, extract once every 3 frames, which is equivalent to extracting 10 images in a 1-second video, that is, discretizing the 20-second irradiation process into 200 images with a time series relationship. For the discretized images, extract one image every 5 images to form a new data set of 40 images. The obtained 40 image data are evenly divided into 4 groups, that is, the irradiation process is divided into 4 stages, each stage is 5s. The time interval between two adjacent images in the same stage is 0.5s. The real label data are all annotated by professional clinical HIFU doctors and ultrasound imaging doctors based on irradiation videos and sliced ​​tissues. Through the method proposed in the present invention, damage is located in each of the above four stages, so as to achieve continuous monitoring of HIFU micro-damage.

[0181] This experiment selected 30 pieces of fresh ex vivo tenderloin tissue as experimental samples for a total of 120 groups, and selected 40 images from the video data collected in each group to construct a data set for analyzing the dynamic changes of the HIFU damage area of ​​interest. In this data set, every 10 images represent a stage, and each piece of ex vivo tissue has experienced 4 HIFU irradiation damage stages, totaling 120 types. The processed data was input into the model through Matlab 2022a software programming, and the obtained multiple damage judgment parameters were combined and output to obtain the corresponding damage localization map. Some experimental results are shown as follows. Figure 4 shown. Figure 4 The three images in the first column are, from top to bottom, the unirradiated image (i.e., the image marked with A), the irradiated ultrasound image (i.e., the image marked with B), and the irradiated tissue section image (i.e., the image marked with C). Figure 4 The second, third, fourth and fifth columns in the figure correspond to the four stages of HIFU irradiation of this piece of pork tissue. Figure 4 A in the figure) and the image after irradiation (i.e. Figure 4 Figure B in the figure shows that there are three damaged areas, a, b, and c. In the first stage, that is, comparing Figure 4 Figure A-1 and Figure 4 In Figure B-1, we can see that the bright spot in the damaged target area c has a tendency to spread outward, and the grayscale of the target area b changes from dark to bright, indicating that damage is occurring in areas b and c. Figure 4 In Figure C-1, the present invention correctly detects two damaged areas b and c, where the green rectangle is the damage location point. Comparing the third, fourth and fifth columns (i.e., the second stage, the third stage and the fourth stage) in turn, it can be seen that the detection algorithm of the present invention can accurately identify the HIFU damage generated in each stage, and there is no false detection point on this sample.

[0182] In order to verify the effectiveness of the algorithm in this paper, this paper conducted comparative experiments on grayscale, light intensity and other detection methods. The results are shown in Table 1.

[0183] Table 1 Damage identification results of different methods

[0184]

[0185] Among them, in Table 1, the "grayscale" method refers to the method of using grayscale alone for damage detection; the "optical flow intensity" method refers to the method of using optical flow intensity alone for damage detection; and the "grayscale + optical flow intensity" method refers to the method of combining both grayscale and optical flow intensity parameters for damage detection.

[0186] In Table 1, the performance of various methods is evaluated based on the precision, recall and F1 score. After comprehensive comparison, the method of the present invention performs best in both precision and F1 score, reaching 89.8% and 85.1% respectively, among which the precision is improved by about 20% compared with various methods. From the perspective of F1 score, the temporal multi-parameter HIFU damage detection model proposed in the present invention achieves a good recall rate while maintaining a high precision rate.

[0187] In order to further evaluate the ability of this algorithm in identifying positive samples, this paper draws the precision-recall rate curve (PR curve) of the model, randomly selects 20 groups of data, and increases the threshold by 0.005 (from 0.005 to 1.000) in steps of 0.005, and records the precision and recall rate of the algorithm at each threshold (that is, the threshold in the threshold filtering after calculating the result). The obtained data pairs form a PR curve, as shown in Figure 5 As shown. By calculating the area under the PR curve (AUC), the comprehensive performance of the model at different thresholds can be fully quantified. Figure 5 The PR curve shown in the figure shows that the AUC value of the method of the present invention is 0.843, and the recall rate is at the optimal accuracy in the interval [0.6, 0.95]. Compared with other PR curves, the curve of the method of the present invention is smoother, indicating that the model has excellent performance in the recognition ability of HIFU micro-lesions, and can maintain good accuracy while ensuring a high recall rate. This result further verifies the effectiveness and stability of the algorithm when processing unbalanced data sets, especially in the recognition task of positive samples. In summary, the method of the present invention has high accuracy and reliability in detecting HIFU micro-lesions, and can be used as an effective method for monitoring damage during HIFU treatment.

[0188] In HIFU treatment, accurate monitoring of the damaged area is crucial to the treatment effect. The time-series multi-parameter HIFU damage detection model proposed in the present invention can effectively identify the damaged area generated during HIFU irradiation by combining the time-series information and multi-parameter characteristics of the ultrasound image. Experimental results show that compared with the traditional grayscale judgment method, the multi-parameter damage recognition method proposed in this paper has an accuracy improvement of 20.3% and a recall rate slightly higher than the traditional grayscale judgment method. Compared with the optical flow intensity method, this model has a significant improvement in accuracy, which shows that the model can maintain a high damage area recognition rate while reducing false detections.

[0189] Although the method in the present invention does not achieve the best recall rate, the difference between the methods is only about 2%. This is because other methods identify a larger number of areas to be verified, and the number of label areas contained therein will also increase, but the majority of these areas contain erroneous points.

[0190] Figure 6 is a comparison chart of the detection results of the method of the present invention and the light part of the optical flow, where Figure 6 Figures A and B in the middle are ultrasound images before and after HIFU irradiation at the current stage; Figure 6 Figure C in the middle is the HIFU damage detection result of the method of the present invention; Figure 6 Figure D in the middle is the damage detection result of the optical flow intensity method with the highest recall rate. Figure 6 The red box is the damage error area, the green box is the correct damage detection area, the white ellipse is the non-damage area, and the blue ellipse is the multi-detection damage area. Figure 6 In Figure D, although one more damaged area is detected, there are also areas that are misdetected. Therefore, the improvement of recall rate in the single parameter method is at the expense of precision rate. Figure 6 As can be seen from Figure C, the result graph detected using the method of the present invention avoids the occurrence of this problem and achieves a better balance between recall rate and precision rate.

[0191] Therefore, the method proposed in this invention is not only innovative in theory, but also has high practical value in practical application, and is expected to become an important tool for damage monitoring in HIFU treatment. Future research can further optimize the model parameters and improve the recall rate to achieve a more comprehensive damage monitoring effect.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A HIFU micro-lesion detection method based on optical flow multi-parameter time segment analysis, characterized in that: include: Acquire HIFU ultrasound monitoring image sequences; Preprocessing the HIFU ultrasound monitoring image sequence to generate a preprocessed image sequence; The preprocessed image sequence is segmented and paired according to a preset segmentation rule to generate multiple image pairs, each of which contains two images in adjacent time series; Among the multiple image pairs, use the multi-resolution pyramid optical flow model to perform multi-resolution optical flow calculation on a first image pair among the multiple image pairs to obtain a first optical flow result corresponding to the first image pair; Performing multi-resolution optical flow calculation on the second image in the first image pair and the first image in the next image pair of the second image pair to generate a first optical flow supplement increment; Using the first optical flow result corresponding to the first image pair and the first optical flow supplement increment, the initial optical flow of the second image pair is obtained; For each image pair except the first image pair and the last image pair among the multiple image pairs, the image pair is taken as the current image pair, and S1-S5 are executed: S1: Calculate the multi-resolution optical flow of the current image pair using the multi-resolution pyramid optical flow model, and generate a first optical flow result corresponding to the current image pair according to the initial optical flow of the current image pair; S2: performing multi-resolution optical flow calculation on the second image in the current image pair and the first image in the next image pair of the current image pair to generate a second optical flow supplement increment; S3: summing the first optical flow result corresponding to the current image pair and the second optical flow supplement increment, and using the summed result as the initial optical flow corresponding to the next image pair of the current image pair; S4: taking the next image pair of the current image pair as the current image pair; S5: Repeat S1-S4 until the current image is the last image pair; Perform multi-resolution optical flow calculation on the last image pair, and obtain the final optical flow result based on the initial optical flow of the last image pair; Perform multi-parameter component extraction on the final optical flow result corresponding to the last image pair among the multiple image pairs to obtain a light intensity component, a grayscale change component and an optical flow direction component; Multi-parameter time-segment analysis is performed based on the light intensity component, grayscale change component, and optical flow direction component to detect HIFU micro-lesions. Multi-resolution optical flow calculation specifically includes: For the two images in the input image pair, a pyramid is established. For each layer of the pyramid, the two images in the image pair are downsampled according to the number of layers of the current pyramid, so as to generate an image group with a resolution corresponding to the number of layers of the pyramid. In the obtained pyramid, the lowest resolution image is at the top layer and the original image is at the bottom layer. Performing optical flow calculation on the image group obtained at each layer to generate an optical flow calculation result corresponding to the layer; multiplying the optical flow calculation result with the grayscale transformation matrix to generate an incremental optical flow corresponding to the layer; Add the incremental optical flow corresponding to each layer in the pyramid model to obtain the total incremental optical flow; Adding the initial optical flow corresponding to the image pair to the total incremental optical flow to obtain the total optical flow corresponding to the image pair; Detection of HIFU micro-damage includes: Get the optical flow direction vector matrix corresponding to the optical flow direction component , Size ; For each pixel , using the template matrix , by sliding the window from the optical flow direction vector matrix Extract submatrix from ,Right now ,in, , , The size is , The size is smaller than The size, is a complex matrix consisting of unit vectors; For each submatrix , calculate the cosine similarity element by element, and get ,in, ， in, is the matrix L2 norm calculation, represents the complex conjugate operation; By calculating the mean, we get the submatrix With the template matrix The similarity between them is recorded as pixel points Similarity , ； The quantized light intensity component, grayscale change component and optical flow direction component are normalized and aggregated to obtain the damage judgment matrix ,Right now ； in, is the weight coefficient, Represents the normalized result of the matrix composed of the similarities of all pixels in the image. Represents the normalized image grayscale difference matrix, which represents the normalized grayscale difference between the last image and the first image in the image sequence. represents the normalized optical flow intensity matrix; According to the damage determination matrix Detection of HIFU micro-damage.

2. The HIFU micro-lesion detection method based on optical flow multi-parameter time segment analysis according to claim 1, characterized in that: The multi-resolution optical flow calculation formula is expressed as: ， in, ， represents the initial optical flow; Indicates The scaling index of the layer pyramid, represents the total number of layers of the pyramid, It represents the topmost layer of the pyramid; Represents the image pair for which multi-resolution optical flow calculation is to be performed conduct Multiple downsampling operations, , and The image pairs The previous and next images in ; Indicates The grayscale transformation matrix after normalization of the layer pyramid, and They are respectively The previous and next images in the layer pyramid, Calculate kernel function for optical flow; Indicates The grayscale difference between the two images in the layer pyramid is calculated.

3. The HIFU micro-lesion detection method based on optical flow multi-parameter time segment analysis according to claim 1, characterized in that: Detection of HIFU micro-damage also includes: Set grayscale conversion filter ; Using grayscale transformation filter Damage judgment matrix Perform filtering and obtain the filtering result ,in, ， Among them, the grayscale transformation filter Represents a matrix composed of 0s and 1s. Specifically, a convolution kernel is used to perform a convolution operation on the grayscale subtraction of the first image and the last image in the image sequence. The result after the convolution is filtered by setting a grayscale threshold. The pixels whose convolution results exceed the grayscale threshold are marked as 1, and the pixels whose convolution results exceed the grayscale threshold are marked as 0. The filtering results Perform threshold filtering to generate damage location results.

4. A HIFU micro-lesion detection system based on optical flow multi-parameter time segment analysis, characterized in that: include: An image acquisition module, used for acquiring a HIFU ultrasound monitoring image sequence; A preprocessing module, used for preprocessing the HIFU ultrasound monitoring image sequence to generate a preprocessed image sequence; A segmentation pairing module is used to segment and pair the preprocessed image sequence according to a preset segmentation rule to generate multiple image pairs, each of which contains two images in adjacent time series; The optical flow estimation module is used to perform the following processes: Among the multiple image pairs, use the multi-resolution pyramid optical flow model to perform multi-resolution optical flow calculation on a first image pair among the multiple image pairs to obtain a first optical flow result corresponding to the first image pair; Performing multi-resolution optical flow calculation on the second image in the first image pair and the first image in the next image pair of the second image pair to generate a first optical flow supplement increment; Using the first optical flow result corresponding to the first image pair and the first optical flow supplement increment, the initial optical flow of the second image pair is obtained; For each image pair except the first image pair and the last image pair among the multiple image pairs, the image pair is taken as the current image pair, and S1-S5 are executed: S1: Calculate the multi-resolution optical flow of the current image pair using the multi-resolution pyramid optical flow model, and generate a first optical flow result corresponding to the current image pair according to the initial optical flow of the current image pair; S2: performing multi-resolution optical flow calculation on the second image in the current image pair and the first image in the next image pair of the current image pair to generate a second optical flow supplement increment; S3: summing the first optical flow result corresponding to the current image pair and the second optical flow supplement increment, and using the summed result as the initial optical flow corresponding to the next image pair of the current image pair; S4: taking the next image pair of the current image pair as the current image pair; S5: Repeat S1-S4 until the current image is the last image pair; Perform multi-resolution optical flow calculation on the last image pair, and obtain the final optical flow result based on the initial optical flow of the last image pair; A component extraction module, used for performing multi-parameter component extraction on a final optical flow result corresponding to a last image pair among the multiple image pairs to obtain a light intensity component, a grayscale change component and an optical flow direction component; The damage detection module is used to perform multi-parameter time-segment analysis based on the light intensity component, grayscale change component and light flow direction component, so as to detect HIFU micro-damage; Multi-resolution optical flow calculation specifically includes: For the two images in the input image pair, a pyramid is established. For each layer of the pyramid, the two images in the image pair are downsampled according to the number of layers of the current pyramid, so as to generate an image group with a resolution corresponding to the number of layers of the pyramid. In the obtained pyramid, the lowest resolution image is at the top layer and the original image is at the bottom layer. Performing optical flow calculation on the image group obtained at each layer to generate an optical flow calculation result corresponding to the layer; multiplying the optical flow calculation result with the grayscale transformation matrix to generate an incremental optical flow corresponding to the layer; Add the incremental optical flow corresponding to each layer in the pyramid model to obtain the total incremental optical flow; Adding the initial optical flow corresponding to the image pair to the total incremental optical flow to obtain the total optical flow corresponding to the image pair; Detection of HIFU micro-damage includes: Get the optical flow direction vector matrix corresponding to the optical flow direction component , Size ; For each pixel , using the template matrix , by sliding the window from the optical flow direction vector matrix Extract submatrix from ,Right now ,in, , , The size is , The size is smaller than The size, is a complex matrix consisting of unit vectors; For each submatrix , calculate the cosine similarity element by element, and get ,in, ， in, is the matrix L2 norm calculation, represents the complex conjugate operation; By calculating the mean, we get the submatrix With the template matrix The similarity between them is recorded as pixel points Similarity , ； The quantized light intensity component, grayscale change component and optical flow direction component are normalized and aggregated to obtain the damage judgment matrix ,Right now ； in, is the weight coefficient, Represents the normalized result of the matrix composed of the similarities of all pixels in the image. Represents the normalized image grayscale difference matrix, which represents the normalized grayscale difference between the last image and the first image in the image sequence. represents the normalized optical flow intensity matrix; According to the damage determination matrix Detection of HIFU micro-damage.

5. The HIFU micro-lesion detection system based on optical flow multi-parameter time segment analysis according to claim 4, characterized in that: The multi-resolution optical flow calculation formula is expressed as: ， in, ， represents the initial optical flow; Indicates The scaling index of the layer pyramid, represents the total number of layers of the pyramid, It represents the topmost layer of the pyramid; Indicates the image pair for which multi-resolution optical flow calculation is to be performed conduct Multiple downsampling operations, , and The image pairs The previous and next images in ; Represents the normalized The grayscale transformation matrix of the layer pyramid, and They are respectively The previous and next images in the layer pyramid, Calculate kernel function for optical flow; Expressing the The grayscale difference between the two images in the layer pyramid is calculated.

6. The HIFU micro-lesion detection system based on optical flow multi-parameter time segment analysis according to claim 4, characterized in that: Detection of HIFU micro-damage also includes: Set grayscale conversion filter ; Using grayscale transform filter Damage judgment matrix Perform filtering and obtain the filtering result ,in, ， Among them, the grayscale transformation filter Represents a matrix composed of 0s and 1s. Specifically, a convolution kernel is used to perform a convolution operation on the grayscale subtraction of the first image and the last image in the image sequence. The result after the convolution is filtered by setting a grayscale threshold. The pixels whose convolution results exceed the grayscale threshold are marked as 1, and the pixels whose convolution results exceed the grayscale threshold are marked as 0. The filtering results Perform threshold filtering to generate damage location results.

Images