Arc array photoacoustic tomography blood vessel recognition and mask generation method based on physical simulation

Abstract

The invention discloses a physical simulation-based arc array photoacoustic tomography blood vessel recognition and mask generation method, which comprises the following steps of: performing optical and acoustic coupling simulation based on a three-dimensional blood vessel network model, and generating a simulated blood vessel reconstruction image and a reference fact mask which is spatially registered with the simulated blood vessel reconstruction image; constructing a training data set based on the simulated blood vessel reconstruction image and the reference fact mask, and training a deep learning network to obtain a blood vessel segmentation model; acquiring a real-time photoacoustic image and a real-time ultrasonic image, and determining position information of the target area according to the real-time ultrasonic image; cutting the real-time photoacoustic image according to the position information, and inputting the cut photoacoustic image into a blood vessel segmentation model to obtain a blood vessel prediction mask; and fusing the blood vessel prediction mask and the cut photoacoustic image to generate a final blood vessel recognition image. According to the invention, the synchronous generation and image fusion functions of the blood vessel mask are realized, and an effective technical means is provided for high-precision real-time photoacoustic image blood vessel recognition.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of blood vessel imaging, and particularly relates to an arc array photoacoustic tomography blood vessel recognition and mask generation method based on physical simulation. BACKGROUND

[0002] Photoacoustic tomography (PAT) is an emerging hybrid imaging technology that has shown great potential in blood vessel visualization. However, the existing technology still faces many challenges. First, in the image segmentation and recognition tasks based on deep learning, it is extremely difficult and costly to obtain a large number of blood vessel ground truth masks that are accurately registered with photoacoustic images, which leads to a serious lack of high-quality training data sets. Second, due to the limited viewing angle and receiving aperture of the traditional linear array probe, serious artifacts and structural distortion are introduced in the imaging process, making it difficult to accurately restore the true size, shape and connectivity of blood vessels, and the image contrast and signal-to-noise ratio are insufficient. At present, the bottleneck of two-dimensional photoacoustic tomography is that the recognition of blood vessel features still highly depends on the experience of users, and the realization of automatic and intelligent objective analysis is the key to development. Therefore, in practical applications, a real-time and accurate blood vessel mask is of great significance for outlining the blood vessel features. SUMMARY

[0003] The application provides an arc array photoacoustic tomography blood vessel recognition and mask generation method based on physical simulation to solve the problems existing in the prior art.

[0004] To achieve the above purpose, the application provides an arc array photoacoustic tomography blood vessel recognition and mask generation method based on physical simulation, which comprises the following steps: Performing optical and acoustic coupling simulation based on a three-dimensional blood vessel network model to generate a simulated blood vessel reconstruction image and a ground truth mask spatially registered therewith; Constructing a training data set based on the simulated blood vessel reconstruction image and the ground truth mask, and training a deep learning network to obtain a blood vessel segmentation model; Acquiring real-time photoacoustic images and real-time ultrasound images, and determining the position information of a target region according to the real-time ultrasound images; Cutting the real-time photoacoustic images according to the position information, and inputting the cut photoacoustic images into the blood vessel segmentation model to obtain a blood vessel prediction mask; Fusing the blood vessel prediction mask and the cut photoacoustic images to generate a final blood vessel recognition image.

[0005] Optionally, the generation of the simulated blood vessel reconstruction image and the ground truth mask spatially registered therewith comprises: acquire a three-dimensional blood vessel network model, and combine the three-dimensional blood vessel network model with tissue optical parameters to perform optical simulation to obtain an initial pressure matrix and a binary matrix representing a geometric position of a blood vessel; process the binary matrix to generate the ground truth mask; perform acoustic simulation on the initial pressure matrix based on a preset arc array sensor model to obtain original radio frequency signal data; perform image reconstruction on the original radio frequency signal data and fuse gradient information to optimize to obtain the simulated blood vessel reconstruction image.

[0006] Optionally, the optical simulation comprises: a Monte Carlo simulation method is used to calculate the propagation and absorption of photons in the tissue, and the initial pressure matrix and the binary matrix are generated synchronously.

[0007] Optionally, the arc array sensor model is a two-way focusing arc array sensor model.

[0008] Optionally, obtaining the simulated blood vessel reconstruction image comprises: using a delay-and-sum algorithm to reconstruct the original radio frequency signal data to obtain a standard photoacoustic image; performing gradient calculation on the original radio frequency signal data and applying a time sequence accumulation weight to perform reconstruction to obtain a gradient photoacoustic image; fusing the standard photoacoustic image and the gradient photoacoustic image to obtain the simulated blood vessel reconstruction image.

[0009] Optionally, the obtaining of the blood vessel segmentation model comprises: performing size padding and adding multi-scale Gaussian noise processing on the simulated blood vessel reconstruction image, and simultaneously performing synchronous copying on the ground truth mask to obtain an enhanced training data set; inputting the enhanced training data set into a U-net network for training to obtain the blood vessel recognition pre-training model.

[0010] Optionally, the training of the U-net network comprises: dividing the enhanced training data set into a training set and a validation set according to a preset ratio; using an enhanced U-Net network structure comprising residual connection, deep supervision, depth separable convolution, hybrid attention module, attention gate and ASPP module; using a composite loss function combining Dice loss and foreground weighted loss for training to obtain the pre-training model.

[0011] Optionally, the determining of the position information of the target region according to the real-time ultrasound image comprises: Image analysis algorithms are used to identify hypoechoic or anechoic areas caused by lesions in real-time ultrasound images. Calculate the center position and range of the region in the horizontal direction to generate the clipping coordinates; The real-time photoacoustic image is cropped according to the cropping coordinates to obtain the lesion area of ​​the photoacoustic image.

[0012] Optionally, generating the final blood vessel recognition image includes: The pixel values ​​of the blood vessel prediction mask are linearly scaled to a preset intensity range; The scaled mask is multiplied pixel by pixel with the lesion area in the photoacoustic image to obtain the final blood vessel recognition image.

[0013] Compared with the prior art, the present invention has the following advantages and technical effects: This invention achieves significant technical results by constructing a complete workflow from physical simulation to deep learning model application. It solves the problem of insufficient high-quality simulation datasets, enabling the automated generation of massive amounts of spatially perfectly registered paired simulation images and baseline fact masks. By customizing a dual-focusing arc array sensor model in acoustic simulation and combining it with a gradient fusion reconstruction algorithm, it effectively suppresses the spectral limitations, artifacts, and structural distortions inherent in traditional linear array probes, significantly improving the quality of simulated reconstructed vascular images. The deep learning model trained on this high-quality simulation dataset possesses powerful vascular feature extraction and anti-interference capabilities. Applying the trained model to real-time images enables simultaneous generation of vascular masks and image fusion, providing an effective technical means for high-precision real-time photoacoustic image vascular recognition. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of a method according to an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram illustrating the results of an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] This invention provides a method for real-time generation of blood vessel identification and masks in arc-array photoacoustic tomography based on physical simulation. The process is as follows: First, a three-dimensional blood vessel network model is acquired. Then, the model is optically simulated to obtain an initial pressure matrix containing the light flux distribution and a binarized matrix accurately representing the blood vessel location. Next, the binarized matrix is ​​processed to output a reference fact mask that is perfectly spatially registered with the blood vessel identification image. Simultaneously, the initial pressure matrix is ​​acoustically simulated based on a preset dual-path focusing arc-array sensor model to obtain raw radio frequency (RF) signal data. Then, these raw RF signal data are used for image reconstruction, and their gradient information is fused for optimized overlay. A simulated vascular reconstruction image is obtained. Based on this, data augmentation processing is performed on the simulated vascular reconstruction image and the baseline fact mask to construct an expanded augmented training dataset. Subsequently, the augmented training dataset is input into the U-net network for training to obtain a pre-trained model. In the real-time application stage, the lesion region is first located based on the real-time ultrasound image to obtain the target location information. Then, the real-time photoacoustic image is cropped based on the location information to obtain the photoacoustic image lesion region. Then, the cropped photoacoustic image lesion region is input into the pre-trained model to obtain the vascular prediction mask in real time. Finally, the vascular prediction mask and the photoacoustic image lesion region are fused to generate the final vascular recognition image.

[0019] Example 1 like Figure 1 As shown, this embodiment provides a method for blood vessel identification and mask generation in arc array photoacoustic tomography based on physical simulation, including the following steps: Optical and acoustic coupling simulation is performed based on a three-dimensional vascular network model to generate simulated vascular reconstruction images and a reference fact mask spatially registered with them. A training dataset is constructed based on simulated blood vessel reconstruction images and benchmark fact masks, and a deep learning network is trained to obtain a blood vessel segmentation model. Acquire real-time photoacoustic images and real-time ultrasound images, and determine the location information of the target area based on the real-time ultrasound images; The real-time photoacoustic image is cropped based on the location information, and the cropped photoacoustic image is input into the blood vessel segmentation model to obtain the blood vessel prediction mask; The blood vessel prediction mask is fused with the cropped photoacoustic image to generate the final blood vessel recognition image.

[0020] Specifically, the following steps are included: Step 1: Obtain a 3D vascular network model; Multiple basic vascular networks with fundamental anatomical structures were generated using 3D modeling software (Blender). Within the MATLAB environment, the MCmatlab toolkit was used to programmatically overlay pathological features onto the imported basic vascular network models. This included creating spherical clusters of randomly scattered microcalcifications, applying Gaussian smoothing filters to the 3D random noise field to simulate non-uniform adipose tissue distribution, generating skin layers (epidermis and dermis) with physiological curvature using large-radius circular arcs and sinusoidal perturbations, and defining a coupling medium above the epidermis.

[0021] Step 2: Perform optical simulation; By combining a three-dimensional vascular network model with preset tissue optical parameters, a three-dimensional tissue optical media property distribution map is obtained. Each media ID representing blood, epidermis, dermis, and fat in the model is assigned precise optical parameters at a wavelength of 1000 nm, including absorption coefficient (µa), scattering coefficient (µs), anisotropy factor (g), and refractive index (n). Based on this optical property distribution map, Monte Carlo optical simulation is run to calculate the propagation and absorption of photons within the tissue, thereby simultaneously generating two spatially perfectly registered core data sets. 1) An initial pressure matrix containing the light flux distribution; 2) A binarized matrix containing only vascular geometry information.

[0022] Step 3: Generate a baseline fact mask; All voxels labeled as blood vessels (medium ID 4) are extracted from the tissue optical medium property distribution map to form a pure three-dimensional blood vessel binarization matrix. The three-dimensional binarization matrix is ​​then subjected to maximum intensity projection along the elevation direction (Z-axis) to generate a two-dimensional mask. Functions such as `imresize` are used to perform geometric transformations and scaling on this image to ensure its resolution matches the final reconstructed image (e.g., 1200x400 pixels). The processed mask is then grayscale normalized, outputting a baseline fact mask that is perfectly registered in space with the simulated blood vessel image.

[0023] Step 4: Perform acoustic simulation; In the MATLAB environment, the k-Wave toolbox was used to construct an acoustic heterogeneous propagation medium model based on the tissue geometry generated by optical simulation, assigning corresponding sound velocities and densities to each medium. An initial pressure matrix was used as the sound source, and a two-way focusing arc array sensor model was deployed. This model was designed according to actual hardware parameters, containing 512 array elements arranged along a 130mm radius of curvature to achieve in-plane focusing; it also has an elevation focusing radius of 135mm (unit elevation of 29mm), aiming for acoustic focusing at a depth of 120mm. The kspaceFirstOrder3D function was used to simulate acoustic wave propagation and obtain the raw RF signal data.

[0024] Step 5: Perform image reconstruction and optimization; Based on the sensor's geometric model, a sampling frequency of 20 MHz, and a medium sound velocity of 1540 m / s, a delay matrix and an angle weight matrix are pre-calculated and generated. The original radio frequency signal data is reconstructed using a delay-and-sum algorithm to obtain a "standard photoacoustic image." Gradient calculations are performed along the time dimension on the original radio frequency signal data, and a time-accumulated weight is applied to amplify the calculated negative gradient signal. This reconstruction process is repeated to obtain a "gradient photoacoustic image." The grayscale values ​​of the two images are directly added and non-negatively processed to fuse them into a high-contrast simulated blood vessel reconstruction image.

[0025] Step Six: Expand and preprocess the dataset; A 1200x400 pixel simulated blood vessel reconstruction image and a baseline fact mask were placed in the center of a 1216x416 pixel canvas, making their size consistent with the preset input size of the U-Net network. For each size-filled simulated blood vessel reconstruction image, Gaussian noise with preset noise variances of 0, 0.01, 0.03, and 0.05 was added, creating four enhanced input images with different noise levels. Each size-filled baseline fact mask was also duplicated four times.

[0026] Step 7: Train the U-Net network; A dataset consisting of paired 1216x416 pixel "simulated blood vessel reconstruction images" and "baseline fact masks" was used, divided into an 80% training set and a 20% validation set. An enhanced U-Net network was trained using this dataset. The network integrates residual connections, deep supervision, depthwise separable convolutions, a hybrid attention module, attention gating, and an ASPP module. A multi-level weighted composite loss function is used for optimization. This function calculates the Dice loss and foreground weighted loss between the predicted masks and the baseline fact mask for both the main output branch and all auxiliary branches. The losses of all branches are then weighted and summed according to preset weights (e.g., 1.0 for the main output and progressively decreasing weights for auxiliary outputs) to obtain the final total loss, which is used to update the network parameters. Training uses a batch size of 12, an Adam optimizer with an initial learning rate of 1e-3, and a learning rate scheduling strategy with an early stopping mechanism (Patience 15). Mixed precision computation is used throughout the training process, and the best-performing pre-trained model on the validation set is saved.

[0027] Step 8: Perform real-time photoacoustic image recognition and mask generation; Image analysis algorithms are used to identify hypoechoic or anechoic dark gray shadow areas caused by lesions in real-time ultrasound images. The center position and extent of this shadow area in the horizontal direction are calculated, generating a set of precise cropping coordinates. These horizontal cropping coordinates are mapped to the photoacoustic image coordinate system, and the synchronously acquired real-time photoacoustic image is precisely cropped to obtain the photoacoustic image-localized lesion area. The cropped photoacoustic image (sized and padded to match the network input) is input into a pre-trained model to generate a vessel prediction mask in real time. The pixel values ​​of the prediction mask are linearly scaled to a preset intensity range (e.g., 0.3 for background, 1.0 foreground), and then multiplied pixel-by-pixel with the photoacoustic image-localized lesion area to finally output a brightly lit vessel recognition image.

[0028] The above method was experimentally verified, and the results are as follows: Figure 2 As shown, the image includes the original photoacoustic image as input, a blood vessel prediction mask processed by the model, and the blood vessel recognition result after overlaying the prediction mask onto the original photoacoustic image. Figure 2 It is evident that the present invention can accurately identify blood vessels.

[0029] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for blood vessel identification and mask generation in arc-array photoacoustic tomography based on physical simulation, characterized in that, The method comprises the following steps: Based on the three-dimensional vascular network model, optical and acoustic coupling simulation is performed to generate a simulated blood vessel reconstruction image and a benchmark mask spatially registered therewith; Based on the simulated blood vessel reconstruction image and the benchmark mask, a training data set is constructed, and a deep learning network is trained to obtain a blood vessel segmentation model; Real-time photoacoustic images and real-time ultrasound images are acquired, and the position information of the target region is determined according to the real-time ultrasound images; The real-time photoacoustic images are cropped according to the position information, and the cropped photoacoustic images are input into the blood vessel segmentation model to obtain a blood vessel prediction mask; The blood vessel prediction mask is fused with the cropped photoacoustic images to generate a final blood vessel identification image.

2. The method of claim 1, wherein, The generation of the simulated blood vessel reconstruction image and the benchmark mask spatially registered therewith comprises: A three-dimensional vascular network model is acquired, and optical simulation is performed on the three-dimensional vascular network model in combination with tissue optical parameters to obtain an initial pressure matrix and a binary matrix representing the geometric position of the blood vessels; The binary matrix is processed to generate the benchmark mask; The initial pressure matrix is subjected to acoustic simulation based on a preset arc array sensor model to obtain original radio frequency signal data; The original radio frequency signal data is subjected to image reconstruction and optimization by fusing gradient information to obtain the simulated blood vessel reconstruction image.

3. The method of claim 2, wherein, The optical simulation comprises: The propagation and absorption of photons in the tissue are calculated by using a Monte Carlo simulation method to synchronously generate the initial pressure matrix and the binary matrix.

4. The method of claim 2, wherein, The arc array sensor model is a double-path focused arc array sensor model.

5. The method of claim 2, wherein, The generation of the simulated blood vessel reconstruction image comprises: The original radio frequency signal data is reconstructed using a delay-and-sum algorithm to obtain a standard photoacoustic image; Gradient calculation is performed on the original radio frequency signal data, and a time sequence accumulation weight is applied, and then reconstruction is performed to obtain a gradient photoacoustic image; The standard photoacoustic image and the gradient photoacoustic image are fused to obtain the simulated blood vessel reconstruction image.

6. The method of claim 1, wherein, The generation of the blood vessel segmentation model comprises: The simulated blood vessel reconstruction image is subjected to size padding and addition of multi-scale Gaussian noise processing, and the benchmark mask is synchronously copied to obtain an enhanced training data set; The enhanced training data set is input into a U-net network for training to obtain the blood vessel identification pre-training model.

7. The method of claim 6, wherein, The training of the U-net network comprises: The enhanced training data set is divided into a training set and a validation set according to a preset ratio; An enhanced U-Net network structure comprising residual connection, deep supervision, depth separable convolution, hybrid attention module, attention gate and ASPP module is adopted; A compound loss function combining Dice loss and foreground weighted loss is used for training to obtain the pre-training model.

8. The method of claim 1, wherein, The determination of the position information of the target region according to the real-time ultrasound images comprises: A low echo or no echo region caused by a lesion in the real-time ultrasound image is identified by an image analysis algorithm; The center position and range of the region in the transverse direction are calculated to generate a cropping coordinate; According to the clipping coordinates, a real-time photoacoustic image is clipped to obtain a lesion region of the photoacoustic image.

9. The method of claim 1, wherein, The generating of the final blood vessel recognition image comprises: linearly scaling pixel values of the blood vessel prediction mask to a preset intensity interval; multiplying the scaled mask and the lesion region of the photoacoustic image pixel by pixel to obtain the final blood vessel recognition image.

Images