Infrared thermal image enhanced real-time rectal blood vessel blood flow detection method

Abstract

The invention discloses an infrared thermal image enhanced real-time rectal blood vessel blood flow detection method. The method comprises the following steps that S1, a thermal image of an operation area is collected through an infrared thermal imager to serve as a to-be-processed medical purpose image; s2, adopting a double-branch collaborative network architecture, adopting an improved U-Net structure for a deep denoising priori network DDPNet, introducing a detail sensing unit DAU, and separating noise and organizational structure features through residual learning; and S3, obtaining a supervision signal in the image, and processing each supervision model to obtain a processed image. The method can effectively improve the accuracy and comprehensiveness of information obtained after image processing in the operation and improve the effectiveness of information acquisition in the operation process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical imaging technology, and in particular to a system and method for real-time monitoring of blood flow during radical resection of rectal cancer surgery by combining infrared thermal imaging technology and real-time image enhancement algorithm. BACKGROUND

[0002] In the medical field, rectal cancer, as one of the common malignant tumors of the gastrointestinal tract, seriously threatens human health. According to relevant data, the incidence of rectal cancer in men is higher than that in women, and the relative high-risk age is 40-80 years old, and the incidence increases with age. At present, surgical resection is still an important means to cure rectal cancer, and radical resection of rectum is suitable for various conditions, such as low rectal cancer, etc.

[0003] In recent years, infrared thermal imaging technology has gradually attracted widespread attention in radical resection of rectal cancer due to its non-invasive and real-time monitoring of tissue blood flow and temperature changes. Its principle is to image according to the infrared radiation difference of the object, which can convert the infrared radiation distribution emitted by the object surface into a visible image. During the operation, this technology can judge the blood flow state of the tissue, and can prevent surgical risks and evaluate the effect of the operation during the operation.

[0004] However, in the actual operation scene, the infrared image faces many severe challenges. On the one hand, there are factors such as the reflection of the surface of surgical instruments, the illumination of surgical lamps, the approach of the hands of medical staff, and the movement caused by the breathing, heartbeat and operation displacement of the tissue itself. These factors cause the noise of the infrared image to increase significantly, and the contrast to decrease greatly, which seriously affects the accuracy of the judgment of the blood flow state during the operation. On the other hand, traditional image enhancement algorithms, such as histogram equalization and wavelet denoising, have exposed a series of problems in the dynamic operation environment. First, a single denoising or enhancement method is difficult to effectively retain image details while suppressing noise. Second, such algorithms lack targeted processing of multi-scale features, which can easily cause loss of high-frequency information, making the clarity and accuracy of the image greatly discounted. Third, traditional algorithms have deficiencies in real-time performance, which cannot meet the timeliness requirements of dynamic monitoring during the operation.

[0005] In addition, although existing deep learning models perform well in static image processing, they have obvious insufficient adaptability when faced with the dynamic characteristics of intraoperative infrared images, such as rapid tissue deformation and time-varying noise. These models have high complexity and are difficult to achieve efficient deployment in medical devices, limiting their application in actual operations.

[0006] In summary, in the actual operation of radical resection of rectal cancer, there is an urgent need for a technical solution that can enhance the quality of infrared images in real time and accurately, and dynamically feedback blood flow information, in order to improve the accuracy and safety of the operation and provide more reliable support for the treatment of patients.

[0007] The information disclosed in the background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0008] This invention addresses the technical problem of poor real-time performance and accuracy of dynamic monitoring using infrared thermography during radical resection of rectal cancer. It discloses a real-time rectal vascular blood flow detection method enhanced by infrared thermography. This method can effectively improve the accuracy and comprehensiveness of information obtained after intraoperative image processing and improve the effectiveness of information acquisition during surgery.

[0009] This application provides a method for real-time detection of rectal blood flow with infrared thermography enhancement.

[0010] The beneficial effects that this application can produce include:

[0011] 1) The infrared thermo-enhanced real-time rectal vascular blood flow detection method provided in this application acquires thermal images of the surgical area using an infrared thermal imager, and uses image enhancement algorithms to improve the clarity of vascular images and the visibility and clarity of vascular edges in real time. Subsequently, by analyzing the perfusion characteristics of the enhanced images, the tissue blood flow status is evaluated to assist surgeons in accurately judging tissue activity and blood supply. The images obtained by this method have the advantages of real-time and high sensitivity, which can significantly improve the accuracy and efficiency of intraoperative judgment, reduce the incidence of complications, and have broad clinical application prospects. Attached Figure Description

[0012] Figure 1 A schematic flowchart of a real-time rectal blood flow detection method with infrared thermography enhancement in at least one embodiment provided in this application;

[0013] Figure 2 A flowchart of step S1 is provided in at least one embodiment of this application;

[0014] Figure 3 A flowchart of step S2 is provided in at least one embodiment of this application;

[0015] Figure 4 A flowchart of step S3 is provided in at least one embodiment of this application;

[0016] Figure 5 A flowchart of step S4 is provided in at least one embodiment of this application;

[0017] Figure 6 A flowchart of step S5 is provided in at least one embodiment of this application; Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] See Figures 1-6 The infrared thermo-enhanced real-time rectal vascular blood flow detection method provided in this application includes the following steps:

[0021] Step S1: Multimodal dynamic registration and feature-level fusion

[0022] To acquire images for medical use, registration and fusion of images from different light sources are performed before dynamic noise estimation and preprocessing to facilitate downstream enhancement tasks. Real-time non-rigid registration is employed using a lightweight deformation field estimation network (DFNet), based on optical flow and thin-plate spline transformation (TPS) to achieve fast registration. It is generated by the following formula:

[0023]

[0024] in, This is a reference frame (which can be a preoperative planning image). For real-time intraoperative frames, the network outputs the control point displacement matrix to generate a dense deformation field through TPS interpolation;

[0025] After registration, infrared (IR) and visible light (RGB) images are used to extract features using a dual-branch encoder to obtain infrared features. and visible light characteristics :

[0026]

[0027] in, Infrared image, Visible light image;

[0028] Fusion features are obtained through attention mechanisms. :

[0029]

[0030] in, For dynamic weights, For adaptive weights, For the Sigmoid function, The fused feature map;

[0031] Attention mechanisms use dynamic weights Control infrared With visible light The core of feature fusion ratio is to adaptively select complementary information from two modalities through a data-driven approach.

[0032] infrared features With visible light characteristics By concatenating the data along the channel dimension, a joint feature representation is formed, providing a foundation for cross-modal interaction and capturing the spatial alignment information of the two modalities. Convolutional layer The number of feature channels after splicing is compressed to generate a spatial attention weight map.

[0033] Step S2: Image fusion using a dual-branch collaborative network architecture (DDPNet+CIMNet)

[0034] A dual-branch collaborative network architecture is adopted, in which the deep denoising prior network (DDPNet) uses an improved U-Net structure and introduces detail-aware units (DAUs) to separate noise and tissue structure features through residual learning. The DAU module is embedded in the encoder stage to dynamically extract local detail residuals.

[0035]

[0036] in, For the input image The output feature map after convolution operation This represents the output value of the center pixel of the convolution kernel. Combined with an adversarial training strategy, the discriminator outputs an adversarial loss to constrain the realism of the generated image.

[0037] The contrast enhancement module (CIMNet) decomposes the image into low-frequency contrast components through multi-scale decoupling and Gaussian pyramid decomposition. ) and high-frequency detail components ( ). No. Generation of low-frequency components in the layer:

[0038]

[0039] in The number of decomposition layers, For the first Low-frequency components of the layer, Let be a Gaussian kernel function with kernel weights of . Indicates the convolution operation; Initial conditions for the downsampling operator (usually intermittent sampling) .

[0040] No. Layer high-frequency detail residual calculation:

[0041]

[0042] in, This is an upsampling operator (bilinear interpolation). For the first Layer high frequency components, For the first -1 layer high-frequency components, For the first Detail residuals in high-frequency layer images.

[0043] Simultaneously, dual-branch enhancement, including contrast and detail branches, further enhances image quality. The contrast branch employs global histogram stretching and local adaptive gamma correction to enhance vascular structures.

[0044]

[0045] in, These are input features. Input features The result after processing by the Clip function and These are weighting coefficients. It is Local Adaptive Gamma Correction.

[0046] The detail branch introduces detail maps from DDPNet for feature stitching and expands the receptive field through dilated convolution.

[0047] Step S3: Multi-scale joint loss function

[0048] To achieve joint optimization of global structure and local details, this invention integrates four monitoring signals, including adversarial loss (…). Multi-scale SSIM loss ), contrast loss ( and loss of detail () The adversarial loss is based on the Wasserstein GAN optimization strategy, which constrains the distribution of the generated image by the discriminator output; the multi-scale SSIM loss measures the image fidelity by multi-scale structural similarity; the contrast loss constrains the similarity between low-frequency components and the target image; and the detail loss forces high-frequency details to align with the target residual.

[0049] Step S4: Real-time dynamic inference optimization technology

[0050] To achieve real-time dynamic inference optimization, this invention employs a lightweight design, performing channel pruning on DDPNet and CIMNet while preserving key feature channels:

[0051]

[0052] in, This represents the total number of channels in the convolutional layer. It is the sum of the absolute values ​​of the convolution kernel weights. To retain the previous It has several important channels, and performs hardware acceleration at the same time. It utilizes CUDA core parallel computing to map convolution operations to GPU thread blocks, enabling fast processing of each frame.

[0053] Specifically, it includes the following steps:

[0054] Step S1: Multi-source image registration and fusion. To provide a good foundation for downstream enhancement tasks, this step registers and fuses multi-source images, and then provides the resulting image to the downstream image enhancement task.

[0055] Step S101: Multi-source data input and preprocessing: In this step of the present invention, the input data are infrared images and visible light images, and RGB visible light images are captured simultaneously. , used for anatomical reference, and the infrared image is normalized to [0,1]:

[0056]

[0057] in, The mean and standard deviation are the values ​​pre-calibrated. For the first The original infrared image of the frame, The infrared image was standardized. The RGB image was converted to the YUV color space, and the luminance channel was extracted. Used for registration.

[0058] Step S102: Dynamic non-rigid registration. This step in the present invention uses static infrared images from the preoperative planning stage. or the first frame during surgery As a reference frame. The Deformation Field Estimation Network (DFNet) is used to input the reference frame. With the current frame The stitched image. Output control point displacement matrix. (N represents the grid size, which is set according to the image dimensions). The network architecture used is the lightweight U-Net, which contains four downsampling layers and skip connections, with the last layer being a fully connected layer for prediction. A global smooth deformation field is generated based on the control point displacement using thin plate spline (TPS) interpolation. :

[0059]

[0060] in, For radial basis functions, the coefficients are... The solution is obtained using the least squares method. Furthermore, a deformation field is applied to resample the current frame during image alignment.

[0061]

[0062] in In the context of image transformation, for the function and images , This means that the deformation field Applied to the current frame image The aligned image is obtained by transforming (resampling) each pixel in the image according to the deformation field. In other words, first, based on the deformation field... Transform the coordinates, and then use the transformed coordinates to find the original image. The corresponding pixel values ​​are obtained from the image and used to construct a new aligned image.

[0063] Optimization is performed by constraining the smoothness of the deformation field using bidirectional optical flow uniformity loss:

[0064]

[0065] in, Indicates starting from the first Frame to the The deformation field transformation of the frame, and then from the first frame... Frame to the The deformation field transformation of the frame is calculated. (here The L2 norm is used to measure the difference from the identity transformation, thereby constraining the smoothness of the deformation field for optimization.

[0066] S103: Multimodal spatial alignment. Since this invention uses infrared and visible light images, a multimodal spatial alignment method is used for registration to provide a better foundation for downstream tasks.

[0067] First, cross-modal registration is performed, using a weighted DFNet to align the infrared and visible light images:

[0068]

[0069] And resample the RGB image:

[0070]

[0071] Next, resolution adaptation is performed by double-cubic downsampling of the RGB image to match the infrared image size.

[0072] S104: Feature-level multimodal fusion. In this step of the invention, a dual-branch encoder is first used to extract features, where the infrared branch uses convolution (Conv+ReLU) to extract blood flow features. The RGB branch (visible light branch) uses a pre-trained ResNet-18 (with fully connected layers removed) to extract anatomical features. .

[0073] Secondly, cross-modal attention fusion is performed, and channel attention weights are calculated. :

[0074]

[0075] in, This is the Sigmoid function.

[0076] Then, weighted feature fusion is performed:

[0077]

[0078] It is a pixel-wise multiplication (Hadamard product) that dynamically adjusts the contribution ratio of the two modalities to achieve complementary feature-level information.

[0079] Finally, the fused features are obtained by transposing the convolution. Upsampled to the original resolution, output fused image .

[0080] The output of step S1 of this invention is a registered and aligned single-mode infrared image. or multimodal fusion image , as input to downstream enhancement modules.

[0081] S2: Dynamic noise estimation and preprocessing. This step in the invention uses the fused infrared image as input. First, normalize to the range [0,1]. Then, calculate the local noise variance map based on a sliding window. Finally, and Concatenate as DDPNet input, i.e. .

[0082] S201: Input image standardization. First, normalization is performed on the input image, then the merged image is... Pixel values ​​are linearly mapped to the range [0,1].

[0083]

[0084] The minimum value of the pixels, The maximum value of the pixels, This is a normalized image; the infrared and visible light channels need to be normalized separately.

[0085] Secondly, spatial smoothing filtering is performed, using a Gaussian filter to suppress high-frequency noise:

[0086]

[0087] S202: Dynamic noise level estimation. First, local variance is calculated by traversing the image using a sliding window, calculating the noise variance for each window:

[0088]

[0089] in, This represents the mean value within the window.

[0090] Secondly, motion compensation is corrected by utilizing the registered deformation field. Jacobian determinant Suppressing noise overestimation in motion regions:

[0091]

[0092] This indicates a locally compressed region, and the noise variance needs to be attenuated according to the degree of deformation.

[0093] Then, the global noise is modeled, the noise variance-mean relationship curve is fitted, and a noise level map is established. :

[0094]

[0095] Among them, coefficient Real-time fitting of all windows using the least squares method Data pair.

[0096] S203: Enhancement of Spatiotemporal Consistency. This step of the invention enhances spatiotemporal consistency, including temporal weighted filtering and spatial guided filtering. Temporal weighted filtering combines the noise maps from the previous few frames to calculate the temporally smoothed noise level:

[0097]

[0098] Spatial guided filtering To guide the image, Maintain smooth edges:

[0099]

[0100] S204: Construction of input features. The preprocessed image is concatenated with the noise map to form the input tensor of DDPNet:

[0101]

[0102] The first channel is a smoothed image, and the second channel is a noise level map.

[0103] S205: Outlier Noise Detection and Suppression. This step of the invention includes outlier noise detection and nonlocal mean correction. Outlier noise detection identifies regions where the noise level exceeds three standard deviations of the global mean. hour, 1, and 0 in all other cases. for The global mean and standard deviation are then calculated. Next, a fast nonlocal mean filter is applied to the labeled region using nonlocal mean correction.

[0104]

[0105] The input tensor constructed in step S2 of this invention is... It contains a smoothed image and a noise level map, which are used by DDPNet for denoising and detail extraction.

[0106] S3: Multi-scale thermal image enhancement, denoising, and detail extraction (DDPNet). In this step of the invention, the encoder path extracts multi-scale features through four levels of downsampling, and each level embeds a DAU module to generate a detail residual map. The decoder upsamples and fuses detail residuals with contextual features to output a preliminary denoised image. and detailed pictures .

[0107] S301: Encoder multi-scale feature extraction. In this step of the invention, the preprocessed noisy level map is first used as input.

[0108] Secondly, four-level downsampling encoding is performed.

[0109] Level 1 is the original resolution sampling encoding. Specifically, it first undergoes a 3×3 convolution, followed by ReLU activation, then a detail-aware unit (DAU) module to extract local detail residuals and suppress noise interference. Finally, max pooling is performed to obtain the output feature map. Levels 2 to 4 are downsampled step by step, repeating the following operation:

[0110]

[0111] Output feature map scale: , , .

[0112] The mathematical expression of the detail perception unit is as follows:

[0113]

[0114] Extract the center pixel value of the 3×3 convolution kernel (i.e., ignore edge response).

[0115] S302: Decoder multi-scale feature fusion. This step in the invention includes four levels of upsampling decoding, specifically as follows:

[0116] First, the input for levels 4 to 3 is... This step involves first transposing the convolution, then concatenating it with the features from level 3, followed by a convolutional layer, and finally ReLU activation. Its output is...

[0117] Next is level 3 to level 2, where the input is... This is concatenated with the level 2 features. The operation is the same as the previous one, and the output is... .

[0118] Then it's from level 2 to level 1, and the input is... Concatenated with the features of level 1, the output is .

[0119] Next comes the final upsampling, followed by transposing the convolution to restore the original resolution, outputting a preliminarily denoised image. .

[0120] Finally, residual connections enhance details, adding detail residuals from the DAU outputs of each layer. Weighted fusion:

[0121]

[0122] Final denoising result:

[0123]

[0124] It is the weight of the detailed residuals after weighted fusion.

[0125] S303: Adversarial Training Optimization. This step in the invention uses a discriminator network, with the following structure: 5 convolutional layers → fully connected layers → Sigmoid output probability. The input is a stitched, denoised image. With noise level map .

[0126] In this step, the adversarial loss is calculated using the Wasserstein GAN-GP loss:

[0127]

[0128] in, This is a random interpolation between real and generated samples. . This refers to the discriminator. For real samples The average score output. This refers to the discriminator. For denoised images The average score output. This refers to the discriminator. For random interpolated samples The average of the squared differences between the magnitude (L2 norm) of the gradient and 1.

[0129] S304: Multi-level supervision and loss calculation. This step in the invention uses multiple losses for supervision and optimization, as detailed below:

[0130] 1. Pixel-level L1 loss, its mathematical expression is:

[0131]

[0132] 2. Multi-scale SSIM loss, its mathematical expression is:

[0133]

[0134] Indicates the downsampling scale. and The size depends on the situation.

[0135] 3. Loss of detail retention on detail residuals Apply sparsity constraints:

[0136]

[0137] 4. The total loss function is

[0138]

[0139] The output of step S3 of this invention includes a denoised image. and detailed residual diagram For use by CIMNet enhancement.

[0140] S4: Contrast-Detail Decoupling Thermal Imaging Enhancement (CIMNet). This step in the invention... Perform a 3-layer Gaussian pyramid decomposition to obtain low-frequency components. With high frequency components Adaptive histogram stretching is applied to the low-frequency components of each layer, using the following formula: ,in, For learnable parameters, The value is not 0. The DDPNet's... It is spliced ​​with high-frequency components from each layer, and the edges are enhanced by dilated convolution. By using channel attention (CALayer) to weighted fuse features at various scales, we obtain... .

[0141] S401: Input Feature Preprocessing. The input data for this step in this invention is a denoised image. (From DDPNet output) and detailed residual map (Detailed features extracted by DDPNet).

[0142] After inputting the data, feature concatenation is performed first, concatenating the detail residuals with the denoised image to form the input tensor:

[0143]

[0144] S402: Multi-scale Gaussian pyramid decomposition. This step in the invention first constructs a Gaussian pyramid by performing three layers of Gaussian downsampling on the input image to generate multi-scale features:

[0145]

[0146] It is the original scale. The original input image (Right now After applying Gaussian blur (smoothing with Gaussian kernel), downsampling is performed. Yes Further Gaussian blurring and downsampling are performed, that is, the Gaussian pyramid operation is repeated on the output of the previous layer.

[0147] Secondly, the Laplace residual is calculated, and high-frequency details of each layer are extracted by upsampling difference:

[0148]

[0149] Output residual ,in For the finest details.

[0150] S403: Contrast Enhancement. This step in the invention mainly focuses on low-frequency optimization, including low-frequency component extraction, adaptive histogram stretching, and local gamma correction. Low-frequency component extraction is performed on each Gaussian image layer. Perform low-pass filtering to obtain the low-frequency components:

[0151]

[0152] Adaptive histogram stretching dynamically adjusts the contrast range of low-frequency components in each layer:

[0153]

[0154] : The mean and standard deviation of the low-frequency components of the current layer. Learnable parameters, predicted via 1×1 convolution.

[0155] Local gamma correction adjusts the gamma value according to the spatial brightness distribution. ):

[0156]

[0157] Predicted by a small CNN, with the input being the features of the current layer.

[0158] S404: Detail Enhancement. This step in the invention mainly involves high-frequency enhancement, including detail feature fusion, dilated convolution enhancement, and detail sharpening.

[0159] Detail feature fusion will integrate the Laplace residuals Detailed residuals with DDPNet splicing:

[0160]

[0161] Dilated convolution enhancement uses multi-scale dilated convolution (dilation=1,2,3) to extract details from multiple receptive fields:

[0162]

[0163] Dilated convolution kernel weights are adaptively fused through an attention mechanism.

[0164] Detail sharpening enhances edges through high-pass filtering and residual concatenation:

[0165]

[0166] The sharpening intensity coefficient is dynamically predicted by the network.

[0167] S405: Multi-scale feature fusion. This step of the invention fuses multi-scale features, first performing cross-scale feature alignment, and then assembling the enhanced low-frequency features of each layer (…). ) and high frequency ( Perform bilinear upsampling to the original resolution.

[0168] Secondly, weighted channel attention (CALayer) is used to calculate the attention weights for features at each scale:

[0169]

[0170] GAP stands for Global Average Pooling, which outputs weights. .

[0171] Then output the weighted fusion:

[0172]

[0173] S406: Residual Connection and Output. This step in the invention includes residual enhancement and dynamic range compression. Residual enhancement involves adding the fused result to the original input while preserving the basic structure:

[0174]

[0175] Dynamic range compression uses a learnable S-curve to prevent over-enhancement.

[0176]

[0177] parameter Predict from global features using fully connected layers.

[0178] The final output of step S4 of this invention is the final enhanced image. Attention weights at various scales (For visualization analysis)

[0179] S5: Hemodynamic parameter calculation. This step in the present invention is based on enhanced images. Calculate spatial gradient The proportionality coefficient was determined through calibration experiments. Calculate blood flow velocity Set a threshold. If the area If the frame rate is below the threshold for several consecutive frames, an audible and visual alarm will be triggered.

[0180] S501: Temperature Field Modeling and Calibration. The first step in this invention is pixel-temperature mapping, establishing a linear relationship between grayscale values ​​and actual temperature through pre-operative calibration experiments.

[0181]

[0182] The calibration coefficients were determined through experiments using a blackbody radiation source.

[0183] Secondly, there is environmental temperature compensation, which corrects the temperature field based on real-time temperature and humidity data from the operating room.

[0184]

[0185] The current ambient temperature. This is the compensation coefficient.

[0186] S502: Spatial gradient calculation. This step in the invention is mainly for calculating the blood flow direction and velocity. Specifically, the temperature gradient calculation uses the Sobel operator to calculate the spatial gradient magnitude and direction:

[0187]

[0188] The gradients are in the horizontal and vertical directions.

[0189] Blood flow velocity estimation is based on the derivation of the flow velocity using a thermal convection-diffusion model:

[0190]

[0191] It is the thermal conductivity of the tissue; It is the specific heat capacity of the tissue; It is the baseline temperature difference (the temperature difference of healthy tissue measured before surgery).

[0192] S503: Time-domain hemodynamic analysis. This step in the invention includes time derivative calculation and blood flow pulse detection. Time derivative calculation involves performing a sliding window time differentiation on a continuous series of temperature fields:

[0193]

[0194] This is the frame interval time.

[0195] Blood flow pulse detection uses wavelet transform to extract the heart rhythm (0.5-3Hz frequency band):

[0196]

[0197] Pulse intensity The time marker is used to indicate a valid blood flow signal.

[0198] S504: Vascular network segmentation and density calculation. This step in the present invention mainly includes vascular structure extraction, vascular density quantification, and branch complexity analysis.

[0199] The extraction of vascular structures uses U-Net segmentation to extract vascular regions from the enhanced image:

[0200]

[0201] Blood vessel density quantization calculates the percentage of blood vessel pixels per unit area.

[0202]

[0203] Branch complexity analysis is based on calculating the fractal dimension using a skeletonized vascular map:

[0204]

[0205] Different scales The number of vascular skeleton nodes below.

[0206] S505: Abnormal Blood Flow Detection and Alarm. This step in the invention is to check for abnormal blood flow during surgery and to trigger an alarm when abnormal blood flow occurs. The main steps are as follows:

[0207] The first step is ischemic region identification, which requires defining dual threshold conditions:

[0208]

[0209] Secondly, there is the bleeding risk warning, which is based on temperature changes and sudden changes in blood vessel density. .

[0210] Then there is the multi-parameter fusion alarm, which uses fuzzy logic to comprehensively assess the risk level:

[0211]

[0212] when An audible and visual alarm will be triggered at any time.

[0213] S506: Real-time visualization and output. This step of the invention outputs a real-time blood flow map, including a blood flow velocity map: HSV color mapping (hue represents direction, saturation represents velocity). Ischemic region map: displayed with a red semi-transparent overlay.

[0214] Meanwhile, the digital dashboard displays key parameters in real time:

[0215]

[0216] The main output of step S5 in this invention is the blood flow velocity field. Blood vessel density It features ischemic area maps and real-time alarm signals. It also integrates a visual interface with DICOM archived data, supporting real-time intraoperative interaction.

[0217] Example

[0218] This invention, combined with the actual operation of a radical resection of rectal cancer (Bacon procedure), details the method for intraoperative intestinal blood supply monitoring using infrared thermography, specifically including the following steps:

[0219] Step 1: System Initialization and Preoperative Configuration. Install the infrared thermal imaging module and the visible light module on a dedicated imaging stand or robotic arm in the operating room, adjusting the angle to cover the entire intestinal operating area. Start the image acquisition system and perform calibration. The infrared module uses an uncooled focal plane array (640×512 resolution, 17μm pixel pitch, thermal sensitivity ≤30mK), and the visible light module is an ultra-high-definition white light industrial camera.

[0220] When initializing the image fusion system, the multi-source dynamic registration technique of this invention is used, employing a real-time non-rigid registration method. A lightweight deformation field estimation network (DFNet) is used to perform spatial registration correction between infrared and visible light images based on optical flow and thin-plate spline transformation (TPS). Deformation field From the formula Generate, where For reference frames (such as preoperative planning images). This is a real-time frame during surgery. The network outputs a control point displacement matrix, which is then interpolated using TPS to generate a dense deformation field, ensuring that pseudo-color layers can be overlaid within the same field of view during surgery. A micro-industrial computer loads software integrating the image enhancement algorithm of this invention. This algorithm incorporates innovative technologies such as multi-source dynamic registration, feature-level fusion, and a dual-branch collaborative network architecture, and is connected to AR glasses to provide intraoperative visual prompts.

[0221] Step Two: Intraoperative Image Acquisition and Real-time Processing After the surgery begins, the system simultaneously acquires long-wave infrared and visible light images, which are then processed in the multimodal image processing module. At this point, the multi-source dynamic registration and feature-level fusion technology of this invention is used to register images from different light sources before dynamic noise estimation and preprocessing. After registration, a dual-branch encoder is used to extract infrared features. and visible light characteristics ,Right now , Then, features are fused using an attention mechanism. ,in , For the Sigmoid function, This method uses adaptive weights. It adaptively selects complementary information from two modalities in a data-driven manner, providing a richer feature base for subsequent image enhancement.

[0222] Then, a dual-branch collaborative network architecture (DDPNet+CIMNet) is used to further process the fused image. The deep denoising prior network (DDPNet) adopts an improved U-Net structure, embedding detail-aware units (DAUs) in the encoder stage. It separates noise from tissue structure features through residual learning, dynamically extracting local detail residuals, as shown in the formula. ,in This represents the output value of the center pixel of the convolution kernel. Combined with an adversarial training strategy, the discriminator outputs an adversarial loss to constrain the realism of the generated image.

[0223] The contrast enhancement module (CIMNet) decomposes the image into low-frequency contrast components through multi-scale decoupling and Gaussian pyramid decomposition. ) and high-frequency detail components ( ). No. The formula for generating low-frequency components in the layer is as follows: ( ),in The number of decomposition layers, For the first Low-frequency components of the layer, Let be a Gaussian kernel function with a standard deviation of . The core weight is , For downsampling operators (usually intermittent sampling), initial conditions . No. Layer high-frequency detail residual calculation is as follows ,in This is an upsampling operator (bilinear interpolation). To address the corresponding detail residuals, the contrast branch employs global histogram stretching and local adaptive gamma correction to enhance the vascular structure, as shown in the formula: ,in It's the Clip function. It employs local adaptive gamma correction; the detail branch incorporates detail maps from DDPNet for feature stitching, and expands the receptive field through dilated convolution. This series of processes enhances the infrared image in real time, clearly highlighting the blood vessel contours and temperature difference boundaries.

[0224] When entering the abdominal cavity and freeing the inferior mesenteric artery and vein, the enhanced infrared image is compared with the background temperature to automatically depict the vessel contour and output a pseudo-color layer to the AR glasses. The system calculates the average temperature of the vessel area in real time and generates a curve. If a sudden temperature gradient occurs ( The system highlights the corresponding area to assist in locating the direction of blood vessels.

[0225] Step 3: Blood supply assessment before intestinal segment transection. The rectum is pulled out, and the surgeon plans to resect the segment 10cm above the tumor margin. The infrared system utilizes the innovative feature of a multi-scale joint loss function to focus on monitoring this area. The multi-scale joint loss function integrates adversarial loss (…). Multi-scale SSIM loss ), contrast loss ( and loss of detail () This approach achieves joint optimization of global structure and local details. The adversarial loss is based on the Wasserstein GAN optimization strategy, with the discriminator output constraining the generated image distribution; the multi-scale SSIM loss measures image fidelity through multi-scale structural similarity; the contrast loss constrains the similarity between low-frequency components and the target image; and the detail loss forces high-frequency details to align with the target residual.

[0226] Based on this, the system collects temperatures at the terminal, sub-terminal, and near-terminal points of the region, compares their thermal distribution maps, and automatically determines the temperature gradient. If the temperature in the terminal region is lower than that in the sub-terminal region... The system displayed a "insufficient blood supply" message and suggested adjusting the transection position. Activating 2x electronic zoom further enhanced the thermal texture details of the intestinal surface using image enhancement technology, assisting the surgeon in confirming the final transection location.

[0227] Step Four: Vessel Dissection and Temperature Response Assessment. After freeing the mesenteric vessels, the dissection operation is performed. The system uses real-time dynamic inference optimization technology to record images before and after dissection and generates a temperature response map. This invention adopts a lightweight design, performing channel pruning on DDPNet and CIMNet to retain key feature channels. The formula is as follows: ,in The sum of the absolute values ​​of the convolution kernel weights, retaining the first... This is a crucial channel. Simultaneously, hardware acceleration is implemented, utilizing CUDA cores for parallel computing to map convolution operations to GPU thread blocks, enabling rapid processing of each frame and thus allowing for fast and accurate image recording and analysis.

[0228] Within 20 seconds of disconnection, the color of the corresponding area of ​​the blood vessel on the temperature distribution map darkens significantly. The system automatically calculates the cooling rate and indicates "Blood flow interruption successful." If no significant cooling occurs, the system indicates "Suspected residual blood flow," prompting the operator to take further action.

[0229] Step 5: Post-intestinal segment suture blood supply monitoring. After the intestinal segment is severed, the proximal colon is pulled out of the anus and fixed, and intestinal suture is performed. The system acquires continuous infrared data of the suture area for 30 seconds and plots the temperature change trend curve of the anastomosis. A multi-scale joint loss function is used to ensure the accuracy and stability of the images. If the temperature rise in the anastomosis area lags behind... seconds, or temperature rise The system prompts "risk of blood supply to the anastomosis area," allowing the surgeon to decide whether to re-suture or dilate blood vessels. An infrared pseudo-color layer is superimposed on the visible light image for the surgeon to review the blood supply status of the anastomosis during the operation. This process is also based on multi-source dynamic registration and feature-level fusion technology to ensure accurate superposition and information complementarity between the two images.

[0230] Step Six: Postoperative Data Storage and Teaching Reuse. The system automatically saves keyframe images (comparison before and after severance, thermal images of the sutured area, and temperature curves) to the local database of the micro-industrial control computer. Medical staff can export surgical data for postoperative analysis and discussion; they can also play infrared and visible light fused images along a timeline for medical student teaching or expert review. After desensitization processing, the saved image data can also be used as a source of training samples for further algorithm optimization, continuously improving the image enhancement and intelligent analysis modules.

[0231] This embodiment uses the method provided in this application to achieve real-time visualization, quantitative judgment, and intelligent prompting of intestinal blood flow status during radical resection of rectal cancer through infrared thermal imaging, real-time image enhancement, multimodal fusion, and intelligent interpretation. This significantly improves the surgeon's accuracy in judging intestinal blood supply, effectively reduces the risk of postoperative complications such as anastomotic necrosis and intestinal necrosis, and also has good teaching and reuse value.

[0232] It should be emphasized that the embodiments described in this specification are merely illustrative of the technical solutions of the present invention, and the specific parameter settings, process details, and structural constructions involved do not constitute a limitation on the scope of patent protection. Those skilled in the art, based on an accurate understanding of the technical features explicitly defined in the claims, and in conjunction with prior art knowledge generally known in the art and the technical content disclosed in this specification, have the right to reasonably optimize and improve the implementation methods of the present invention. These optimizations and improvements include, but are not limited to, equivalent substitution of technical elements, adjustment of the order of process steps, and organic combination and expansion of technical features. Any innovative technical solution based on the core technical concept of the present invention, as long as its achieved technical effect has the same effect as the necessary technical features recorded in the claims, shall be deemed to fall within the legal protection scope defined by the claims of this patent.

Claims

1. A method of real-time rectal vascular blood flow detection with infrared thermographic enhancement, characterized in that, Comprising the following steps: Step S1: acquire the thermal image of the surgical area by an infrared thermal imager as a medical image to be processed, and register and fuse different light source images before dynamic noise estimation and preprocessing, adopt real-time non-rigid registration, use a lightweight deformation field estimation network DFNet, register based on optical flow method and thin plate spline transformation, after registration, use a double-branch encoder to extract features from infrared IR and visible light RGB images to obtain infrared features and visible light features , control the weight α of the obtained infrared features and visible light features by an attention mechanism to generate a spatial attention weight map; Step S2: using a double-branch collaborative network architecture, the deep denoising prior network DDPNet adopts an improved U-Net structure, introduces a detail perception unit DAU, and separates noise and organizational structure features through residual learning; in the encoder stage, the DAU module is embedded to dynamically extract local detail residuals; the contrast enhancement module CIMNet uses multi-scale decoupling and a Gaussian pyramid decomposition to decompose the image into a low-frequency contrast component C and a high-frequency detail component D; after decomposition, the contrast branch and the detail branch of the low-frequency contrast component and the high-frequency detail component are enhanced respectively; Step S3: obtaining the supervision signal in the image and processing each supervision model to obtain a processed image.

2. The method of claim 1, wherein the method is performed in real time. Further comprising the following steps: Step S4: channel pruning is performed on DDPNet and CIMNet to retain key feature channels P to realize real-time dynamic inference optimization; CUDA core parallel computing is used to map convolution operations to GPU thread blocks to realize fast processing of each frame.

3. The method of claim 1, wherein the method is performed in real time. Step S1 includes the following steps: Step S101: Take the infrared image and the visible light image as the image to be processed, the RGB visible light image captured synchronously , for anatomical structure reference, and the infrared image is normalized to [0, 1], the RGB image is converted to YUV color space, and the luminance channel is extracted for registration; Step S102: preoperative planning stage static infrared image Or intraoperative first frame As a reference frame, using deformation field estimation network (DFNet), input reference frame And the splicing image of the current frame Output control point displacement matrix The network structure used is lightweight U-Net, containing 4 down-sampling layers and skip connection, and the last layer is a fully connected layer for prediction Thin plate spline (TPS) interpolation method is used to generate global smooth deformation field according to the control point displacement wherein, is a radial basis function, and coefficients are solved by least square method, x is the x-axis coordinate point in the image, y is the y-axis coordinate point in the image, and the current frame is resampled according to the following formula by applying the deformation field when the images are aligned wherein, In the context of image warping, for functions and images the warped image is obtained by transforming (resampling) each pixel in the image according to the deformation field, Optimization is performed by using bidirectional optical flow consistency loss to constrain the smoothness of the deformation field: wherein, represents a morphing field transformation from the first frame to the second frame, and a morphing field transformation from the second frame to the third frame, the optimization constraining the smoothness of the morphing fields by computing a difference to the identity transformation. Step S103: registration is performed by using a multi-modal space alignment method; Step S104: Sample dual-branch encoder extracts features, where the infrared branch uses convolution (Conv+ReLU) to extract blood flow features , the pre-trained ResNet-18 removes the fully connected layer to extract anatomical features , Second, cross-modal attention fusion, calculate channel attention weight : wherein, is a sigmoid function; Then weighted feature fusion is performed: wherein, is a pixel-wise multiplication Hadamard product, dynamically adjusts the contribution proportion of the two modalities, realizes feature-level information complementation, and finally transposes the convolution to up-sample the fusion features to the original resolution to output a fusion image .

4. The method of claim 3, wherein the method is performed in real time. Step S103 specifically includes: First, cross-modal registration is performed, and the DFNet with shared weights is used to align the infrared and visible light images: And the RGB image is resampled: Second, the resolution is adapted, and the RGB image is bicubic.

5. The method of claim 1, wherein the method is performed in real time. In step S2, global histogram stretching and local adaptive gamma correction operations are used to enhance the contrast of the blood vessel structure in the component image in the contrast branch enhancement: The detail branch introduces a detail map of DDPNet for feature concatenation, and the receptive field is expanded through a dilated convolution.

6. The method of claim 1, wherein the method is performed in real time. Step S2 includes the following steps: Step S201: Normalization processing is performed on the image to be processed, and pixel values of the fused image are linearly mapped to the range [0, 1]: Step S202: The normalized image is subjected to a logarithmic transformation wherein, a minimum value of the pixels of the image, a maximum value of the pixels of the image, is a normalized image, respectively normalizing the infrared and visible light channel information of the image to be processed; Second, spatial smoothing filtering is performed, and a Gaussian filter is used to suppress high-frequency noise: S202: dynamic noise level estimation, calculating the local variance to traverse the image with a sliding window to calculate the noise variance of each window: wherein, is the mean value within the window; Second, the modified motion compensation utilizes the Jacobian determinant of the registration deformation field to suppress motion region noise overestimation:​ wherein, represents the local noise variance of the image at the pixel position, represents a local compression region, where the noise variance needs to be attenuated according to the degree of deformation; Then, the global noise is modeled, a noise variance-mean relationship curve is fitted, and a noise level map is established : where the coefficients By least square method, all windows of Data pairs; S203: performing time domain weighted filtering and space domain guided filtering on the obtained image, wherein the time domain weighted filtering combines the noise image of the previous several frames to calculate the noise level after time domain smoothing : wherein, is the first frame noise map; is the first frame noise map; is the first frame noise map; spatially guided filtering to is the guide image, and edge preservation is maintained: S204: the preprocessed image and the noise map are spliced to form the input tensor of DDPNet: Among them, the first channel is the smoothed image, and the second channel is the noise level map; S205: Sample outlier noise detection Mark the region whose noise level exceeds 3 times the standard deviation of the global mean, if Time, 1, the rest is 0, where, is The global mean, is The global standard deviation, the fast non-local mean filtering of the marked region is carried out by non-local mean correction: Constructing input tensor , containing a smoothed image and a noise level map for DDPNet to perform denoising and detail extraction.

7. The method of claim 1, wherein the method is performed in real time. Step S3 includes the following steps: S301: use the pre-processed level map with noise as input ; Second, four-level down-sampling coding is performed: Level 1 is original resolution sampling coding, the specific operation is first through 3x3 convolution, secondly using ReLU activation, then through detail-aware unit (DAU) module to extract local detail residual, suppress noise interference, and finally maximum pooling is performed, and the feature map is output: Levels 2 to 4 are progressively down-sampling, repeating the following operations: Output feature map scale: , , , The mathematical expression of the detail perception unit used is: wherein, Extract the 3x3 kernel center pixel value (i.e. ignore edge response); S302: First, level 4 to level 3, the input is , transpose convolution, concatenation with the features of level 3, through the convolution layer, using ReLU activation, output ; Second, level 3 to level 2 input concatenate with level 2 features through a convolutional layer, using ReLU activation output ; Then, level 2 to level 1 input , spliced output with level 1 features , final up-sampling, transposed convolution to restore to original resolution, output preliminary denoised image ; Finally, the residual connection reinforces the details, with the details residual output by each layer of DAU Weighted fusion: Correcting the denoising result: in, It is the weight of the detailed residual obtained by weighted fusion. Indicates the first The basic denoising results of the frame, This represents a detailed residual plot obtained based on noise estimation and guidance information. S303: use the discriminator network to optimize the adversarial training, the resulting structure is 5 layers of convolution→fully connected layer→sigmoid output probability, the input is the spliced denoising image with noise level map ; The adversarial loss is calculated in this step, and the WassersteinGAN-GP is used to calculate the loss: wherein, the average score of the output, refers to the discriminator for real samples the average score of the output, refers to the discriminator for denoised images the average score of the output, refers to the discriminator for random interpolated samples the average of the square of the difference between the modulus (L2 norm) of the gradient of the output and 1; S304: multiple losses are used for multi-level supervision and loss calculation, specifically as follows:

1. Pixel-level L1 loss, the mathematical expression is: wherein, denotes the denoised image of the t-th frame, 2. Multi-scale SSIM loss, the mathematical expression is: wherein, denotes the down-sampling scale, and denotes the denoised image and the sharp reference image the local mean at the th scale, and the local standard deviation at the th scale, and depend on the size of the image.

3. Detail preserving loss on detail residual Applying sparsity constraints:

4. The total loss function is: wherein, represents an adversarial loss, represents a pixel-level loss, represents a multi-scale structural similarity loss, represents a detail residual enhancement loss; Output: denoised image and detail residual map for use by the CIMNet enhancement.

8. The method of claim 2, wherein the method is performed in real time. Step S4 includes the following steps: S401: input data: denoised image , and detail residual map , denoised image from DDPNet output, detail residual map detail features extracted from DDPNet; After inputting the data, first, feature concatenation is performed to concatenate the detail residuals and the denoised image into an input tensor: ； S402: a Gaussian pyramid is constructed, and the input image is 3-layer Gaussian down-sampled to generate multi-scale features: wherein, is the original scale, is the original input image (i.e. ) is Gaussian blurred by a Gaussian kernel smoothing and down-sampled, is the output of the previous layer is further Gaussian blurred and down-sampled, i.e. the Gaussian pyramid operation is repeated on the output of the previous layer; Second, the Laplacian residual is calculated by upsampling the difference as shown below Extracting high-frequency details of each layer: wherein, represents a feature map representation at the th scale, represents up-sampling the th coarser feature map to the resolution of the th scale, outputting a residual wherein, is the finest scale detail; S403: contrast enhancement is performed by low frequency component extraction, adaptive histogram stretching and local gamma correction, the low frequency component extraction is to each layer Gaussian image Low-pass filtering is performed to obtain a low frequency component: The adaptive histogram stretching dynamically adjusts the contrast range of each layer of low-frequency components: wherein, : mean and standard deviation of the low frequency components of the current layer; : learnable parameters, predicted by 1x1 convolution, is a very small constant for numerical stability; Local gamma correction adjusts gamma values according to spatial luminance distribution :​ wherein, Predicted by small CNN, input is current layer feature, is the first layer enhanced feature map after normalization and learnable scale adjustment; is S404: high frequency enhancement is performed through detail feature fusion, hollow convolution enhancement and detail sharpening, the detail feature fusion pulls the Laplacian residual and the detail residual of DDPNet to obtain : Where k represents the multi-scale level index; The cavity convolution enhancement adopts multi-scale cavity convolution (dilation=1, 2, 3) to extract multi-receptive field details: Wherein, d represents the cavity rate of the convolution kernel; The cavity convolution kernel weight is adaptively fused through the attention mechanism; The detail sharpening strengthens the edge through high-pass filtering and residual connection: wherein, denotes the feature map of the kth layer after the expansion convolution enhancement, and a is a sharpening intensity coefficient, which is dynamically predicted by the network. S405: First, cross-scale feature alignment is performed, and bilinear up-sampling is performed on each layer of the enhanced low frequency ( ) and high frequency ( ) to the original resolution; Secondly, the weighted channel attention calculates the attention weight of each scale feature: wherein, is the first layer enhanced feature map, is the first layer enhanced feature map after dynamic sharpening; GAP is global average pooling, output weights ; Then, the output weighted fusion is as follows: S406: residual connection and output by residual reinforcement and dynamic range compression, residual reinforcement is to add the fusion result to the original input, and the basic structure is retained : wherein, is the first frame after the basic denoising processing of the image; The dynamic range compression is applied to the learnable S-shaped curve to prevent over-enhancement: wherein the parameters Predicted from global features by a fully connected layer; output is the final augmented image and the scale-specific attention weight maps for visualizing the analysis.

9. The method of claim 2, wherein the method is performed in real time. Further comprising step S5: based on the enhanced image Computing spatial gradient , wherein is the enhanced infrared thermal image, x is the horizontal coordinate of the image, y is the vertical spatial coordinate of the image, the proportional coefficient is determined through calibration experiment , calculating blood flow velocity , wherein is the temperature distribution value corresponding to the image, the threshold value is set , if the area is lower than the threshold value for several continuous frames, triggering the audible and light alarm to alarm and reacquiring the image to be processed.

10. The method of claim 9, wherein the method is performed in real time. Step S5 includes the following steps: Step S501: First, construct a pixel-temperature mapping, establish a linear relationship between the gray value and the actual temperature through preoperative calibration experiments : wherein is a calibration coefficient determined by a black body radiation source experiment; Secondly, the environmental temperature compensation is carried out, and the temperature field is corrected according to real-time temperature and humidity data of the operating room : wherein, is the current ambient temperature, is a compensation factor, is a reference ambient temperature; Step S502: Sample space gradient to calculate blood flow direction and speed in the image; First, the spatial gradient magnitude is calculated using the Sobel operator with the direction : wherein are the horizontal and vertical directional gradients; Based on the heat convection-diffusion model, the blood flow velocity is derived: wherein, is the tissue thermal conductivity; is the tissue specific heat capacity; is the baseline temperature difference, the baseline temperature difference being the temperature difference between the preoperative measurement of the healthy tissue flavor and the temperature of the image source assembly from which the temperature is to be measured; Step S503: Time derivative calculation and blood flow pulse detection are performed for time-domain hemodynamic analysis; The time derivative calculation is a sliding window time differentiation on a plurality of continuous frames of temperature fields: wherein, is the frame interval time; is the frame interval time; frame, pixel position corresponding enhanced temperature value; is the frame interval time; frame, pixel position corresponding enhanced temperature value; Applying wavelet transform to extract heartbeat rhythm pulse signal in 0.5-3 Hz band Carrying out blood flow pulse detection: wherein f is the wavelet transform amplitude signal; to enhance the temperature sequence; is a complex exponential wavelet, suitable for detecting pulse signals in the frequency range (0.5-3 Hz); Pulse intensity Time stamps are marked as valid blood flow signals, otherwise as invalid blood flow signals; Step S504: Vessel network segmentation and density calculation are performed by vessel structure extraction, vessel density quantification, and branch complexity analysis; U-Net segmentation is used to enhance the blood vessel region in the image to perform blood vessel structure extraction: Wherein, H is the height of the mask image; W is the width of the mask image; The blood vessel density is obtained by calculating the proportion of blood vessel pixels in a unit area : Branching complexity analysis is based on calculating fractal dimension of skeletonized vessel map : wherein, the number of vessel skeleton nodes at different scales is different. Step S505: Abnormal blood flow detection and alarm, check the abnormal blood flow according to the obtained image during the operation, and alarm when the abnormal blood flow occurs, including the following steps: First, the ischemic region is identified and the following double threshold condition is defined: wherein, for the pixel position corresponding blood flow velocity; for the pixel position corresponding spatial temperature gradient magnitude; Secondly, based on the temperature change and the blood vessel density mutation, the bleeding risk is evaluated, and a warning is given according to the evaluation result and the blood vessel density mutation calculation: ; wherein t is the observation time; T is the temperature; when the temperature change and the blood vessel density does not meet the above conditions, a warning alarm is given, and if the above result is met, no warning alarm is given; Then, multi-parameter fusion alarm, using fuzzy logic comprehensive evaluation risk level : wherein S is the ischemic region and K is the quantitative blood flow index map; when the risk level is assessed a sound and light alarm is triggered. Step S506: Real-time output of blood flow velocity map, HSV color mapping information and ischemic area map; In the HSV color mapping information, the direction is represented by the hue, and the speed is represented by the saturation; In the ischemic area map, the ischemic area is displayed in red semi-transparency superposition; Digital instrument panel displays key parameters in real time : wherein, is a real-time detection meter; is an ischemic region score; By outputting blood flow velocity fields , vascular density , ischemic area map, real-time alarm signal, integrated visualization interface and DICOM archival data, support for intraoperative real-time interaction.

Images