A method to improve the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN

Abstract

The present invention provides a method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN. The method is characterized by comprising behavioral feature analysis of a speckle motion image, designing a blood flow velocity model based on spatiotemporal feature analysis, and testing the predictive ability of the model based on experimental data from a rotating scattering plate. The method uses a 3D-CNN model with excellent performance to extract the spatiotemporal features of three-dimensional speckle image pixels representing different blood flow velocities using a 3D convolution kernel. By fully exploiting the spatiotemporal pixel information of the speckle motion image under single exposure, the pixel features of the speckle motion image under different flow velocities are extracted. A characteristic function is established between the speckle motion image obtained using a single exposure technique and the flow velocity measured using a multiple exposure method. This method accurately predicts the flow velocity represented by the speckle image, thereby improving the measurement accuracy of LSCI and enabling linear measurement. The method becomes a low-cost, high-resolution, real-time linear imaging technology with better clinical application.

Description

Technical Field

[0001] The present invention relates to the field of laser speckle contrast blood flow velocity detection technology optimization, and in particular to a method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN. Background Art

[0002] Measuring surface microcirculatory blood flow velocity is a highly sought-after indicator in clinical practice and is widely used in areas such as cardiovascular health assessment, skin cancer detection, skin burn severity assessment, and the prevention of diabetic plantar ulcers. Laser Speckle Contrast Imaging (LSCI) is a technology that can non-invasively image blood flow within tissues. It is widely used in clinical blood flow velocity detection and has the advantages of being non-invasive, real-time, simple, and low-cost. However, this technology is based on the basic principle of dynamic light scattering and uses various approximate models to approximate the scattered light autocorrelation function to obtain blood flow velocity. This process is affected by numerous uncertainties such as static scatterers, speckle size, and exposure time, which severely reduces the accuracy of the model, making it impossible to perform linear imaging. This leads to significant errors in the measurement of actual blood flow velocity, severely impacting its clinical application. Therefore, improving the accuracy of blood flow velocity measurement is a key issue that this technology urgently needs to address.

[0003] Currently, some progress has been made in improving LSCI imaging accuracy, primarily in two categories: model-improvement methods based on eliminating numerous uncertainties, and multi-exposure LSCI (MELSCI) methods based on fitting algorithms. While the former optimizes the single-exposure speckle model based on factors such as static scatterers and speckle size, improving measurement accuracy to a certain extent, the model is still affected by other factors that cannot be eliminated using mathematical methods, and thus cannot achieve linear measurement of blood flow velocity. MELSCI no longer relies on model calculations, but instead obtains blood flow velocity through fitting using contrast at different exposure times. This effectively improves LSCI imaging accuracy and makes it a technique capable of linear imaging. However, this method significantly increases system cost, and the fitting process also increases imaging time by two to three times, reducing LSCI's real-time imaging capabilities. Consequently, mainstream LSCI instruments on the market are still developed based on the single-exposure model.

[0004] In recent years, prediction methods based on artificial intelligence (AI) have been widely used, providing a new means to improve the measurement accuracy of LSCI. However, current AI applications in this field are limited and face challenges such as lack of data continuity and increased system costs. Therefore, a new AI-based method has been proposed to effectively improve the measurement accuracy of LSCI without increasing system costs or compromising real-time performance. This method is of great significance for establishing LSCI as a low-cost, high-precision, real-time linear imaging technology. Furthermore, this method can be applied to in vivo experiments for testing and optimization, promoting the development of this technology. Summary of the Invention

[0005] The purpose of the present invention is to propose a method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solutions:

[0007] A 3D-CNN-based method for improving the accuracy of laser speckle contrast blood flow imaging is presented. This method focuses on single-exposure LSCI, which features low cost, a simple system, and fast imaging speed. Since MELSCI has been demonstrated to be a linear blood flow measurement technique, the method uses a 3D-CNN, which excels in spatiotemporal feature extraction, to establish a mapping model between the LSCI technique's single-exposure speckle motion image and the flow velocity measured by MELSCI. The model's accuracy is then improved and tested using experimental data from a rotating scattering plate.

[0008] The specific principle is as follows: First, the speckle animations of the rotating scattering plate representing different speeds are used as the input data of the 3D-CNN. Since the speckle animations are multi-frame motions in terms of time series, and each frame of the speckle image contains a large amount of speckle pixel value information in terms of spatial sequence, the accuracy of velocity prediction depends on obtaining rich time series information and spatial pixel value information of the single-frame speckle image. Therefore, the three-dimensional convolution kernel in the 3D-CNN is used to extract the gradient of the three-dimensional pixel matrix. Then, a network composed of convolution, pooling and full connection is used to extract features and change the dimension. The blood flow velocity prediction value represented by each speckle animation is obtained respectively. Finally, the blood flow velocity prediction value of the speckle animation is output to complete the prediction. This method can accurately measure the velocity corresponding to the speckle animation, effectively improving the measurement accuracy of LSCI.

[0009] Specifically, it includes five parts: rotating scattering plate experimental device, data preprocessing, speckle image behavior feature analysis, model parameter setting and training, and model accuracy detection. The rotating scattering plate experimental device consists of a servo motor, a uniform diffuse reflection plate, and a frosted glass reflector. During the experiment, the motor drives the diffuse reflection plate to rotate behind the frosted glass at a certain angular velocity, simulating the flow of red blood cells under the epidermis in human tissue. When the motor speed (angular velocity) is known, the linear velocity on the diffuse reflection plate is the product of the angular velocity and the radius of the point. Therefore, in the speckle image obtained by the experimental device, the linear velocity of each point is fixed and can be calculated as a certain amount of data. Using it as the training target of the model can achieve quantitative evaluation of the model. The experimental device has a total of 10 angular velocity settings.

[0010] The data processing specifically includes the following:

[0011] Data augmentation is achieved by sliding a 32x32 square window from the 32nd pixel to the 320th pixel in steps of 32 to change the radius. Each pixel represents a radius of 0.037 mm, resulting in nine radii. Seventeen sets of data are collected for each radius, resulting in 9x17x10 = 1530 data points for training. This data is partitioned into a training set: validation set: test set ratio of 7:1:2.

[0012] The speckle image behavior feature analysis specifically includes the following contents:

[0013] The pixel values ​​of a single-frame speckle image contain rich spatial information, and the temporal variation of pixel values ​​at a fixed location is closely related to velocity. This means that predicting the motion velocity of a speckle animation depends not only on the spatial pixel values ​​of a single frame but also on the variation in pixel values ​​at fixed locations between frames. The greater the variation in pixel values, the faster the corresponding speckle pattern changes, indicating a higher blood flow velocity. Therefore, velocity prediction is strongly correlated with the spatial and temporal characteristics of speckle animations. Comprehensive and in-depth learning of these features can better exploit the spatiotemporal dynamics of speckle animations.

[0014] The model parameter setting and training specifically include the following:

[0015] The 3D-CNN network in the present invention has five layers. The first layer is the convolution layer C1, which directly accepts the input speckle data and uses a 3*3*5 convolution kernel; the second layer is the maximum pooling layer P2, with a pooling kernel of 2*2*2; the third layer is the convolution layer C3, which uses a 3*3*5 convolution kernel; the fourth layer is the maximum pooling layer P4, with a pooling kernel of 2*2*2; the fifth layer is the convolution layer C5, which uses a 3*3*3 convolution kernel; the sixth layer is the fully connected layer F6, which reconstructs the three-dimensional feature map output by C5 into a single velocity data, with an input dimension of 128 and an output dimension of 1; the convolution layer activation function of this model is the ReLU function, the optimizer uses Adam to accelerate the convergence speed, the batch size is 32, the initial learning rate is 0.001, and the loss function is the MSE function; MSE and MAPE are used as evaluation indicators.

[0016] The model accuracy test specifically includes the following contents:

[0017] The speckle image when the sliding window is moved to positions 97-128 was used as prediction data. This position is roughly in the middle of the entire speckle animation, corresponding to a radius of approximately (97+16)*0.037=4.2mm. The linear velocity varies moderately with the angular velocity, and the overall coverage range is relatively wide, ranging from 4.2*0.047 to 4.2*0.952=0.2mm / s to 4.0mm / s. There are 10 speeds in total, which can effectively test the training level and accuracy of the 3D-CNN model.

[0018] The LSCI images obtained by the rotating scattering plate phantom experimental apparatus are all 320*32*1024 speckle animations. Therefore, the speckle pixel values ​​of each speckle animation can be regarded as a three-dimensional matrix. The angular velocity of the rotating scattering plate phantom experimental apparatus is controlled by changing the rotation speed of the servo motor. A total of 10 angular velocities are set in the present invention, as shown in Table 1.

[0019] Table 1 Angular velocity

[0020]

[0021] The formulas for the two evaluation methods of the 3D-CNN are:

[0022]

[0023]

[0024] In Equation 1, m is the number of samples, y i is the actual value, is the predicted value; n is the number of samples in equation 2, y i is the actual value, is the predicted value.

[0025] The speckle dynamic images used for model detection correspond to 10 speeds as shown in Table 2. These ten speeds are obtained by fixing the radius to 4.2 mm.

[0026] Table 2 LSCI image true speed

[0027]

[0028] The size of the input speckle animation for training was set to 32*32*15. This time series length was obtained through extensive manual experimentation and experience, balancing the difficulty of model training with the completeness of the predicted information. Furthermore, the 32*32 square space size is more suitable for feature extraction of the convolution kernel.

[0029] The specific configuration of the 3D-CNN is shown in Table 3.

[0030] Table 2 3D-CNN parameter configuration

[0031]

[0032] In the LSCI speckle animation, the pixel values ​​of a single frame contain rich spatial information, and the temporal variation of pixel values ​​at a fixed location is closely related to velocity. This means that predicting the motion speed of a speckle animation depends not only on the spatial pixel values ​​of a single frame but also on the variation in pixel values ​​at fixed locations between frames. Therefore, velocity prediction is strongly correlated with the spatial and temporal characteristics of the speckle animation.

[0033] The pixel value matrix of each speckle animation is a complete three-dimensional matrix. Therefore, the 3D-CNN used in this paper can mine the spatiotemporal features of the above matrix sequence while ensuring the integrity of spatiotemporal information, perform overall feature extraction, and organically complete the prediction.

[0034] In 3D-CNN, the area covered by the convolution kernel is called the receptive field. This is also the size of the area on the input image where pixels in the feature map output by each layer of the network are mapped. Therefore, there is a strong spatiotemporal correlation between the pixel values ​​in the same receptive field. Starting from the input layer, 3D-CNN uses a three-dimensional convolution kernel to extract spatiotemporal information from the pixel matrix, summarizes the features, and forms a new feature map, which is then passed to the next layer for further processing.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention uses a 3D-CNN model with excellent performance in the field of spatiotemporal feature extraction. It extracts the spatiotemporal features of pixels in three-dimensional speckle images (speckle animations) representing different blood flow velocities through a 3D convolution kernel. By fully exploiting the spatiotemporal pixel information of the speckle animations under single exposure, the pixel features of the speckle animations under different flow velocities are extracted. A characteristic function is established between the speckle animations obtained using single-exposure technology and the flow velocities measured using multiple exposure methods. This allows for accurate prediction of the flow velocity represented by the speckle image, thereby improving the measurement accuracy of LSCI and enabling linear measurement. This technology becomes a low-cost, high-resolution, real-time linear imaging technology with better clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the blood velocity prediction model;

[0038] Figure 2 Schematic diagram of the rotating scattering plate phantom;

[0039] Figure 3 LSCI images acquired for the rotating scatter plate phantom;

[0040] Figure 4 Select the method for the radius of the input data;

[0041] Figure 5 is the pixel position distribution map of a single-frame speckle image;

[0042] Figure 6 is the spatiotemporal pixel distribution map of the speckle image;

[0043] Figure 7 For the 3D-CNN prediction process;

[0044] Figure 8 Method for selecting speckle images for test data. DETAILED DESCRIPTION

[0045] In order to clarify the technical problems, technical solutions, implementation processes and performance demonstrations, the present invention is further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are for explanation only. The present invention is not intended to limit the present invention. Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings do not have to be drawn to scale unless otherwise specified.

[0046] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0047] In addition, numerous specific details are provided in the following detailed description to better illustrate the present disclosure. Those skilled in the art will appreciate that the present disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art are not described in detail in order to highlight the main points of the present disclosure.

[0048] Example 1

[0049] A 3D-CNN-based method for improving the accuracy of laser speckle contrast blood flow imaging is presented. This method focuses on single-exposure LSCI, which features low cost, a simple system, and fast imaging speed. Since MELSCI has been demonstrated to be a linear blood flow measurement technique, the method uses a 3D-CNN, which excels in spatiotemporal feature extraction, to establish a mapping model between the LSCI technique's single-exposure speckle motion image and the flow velocity measured by MELSCI. The model's accuracy is then improved and tested using experimental data from a rotating scattering plate.

[0050] The specific principle is as follows: First, the speckle animations of the rotating scattering plate representing different speeds are used as the input data of the 3D-CNN. Since the speckle animations are multi-frame motions in terms of time series, and each frame of the speckle image contains a large amount of speckle pixel value information in terms of spatial sequence, the accuracy of velocity prediction depends on obtaining rich time series information and spatial pixel value information of the single-frame speckle image. Therefore, the three-dimensional convolution kernel in the 3D-CNN is used to extract the gradient of the three-dimensional pixel matrix. Then, a network composed of convolution, pooling and full connection is used to extract features and change the dimension. The blood flow velocity prediction value represented by each speckle animation is obtained respectively. Finally, the blood flow velocity prediction value of the speckle animation is output to complete the prediction. This method can accurately measure the velocity corresponding to the speckle animation, effectively improving the measurement accuracy of LSCI.

[0051] Specifically, it includes five parts: rotating scattering plate experimental device, data preprocessing, speckle image behavior feature analysis, model parameter setting and training, and model accuracy detection. The rotating scattering plate experimental device consists of a servo motor, a uniform diffuse reflection plate, and a frosted glass reflector. During the experiment, the motor drives the diffuse reflection plate to rotate behind the frosted glass at a certain angular velocity, simulating the flow of red blood cells under the epidermis in human tissue. When the motor speed (angular velocity) is known, the linear velocity on the diffuse reflection plate is the product of the angular velocity and the radius of the point. Therefore, in the speckle image obtained by the experimental device, the linear velocity of each point is fixed and can be calculated as a certain amount of data. Using it as the training target of the model can achieve quantitative evaluation of the model. The experimental device has a total of 10 angular velocity settings.

[0052] The data processing specifically includes the following:

[0053] Data augmentation is achieved by sliding a 32x32 square window from the 32nd pixel to the 320th pixel in steps of 32 to change the radius. Each pixel represents a radius of 0.037 mm, resulting in nine radii. Seventeen sets of data are collected for each radius, resulting in 9x17x10 = 1530 data points for training. This data is partitioned into a training set: validation set: test set ratio of 7:1:2.

[0054] The speckle image behavior feature analysis specifically includes the following contents:

[0055] The pixel values ​​of a single-frame speckle image contain rich spatial information, and the temporal variation of pixel values ​​at a fixed location is closely related to velocity. This means that predicting the motion velocity of a speckle animation depends not only on the spatial pixel values ​​of a single frame but also on the variation in pixel values ​​at fixed locations between frames. The greater the variation in pixel values, the faster the corresponding speckle pattern changes, indicating a higher blood flow velocity. Therefore, velocity prediction is strongly correlated with the spatial and temporal characteristics of speckle animations. Comprehensive and in-depth learning of these features can better exploit the spatiotemporal dynamics of speckle animations.

[0056] The model parameter setting and training specifically include the following:

[0057] The 3D-CNN network in the present invention has five layers. The first layer is the convolution layer C1, which directly accepts the input speckle data and uses a 3*3*5 convolution kernel; the second layer is the maximum pooling layer P2, with a pooling kernel of 2*2*2; the third layer is the convolution layer C3, which uses a 3*3*5 convolution kernel; the fourth layer is the maximum pooling layer P4, with a pooling kernel of 2*2*2; the fifth layer is the convolution layer C5, which uses a 3*3*3 convolution kernel; the sixth layer is the fully connected layer F6, which reconstructs the three-dimensional feature map output by C5 into a single velocity data, with an input dimension of 128 and an output dimension of 1; the convolution layer activation function of this model is the ReLU function, the optimizer uses Adam to accelerate the convergence speed, the batch size is 32, the initial learning rate is 0.001, and the loss function is the MSE function; MSE and MAPE are used as evaluation indicators.

[0058] The model accuracy test specifically includes the following contents:

[0059] The speckle image when the sliding window is moved to positions 97-128 was used as prediction data. This position is roughly in the middle of the entire speckle animation, corresponding to a radius of approximately (97+16)*0.037=4.2mm. The linear velocity varies moderately with the angular velocity, and the overall coverage range is relatively wide, ranging from 4.2*0.047 to 4.2*0.952=0.2mm / s to 4.0mm / s. There are 10 speeds in total, which can effectively test the training level and accuracy of the 3D-CNN model.

[0060] The LSCI images obtained by the rotating scattering plate phantom experimental apparatus are all 320*32*1024 speckle animations. Therefore, the speckle pixel values ​​of each speckle animation can be regarded as a three-dimensional matrix. The angular velocity of the rotating scattering plate phantom experimental apparatus is controlled by changing the rotation speed of the servo motor. A total of 10 angular velocities are set in the present invention, as shown in Table 1.

[0061] Table 1 Angular velocity

[0062]

[0063]

[0064] The formulas for the two evaluation methods of the 3D-CNN are:

[0065]

[0066]

[0067] In Equation 1, m is the number of samples, y i is the actual value, is the predicted value; n is the number of samples in equation 2, y i is the actual value, is the predicted value.

[0068] The speckle dynamic images used for model detection correspond to 10 speeds as shown in Table 2. These ten speeds are obtained by fixing the radius to 4.2 mm.

[0069] Table 2 LSCI image true speed

[0070]

[0071] The size of the input speckle animation for training was set to 32*32*15. This time series length was obtained through extensive manual experimentation and experience, balancing the difficulty of model training with the completeness of the predicted information. Furthermore, the 32*32 square space size is more suitable for feature extraction of the convolution kernel.

[0072] The specific configuration of the 3D-CNN is shown in Table 3.

[0073] Table 2 3D-CNN parameter configuration

[0074]

[0075]

[0076] In the LSCI speckle animation, the pixel values ​​of a single frame contain rich spatial information, and the temporal variation of pixel values ​​at a fixed location is closely related to velocity. This means that predicting the motion speed of a speckle animation depends not only on the spatial pixel values ​​of a single frame but also on the variation in pixel values ​​at fixed locations between frames. Therefore, velocity prediction is strongly correlated with the spatial and temporal characteristics of the speckle animation.

[0077] The pixel value matrix of each speckle animation is a complete three-dimensional matrix. Therefore, the 3D-CNN used in this paper can mine the spatiotemporal features of the above matrix sequence while ensuring the integrity of spatiotemporal information, perform overall feature extraction, and organically complete the prediction.

[0078] In 3D-CNN, the area covered by the convolution kernel is called the receptive field. This is also the size of the area on the input image where pixels in the feature map output by each layer of the network are mapped. Therefore, there is a strong spatiotemporal correlation between the pixel values ​​in the same receptive field. Starting from the input layer, 3D-CNN uses a three-dimensional convolution kernel to extract spatiotemporal information from the pixel matrix, summarizes the features, and forms a new feature map, which is then passed to the next layer for further processing.

[0079] The following will provide a more detailed and complete description of the method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN based on the accompanying drawings.

[0080] The overall prediction schematic diagram of the present invention is as follows Figure 1 As shown in the figure, the model consists of three steps. First, the 3D pixel matrix of the 3D speckle animation is used as the input dataset of the model. Data preprocessing is performed, and the matrix is ​​cut into image sizes and frames suitable for model training. Data augmentation is performed by selecting a radius through a sliding window. Each speckle animation data is matched with a corresponding speed label. The data is then appropriately divided and randomly shuffled to form the final input dataset. This data is then input into a 3D-CNN with set parameters for model training. The 3D convolution, pooling, and fully connected layers in the model extract features and transform the dimensionality of the input data. Finally, the speed corresponding to the speckle animation is obtained and output, completing the training. After training, the model's predictive ability is tested on appropriate speckle animation data, demonstrating that the model has good predictive ability for the corresponding speed of the speckle animation.

[0081] The three-dimensional speckle dynamic image of the present invention is obtained by a rotating scattering plate experimental device. Figure 2 As shown in the figure, the device consists of a servo motor, a uniform diffuse reflector, and a frosted glass reflector. During the experiment, the motor drives the diffuse reflector to rotate behind the frosted glass at a certain angular velocity, simulating the flow of red blood cells under the epidermis in human tissue. When the motor speed (angular velocity) is known, the linear velocity on the diffuse reflector is the product of the angular velocity and the radius of the point, which is a certain amount of data. Compared with using the actual blood flow velocity as the training target, using this quantitative data as the training target of the model can achieve quantitative evaluation of the model. The experimental device is set with a total of 10 angular velocities, and each angular velocity is measured 17 times.

[0082] The original speckle dynamic image of the present invention is as follows Figure 3 As shown in the figure, each speckle animation has a size of 320*32*1024.

[0083] The radius selection method of the present invention is as follows Figure 4 As shown. A 32*32 square sliding window is used to slide from the 1st pixel to the 320th pixel from top to bottom with a step size of 32 to change the radius size. The radius length represented by each pixel is 0.037mm, 9 radii can be obtained, and 17 sets of data are collected for each angular velocity, so 9*17*10=1530 data can be obtained for training. It can be seen that the angular velocity distribution is roughly between 0.05 and 1 rad / s, the radius range is roughly between 2mm and 12mm, and the corresponding speed range is roughly between 0.1mm / s and 12mm / s, which has a wide range and continuity, which is of accuracy significance for model training. The above data is divided into training set: validation set: test set = 7:1:2, and used for model training and testing for 1000 iterations.

[0084] The three-dimensional pixel matrix of the present invention is obtained from the LSCI dynamic image. Assume that the pixel position in a single-frame speckle image is represented by a two-dimensional coordinate (i, j), such as Figure 5 As shown in the figure. Where i and j represent the row and column numbers of the pixel position, respectively. These pixels form a single-frame speckle image with m rows and n columns. In the time dimension, for a pixel at a fixed position, the pixel value is a typical one-dimensional sequence in the time series; in the spatial dimension, for any frame number f, the pixel value at position (i, j) is represented by p(i, j)f, and the pixel data of the entire pixel matrix can be described as an m*n matrix V:

[0085]

[0086] This matrix contains rich spatiotemporal feature information, and each matrix V corresponds to a velocity v. The entire spatiotemporal sequence is a multi-frame motion, such as Figure 6Therefore, the spatiotemporal sequence composed of pixel values ​​representing different speeds can describe the speed through the variation characteristics of the array pixel values. The pixel value sequence composed of N speed speckle images is {V1, V2, ..., V N}.

[0087] During the model training process described in the present invention, the optimizer used Adam to accelerate convergence, the batch size was 32, the initial learning rate was 0.001, and the loss function was the MSE function. MSE and MAPE were used as evaluation indicators.

[0088] The prediction process of the 3D-CNN of the present invention is as follows Figure 7 As shown in the figure. First, 10 32*32*15 3D speckle pixel matrix data are input into the 3D-CNN model and gradient extraction is performed. Then, a 3*3*5 convolution kernel is used for feature extraction, with padding set to 1 and stride set to 1, resulting in a 32*32*13 feature image, where 32 = (32-3+2) / 1+1 and 13 = (15-5+2) / 1+1. Then, maximum pooling is performed with a pooling kernel size of 2*2*2, a stride of 2, and padding set to 1, resulting in a 17*17*7 feature image, where 17 = (32-2+2) / 2+1 and 7 = (13-2+2) / 2+1. Then, the same convolution and pooling are performed again. Finally, a third convolution layer with a convolution kernel size of 3*3*3 is performed to obtain a 9*9*3 feature image. Finally, a fully connected layer is used for dimensionality reduction to obtain 10 data, which are the predicted speed values ​​corresponding to this set of speckle animations.

[0089] The input data for predicting the performance of the test model of the present invention is: the speckle image when the sliding window is moved to the 97-128 position, such as Figure 8 This position is basically in the middle of the entire speckle pattern, with a corresponding radius of approximately (97+16)*0.037=4.2mm. The linear velocity varies moderately with the angular velocity, and the overall coverage is wide, which can effectively detect the training level and accuracy of the 3D-CNN model. The range is 4.2*0.047~4.2*0.952=0.2mm / s~4.0mm / s, with a total of 10 speeds.

[0090] The specific implementation method of the present invention is as follows:

[0091] 1. First, a 320*32*1024 size speckle dynamic image is obtained by rotating the scattering plate phantom experimental device. A total of 10 angular velocities are set, and each angular velocity is measured 17 times.

[0092] 2. Perform data preprocessing on the acquired data.

[0093] (1) Through a large number of manual experiments, we selected a 32*32*15 size speckle image as the input data of the 3D-CNN model. This size setting can achieve the most complete prediction information without increasing the difficulty of model training, and is more suitable for feature extraction of convolution kernels.

[0094] (2) To obtain an appropriate amount of data, data augmentation is performed by selecting a radius. A 32*32 square sliding window is used, sliding from the 32nd pixel to the 320th pixel with a step size of 32 to change the radius. 9 radii can be obtained, and 17 sets of data are collected for each radius. Therefore, 9*17*10=1530 data can be obtained for training.

[0095] (3) The linear velocity corresponding to the above speckle animation data, that is, the product of angular velocity and radius, is used as the predicted label of the data, and the above data is randomly shuffled and divided into data with a division ratio of training set: validation set: test set = 7:2:1.

[0096] 3. Configure the parameters of the 3D-CNN model according to Table 1, which includes three convolutional layers, two pooling layers, and one fully connected layer.

[0097] 4. Input the dataset into the above 3D-CNN model for model training. Set the optimizer to Adam, batch size to 32, learning rate to 0.001, loss function to MSE, evaluation indicators to MSE and MAPE, set the number of training times to 1000, and save the best model.

[0098] 5. Test the prediction performance of the trained model.

[0099] (1) Input data: The speckle image when the sliding window moves to positions 97-128 is selected as the prediction data. This position is basically in the middle of the entire speckle pattern, with a corresponding radius of approximately (97+16)*0.037=4.2mm, and a corresponding linear velocity of 4.2*0.047~4.2*0.952=0.2mm / s~4.0mm / s, with a total of 10 speeds.

[0100] (2) The evaluation method of the model is the linearity between the predicted value and the true value of the speed corresponding to the input speckle pattern. The closer the linearity is to 1, the better the linearity is, and the stronger the model's ability to predict the speed corresponding to the speckle pattern is.

[0101] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN, characterized in that: First, the speckle animation of the rotating scattering plate representing different speeds is used as the input data of 3D-CNN. Since the speckle animation belongs to multi-frame motion from the perspective of time series, and each frame of the speckle image contains a lot of speckle pixel value information from the perspective of spatial sequence, the accuracy of velocity prediction depends on obtaining rich time series information and spatial pixel value information of the single-frame speckle image. Therefore, the three-dimensional convolution kernel in 3D-CNN is used to extract the gradient of the three-dimensional pixel matrix, and the network composed of convolution, pooling and full connection is used to extract features and change the dimension to obtain the blood flow velocity prediction value represented by each speckle animation. Finally, the blood flow velocity prediction value of the speckle animation is output to complete the prediction. Specifically, it includes five steps. The experimental setup consists of a servo motor, a uniform diffuse reflector, and a frosted glass reflector. During the experiment, the motor drives the diffuse reflector to rotate behind the frosted glass at a certain angular velocity, simulating the subcutaneous flow of red blood cells in human tissue. Given a known motor speed, the linear velocity on the diffuse reflector is the product of the angular velocity and the radius of that point. Therefore, in the speckle pattern obtained by this experimental setup, the linear velocity of each point is fixed and can be calculated as a certain amount of data. Using this as a training target for the model allows for quantitative evaluation of the model. The model parameter setting and training specifically include the following: The 3D-CNN network consists of five layers. The first layer is the convolution layer C1, which directly accepts the input speckle data and uses a 3*3*5 convolution kernel; the second layer is the maximum pooling layer P2, with a pooling kernel of 2*2*2; the third layer is the convolution layer C3, with a 3*3*5 convolution kernel; the fourth layer is the maximum pooling layer P4, with a pooling kernel of 2*2*2; the fifth layer is the convolution layer C5, with a 3*3*3 convolution kernel; the sixth layer is the fully connected layer F6, which reconstructs the three-dimensional feature map output by C5 into a single velocity data, with an input dimension of 128 and an output dimension of 1; the convolution layer activation function of this model is the ReLU function, the optimizer uses Adam to speed up the convergence, the batch size is 32, the initial learning rate is 0.001, and the loss function is the MSE function; MSE and MAPE are used as evaluation indicators.

2. The method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN according to claim 1, characterized in that: The data processing specifically includes the following: A 32*32 square sliding window is used to slide from the 32nd pixel to the 320th pixel from top to bottom in a step of 32 to change the radius size to achieve data enhancement. The radius length represented by each pixel is 0.037mm, 9 radii can be obtained, and 17 sets of data are collected for each radius, so 9*17*10=1530 data can be obtained for training; the above data are divided into training set: validation set: test set = 7:1:

2.

3. The method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN according to claim 1, characterized in that: The speckle image behavior feature analysis specifically includes the following contents: The pixel values ​​of a single-frame speckle image contain rich spatial information, and the temporal variation of pixel values ​​at fixed positions is closely related to velocity. This means that the prediction of the motion speed of a speckle animation is not only related to the spatial pixel values ​​of a single frame, but also to the variation of pixel values ​​at fixed positions between frames. The greater the variation in pixel values, the faster the corresponding speckle changes, that is, the faster the blood flow velocity. Therefore, the prediction of velocity is highly correlated with the spatial and temporal characteristics of speckle animations. Comprehensive and in-depth learning of these features can better explore the spatiotemporal dynamic characteristics of speckle animations.

4. The method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN according to claim 1, characterized in that: The model accuracy test specifically includes the following contents: The speckle image when the sliding window is moved to positions 97-128 is used as prediction data. This position is basically in the middle of the entire speckle animation, corresponding to a radius of approximately (97+16)*0.037=4.2mm. The linear velocity varies moderately with the angular velocity, and the overall coverage range is relatively wide, ranging from 4.2*0.047 to 4.2*0.952=0.2mm / s to 4.0mm / s. There are 10 speeds in total, which can effectively test the training level and accuracy of the 3D-CNN model.

5. The method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN according to claim 1, characterized in that: The LSCI images obtained by the rotating scattering plate experimental device are all 320*32*1024 speckle animations, so the speckle pixel values ​​of each speckle animation can be regarded as a three-dimensional matrix. The angular velocity of the rotating scattering plate phantom experimental device is controlled by changing the rotation speed of the servo motor, and a total of 10 angular velocities are set.

6. The method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN according to claim 1, characterized in that: The formulas for the two evaluation methods of the 3D-CNN are: Equation 1 Equation 2 In Equation 1, m is the number of samples, is the actual value, is the predicted value; n in equation 2 is the number of samples, is the actual value, is the predicted value.

7. The method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN according to claim 1, characterized in that: The speckle dynamic images used for model detection correspond to 10 speed magnitudes respectively, and these ten speeds are obtained by fixing the radius to 4.2 mm.

8. The method for improving the accuracy of laser speckle contrast blood flow imaging based on 3D-CNN according to claim 1, characterized in that: The size of the input speckle animation for training is set to 32*32*15. This time series length is obtained through a large number of manual experiments and experience, and a trade-off is made between the difficulty of model training and the completeness of prediction information. In addition, the 32*32 square space size is more suitable for feature extraction of the convolution kernel.

Images