A label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics

Abstract

This invention discloses a label-free real-time blood vessel tracking method driven by dynamic blood oxygenation features, relating to the fields of medical image processing and computer vision. The method includes: acquiring dynamic blood oxygenation signals from a target tissue region; preprocessing the blood oxygen concentration change data; constructing a spatiotemporal matrix of blood oxygenation features; performing adaptive feature extraction using an improved 3D convolutional neural network; establishing a multi-scale blood vessel contour prediction model; dynamically predicting and updating the blood vessel contour using a Kalman filter algorithm; outputting the real-time trajectory of the blood vessel center and boundary contours; and introducing a dynamic feedback mechanism to achieve continuous and stable blood vessel tracking. This invention achieves highly secure blood vessel tracking, significantly improving tracking accuracy, robustness, and real-time performance under tissue movement and complex background interference by utilizing dynamic blood oxygenation features.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and computer vision technology, and more specifically to a label-free real-time blood vessel tracking method driven by dynamic blood oxygenation features. Background Technology

[0002] Precise and real-time location and tracking of blood vessels are crucial in clinical diagnosis (such as vascular lesion examination), surgical navigation (such as minimally invasive surgery), and biomedical research.

[0003] Existing vascular imaging and tracking technologies primarily rely on external markers (such as fluorescent contrast agents) or morphology-based image processing techniques. However, contrast agent-based methods are invasive, may cause allergic reactions, and cannot be used continuously for extended periods, resulting in low safety. Traditional morphology-based image processing techniques (such as grayscale and texture) lack robustness and are susceptible to interference from tissue movement, lighting changes, and tissue coverage, leading to decreased tracking accuracy or even target loss in complex scenarios. Furthermore, existing methods often struggle to effectively distinguish between arteries and veins, and their processing speed is insufficient to meet the demands of high real-time surgical navigation.

[0004] Therefore, proposing a label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics to solve the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a label-free real-time vascular tracking method driven by dynamic blood oxygenation features, which achieves highly secure vascular tracking. By utilizing dynamic blood oxygenation features, it significantly improves tracking accuracy, robustness, and real-time performance under tissue motion and complex background interference.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics includes the following steps: S1. Use near-infrared spectroscopy imaging equipment to collect dynamic blood oxygenation signals of the target tissue area and obtain time-series data on changes in blood oxygenation concentration, including dynamic changes in oxyhemoglobin concentration and deoxyhemoglobin concentration. S2. Preprocess the blood oxygen concentration change data to remove motion artifacts and environmental noise, and obtain standardized blood oxygen characteristic data. S3. Construct a spatiotemporal matrix of blood oxygen features. Reorganize the standardized blood oxygen feature data according to the time and space dimensions, and extract the blood oxygen fluctuation features of each spatial pixel in the time series. The blood oxygen fluctuation features include fluctuation frequency, fluctuation amplitude and phase difference. S4. Based on the spatiotemporal matrix of blood oxygen features, adaptive feature extraction is performed through an improved three-dimensional convolutional neural network to select a subset of blood oxygen features with vascular specificity. The vascular specific features meet the conditions that the fluctuation frequency matches the human hemodynamic cycle and the amplitude is higher than the preset threshold of the surrounding non-vascular tissue. S5. Establish a multi-scale vascular contour prediction model. Using a subset of vascular-specific features as input, combine the Kalman filter algorithm to dynamically predict and update the vascular contour, and output the real-time vascular center trajectory and boundary contour. S6. A dynamic feedback mechanism is introduced to adjust the feature extraction weights and model parameters based on the blood vessel tracking deviation between adjacent frames, thereby achieving continuous and stable blood vessel tracking.

[0007] Optionally, the near-infrared spectral imaging device in S1 can be a laser speckle contrast imaging device or a spatial frequency domain imaging device.

[0008] Optionally, in S2, adaptive filtering, baseline correction, and spatiotemporal registration preprocessing are performed on the blood oxygen concentration change data to remove motion artifacts and environmental noise, resulting in standardized blood oxygen feature data. Spatiotemporal registration employs image registration algorithms based on phase correlation or optical flow to compensate for image shifts caused by tissue motion.

[0009] Optionally, the specific content of S3 is as follows: The blood oxygen concentration values ​​of each pixel in a sequence of images with a spatial resolution of M×N pixels collected over a period of time T are reconstructed into a three-dimensional data structure with dimensions (M,N,T), where the first two dimensions (M,N) represent spatial location and the third dimension (T) represents time series. A fast Fourier transform is performed on the time-series blood oxygenation signal of each pixel to convert the signal from the time domain to the frequency domain, and the frequency component with the highest energy in the power spectrum is determined as the fluctuation frequency of the pixel. After extracting the fluctuation frequency, the fluctuation frequency is compared with a preset frequency range, which corresponds to the heart rate range of a human at rest or in a specific state; if the fluctuation frequency is within the preset frequency range, the pixel features are retained. Bandpass filtering is applied to the time-series blood oxygenation signal of each pixel to retain the frequency band components related to heart rate; the difference between the peak and trough values ​​of the filtered signal in one or more complete cycles is calculated as the quantized value of the fluctuation amplitude. A reference signal is selected, and Hilbert transforms are performed on the reference signal and the signal of each pixel to be measured to obtain the instantaneous phase; the difference between the instantaneous phase of each pixel to be measured and the instantaneous phase of the reference signal is calculated as the phase difference; The three feature values ​​of each pixel—fluctuation frequency, fluctuation amplitude, and phase difference—are fused into a multidimensional feature vector.

[0010] Optionally, the specific content of S4 is as follows: A three-dimensional convolutional neural network, in which the convolutional kernel slides across two spatial dimensions and one temporal dimension to simultaneously extract the spatiotemporal correlation features of blood oxygenation signals; The improved 3D convolutional neural network incorporates an attention mechanism module; the attention mechanism module is configured to automatically generate a weight map by analyzing input features to enhance attention to the channels and spatial regions of amplitude and frequency information in blood oxygen fluctuation features, and suppress the response of noise and non-vascular tissue regions. The vascular-specific characteristics meet the conditions that the fluctuation frequency matches the human hemodynamic cycle and the amplitude is higher than the preset threshold of the surrounding non-vascular tissue.

[0011] Optionally, the attention mechanism module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule is used to learn and assign greater weights to feature channels that are highly correlated with frequency and amplitude features. The spatial attention submodule is used to highlight the spatial location of blood vessels on the feature map.

[0012] Optionally, the condition for matching the fluctuation frequency with the human hemodynamic cycle is: introduce a frequency constraint term into the network's loss function, and the frequency constraint term penalizes the prediction results of pixels whose dominant frequency components deviate from the preset physiological heart rate range.

[0013] Optionally, the condition for the amplitude to be higher than the preset threshold of the surrounding non-vascular tissue is as follows: In the last layer of the improved three-dimensional convolutional neural network, post-processing logic is used to compare the predicted amplitude of each pixel with the average amplitude of the non-vascular background region in the neighborhood, and the pixel is determined to be a vascular point only when the amplitude of the pixel is significantly higher than the background amplitude.

[0014] Optionally, the training process of the improved convolutional neural network adopts a multi-task learning strategy, and the output includes at least: a binary segmentation map for determining whether a pixel is a blood vessel; an amplitude map for regressing the blood oxygen fluctuation amplitude value of the pixel; and a frequency map for regressing the dominant fluctuation frequency of the pixel.

[0015] Optionally, the specific content of S5 is as follows: The established multi-scale blood vessel contour prediction model is a convolutional neural network with an encoder-decoder structure. The encoder extracts and compresses the features of the input feature map step by step through multi-layer convolution and pooling operations to obtain high-level features containing global context information. The decoder restores spatial details step by step through multi-layer upsampling and skip connection operations, and finally outputs a blood vessel contour probability map with the same size as the input image, where the value of each pixel represents the probability of belonging to a blood vessel.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention provides a label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics, which has the following beneficial effects: This invention achieves highly secure vascular tracking by utilizing dynamic blood oxygenation characteristics, significantly improving tracking accuracy, robustness, and real-time performance under tissue movement and complex background interference. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 The flowchart of a label-free real-time vascular tracking method driven by dynamic blood oxygenation features provided by the present invention is shown. Detailed Implementation

[0019] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 As shown, this invention discloses a label-free real-time vascular tracking method driven by dynamic blood oxygenation features, comprising the following steps: S1. Use near-infrared spectroscopy imaging equipment to collect dynamic blood oxygenation signals of the target tissue area and obtain time-series data on changes in blood oxygenation concentration, including dynamic changes in oxyhemoglobin concentration and deoxyhemoglobin concentration. S2. Preprocess the blood oxygen concentration change data to remove motion artifacts and environmental noise, and obtain standardized blood oxygen characteristic data. S3. Construct a spatiotemporal matrix of blood oxygen features. Reorganize the standardized blood oxygen feature data according to the time and space dimensions, and extract the blood oxygen fluctuation features of each spatial pixel in the time series. The blood oxygen fluctuation features include fluctuation frequency, fluctuation amplitude and phase difference. S4. Based on the spatiotemporal matrix of blood oxygen features, adaptive feature extraction is performed through an improved three-dimensional convolutional neural network to select a subset of blood oxygen features with vascular specificity. The vascular specific features meet the conditions that the fluctuation frequency matches the human hemodynamic cycle and the amplitude is higher than the preset threshold of the surrounding non-vascular tissue. S5. Establish a multi-scale vascular contour prediction model. Using a subset of vascular-specific features as input, combine the Kalman filter algorithm to dynamically predict and update the vascular contour, and output the real-time vascular center trajectory and boundary contour. S6. A dynamic feedback mechanism is introduced to adjust the feature extraction weights and model parameters based on the blood vessel tracking deviation between adjacent frames, thereby achieving continuous and stable blood vessel tracking.

[0021] Furthermore, the near-infrared spectral imaging device in S1 employs either a laser speckle contrast imaging device or a spatial frequency domain imaging device.

[0022] Furthermore, in S2, adaptive filtering, baseline correction, and spatiotemporal registration preprocessing are performed on the blood oxygen concentration change data to remove motion artifacts and environmental noise, resulting in standardized blood oxygen feature data. Spatiotemporal registration employs image registration algorithms based on phase correlation or optical flow to compensate for image shifts caused by tissue motion.

[0023] Furthermore, the specific content of S3 is as follows: The blood oxygen concentration values ​​of each pixel in a sequence of images with a spatial resolution of M×N pixels collected over a period of time T are reconstructed into a three-dimensional data structure with dimensions (M,N,T), where the first two dimensions (M,N) represent spatial location and the third dimension (T) represents time series. A fast Fourier transform is performed on the time-series blood oxygenation signal of each pixel to convert the signal from the time domain to the frequency domain, and the frequency component with the highest energy in the power spectrum is determined as the fluctuation frequency of the pixel. After extracting the fluctuation frequency, the fluctuation frequency is compared with a preset frequency range, which corresponds to the heart rate range of a human at rest or in a specific state; if the fluctuation frequency is within the preset frequency range, the pixel features are retained. Bandpass filtering is applied to the time-series blood oxygenation signal of each pixel to retain the frequency band components related to heart rate; the difference between the peak and trough values ​​of the filtered signal in one or more complete cycles is calculated as the quantized value of the fluctuation amplitude. A reference signal is selected, and Hilbert transforms are performed on the reference signal and the signal of each pixel to be measured to obtain the instantaneous phase; the difference between the instantaneous phase of each pixel to be measured and the instantaneous phase of the reference signal is calculated as the phase difference; The three feature values ​​of each pixel—fluctuation frequency, fluctuation amplitude, and phase difference—are fused into a multidimensional feature vector.

[0024] Specifically, the reference signal is selected from one of the following: Select the signal of a pixel located on the main trunk of a large blood vessel within the imaging field of view; The average value of the signal of all pixels within the entire imaging field of view; An externally acquired synchronous physiological signal, such as a photoplethysmography signal or an electrocardiogram signal.

[0025] Specifically, the blood vessel types are initially distinguished based on phase difference: pixels with a phase difference close to zero, that is, almost in phase with the reference signal, are initially identified as arterial blood vessel regions; pixels with a phase difference significantly greater than zero, that is, lagging behind the reference signal, are initially identified as venous blood vessel regions.

[0026] Furthermore, the specific content of S4 is as follows: A three-dimensional convolutional neural network, in which the convolutional kernel slides across two spatial dimensions and one temporal dimension to simultaneously extract the spatiotemporal correlation features of blood oxygenation signals; The improved 3D convolutional neural network incorporates an attention mechanism module; the attention mechanism module is configured to automatically generate a weight map by analyzing input features to enhance attention to the channels and spatial regions of amplitude and frequency information in blood oxygen fluctuation features, and suppress the response of noise and non-vascular tissue regions. The vascular-specific characteristics meet the conditions that the fluctuation frequency matches the human hemodynamic cycle and the amplitude is higher than the preset threshold of the surrounding non-vascular tissue.

[0027] Specifically, the blood oxygenation feature subset with blood vessel specificity is the set of pixels predicted as blood vessels in the binary segmentation map. The determination of the subset combines regression results from amplitude map and frequency map.

[0028] Furthermore, the attention mechanism module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule is used to learn and assign greater weights to feature channels that are highly correlated with frequency and amplitude features. The spatial attention submodule is used to highlight the spatial location of blood vessels on the feature map.

[0029] Furthermore, the condition for matching the fluctuation frequency with the human hemodynamic cycle is as follows: a frequency constraint term is introduced into the network's loss function, and the frequency constraint term penalizes the prediction results of pixels whose dominant frequency components deviate from the preset physiological heart rate range.

[0030] Specifically, the preset physiological heart rate range is 0.5Hz-3.0Hz, but it can also be set according to specific needs.

[0031] Furthermore, the condition for the amplitude to be higher than the preset threshold of the surrounding non-vascular tissue is as follows: in the last layer of the improved three-dimensional convolutional neural network, post-processing logic is used to compare the predicted amplitude of each pixel with the average amplitude of the non-vascular background region in the neighborhood. Only when the amplitude of the pixel is significantly higher than the background amplitude is it determined to be a vascular point.

[0032] Specifically, the range in which the amplitude of a pixel is significantly higher than the amplitude of the background is more than 1.5 to 2 times the standard deviation, and can also be set according to specific needs.

[0033] Furthermore, the training process of the improved convolutional neural network adopts a multi-task learning strategy, and the output includes at least: a binary classification segmentation map for determining whether a pixel is a blood vessel; an amplitude map for regressing the blood oxygen fluctuation amplitude value of the pixel; and a frequency map for regressing the dominant fluctuation frequency of the pixel.

[0034] Furthermore, the specific details of S5 are as follows: The established multi-scale blood vessel contour prediction model is a convolutional neural network with an encoder-decoder structure. The encoder extracts and compresses the features of the input feature map step by step through multi-layer convolution and pooling operations to obtain high-level features containing global context information. The decoder restores spatial details step by step through multi-layer upsampling and skip connection operations, and finally outputs a blood vessel contour probability map with the same size as the input image, where the value of each pixel represents the probability of belonging to a blood vessel.

[0035] Specifically, skip connections concatenate the feature maps of the same scale in the encoder network with the upsampled feature maps in the decoder to fuse deep semantic information with shallow positional detail information, thereby improving the prediction accuracy of blood vessel boundary contours. Specifically, dynamic prediction and updating using the Kalman filter algorithm includes the following sub-steps: After obtaining the blood vessel contour of the initial frame, the coordinates of the key points or control points of the contour on the blood vessel centerline are defined as the state variables of the Kalman filter and initialized. When a new frame arrives, based on a preset motion model (uniform speed model or constant acceleration model), the Kalman filter is used to predict the predicted position of the blood vessel contour points in the current frame according to the state of the previous frame. The contour observation results directly generated by the multi-scale vascular contour prediction model based on the blood oxygen feature subset in the current frame are fused with the predicted location to obtain the optimal estimated state of the vascular contour points in the current frame, and the parameters of the Kalman filter are updated.

[0036] The data fusion is based on the prediction covariance estimated by the Kalman filter itself and the observation noise covariance of the multi-scale vascular contour prediction model. When the observation noise covariance is small, it means that the blood oxygen feature quality of the current frame is high and the observation is reliable, so the fusion result is more inclined to the observation value. When the observation noise covariance is large, it means that the current frame may be severely disturbed, so the fusion result is more inclined to the prediction value to maintain the stability of tracking.

[0037] Specifically, the real-time vascular center trajectory is obtained by processing the final determined vascular contour using a skeleton extraction algorithm; the boundary contour is obtained by thresholding the vascular contour probability map and then smoothing it using morphological operations. The process noise covariance and observation noise covariance of the Kalman filter are not fixed, but are adaptively adjusted according to the tracking deviation output by the dynamic feedback mechanism to achieve online optimization of filter performance.

[0038] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0039] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics, characterized in that, Includes the following steps: S1. Use near-infrared spectroscopy imaging equipment to collect dynamic blood oxygenation signals of the target tissue area and obtain time-series data on changes in blood oxygenation concentration, including dynamic changes in oxyhemoglobin concentration and deoxyhemoglobin concentration. S2. Preprocess the blood oxygen concentration change data to remove motion artifacts and environmental noise, and obtain standardized blood oxygen characteristic data. S3. Construct a spatiotemporal matrix of blood oxygen features. Reorganize the standardized blood oxygen feature data according to the time and space dimensions, and extract the blood oxygen fluctuation features of each spatial pixel in the time series. The blood oxygen fluctuation features include fluctuation frequency, fluctuation amplitude and phase difference. S4. Based on the spatiotemporal matrix of blood oxygen features, adaptive feature extraction is performed through an improved three-dimensional convolutional neural network to select a subset of blood oxygen features with vascular specificity. The vascular specific features meet the conditions that the fluctuation frequency matches the human hemodynamic cycle and the amplitude is higher than the preset threshold of the surrounding non-vascular tissue. S5. Establish a multi-scale vascular contour prediction model. Using a subset of vascular-specific features as input, combine the Kalman filter algorithm to dynamically predict and update the vascular contour, and output the real-time vascular center trajectory and boundary contour. S6. A dynamic feedback mechanism is introduced to adjust the feature extraction weights and model parameters based on the blood vessel tracking deviation between adjacent frames, thereby achieving continuous and stable blood vessel tracking.

2. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 1, characterized in that, The near-infrared spectral imaging equipment in S1 employs either laser speckle contrast imaging or spatial frequency domain imaging.

3. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 1, characterized in that, In S2, adaptive filtering, baseline correction, and spatiotemporal registration preprocessing are performed on the blood oxygen concentration change data to remove motion artifacts and environmental noise, resulting in standardized blood oxygen feature data. Spatiotemporal registration employs image registration algorithms based on phase correlation or optical flow to compensate for image shifts caused by tissue motion.

4. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 1, characterized in that, The specific content of S3 is as follows: The blood oxygen concentration values ​​of each pixel in a sequence of images with a spatial resolution of M×N pixels collected over a period of time T are reconstructed into a three-dimensional data structure with dimensions (M,N,T), where the first two dimensions (M,N) represent spatial location and the third dimension (T) represents time series. A fast Fourier transform is performed on the time-series blood oxygenation signal of each pixel to convert the signal from the time domain to the frequency domain, and the frequency component with the highest energy in the power spectrum is determined as the fluctuation frequency of the pixel. After extracting the fluctuation frequency, the fluctuation frequency is compared with a preset frequency range, which corresponds to the heart rate range of a person at rest or in a specific state. If the fluctuation frequency is within the preset frequency range, the pixel features are preserved; Bandpass filtering is performed on the time-series blood oxygenation signal of each pixel to retain the frequency band components related to heart rate; The difference between the peak and trough values ​​of the filtered signal within one or more complete cycles is calculated and used as the quantized value of the fluctuation amplitude. A reference signal is selected, and Hilbert transforms are performed on the reference signal and the signal of each pixel to be measured to obtain the instantaneous phase; the difference between the instantaneous phase of each pixel to be measured and the instantaneous phase of the reference signal is calculated as the phase difference; The three feature values ​​of each pixel—fluctuation frequency, fluctuation amplitude, and phase difference—are fused into a multidimensional feature vector.

5. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 1, characterized in that, The specific content of S4 is as follows: A three-dimensional convolutional neural network, in which the convolutional kernel slides across two spatial dimensions and one temporal dimension to simultaneously extract the spatiotemporal correlation features of blood oxygenation signals; The improved 3D convolutional neural network incorporates an attention mechanism module; The attention mechanism module is configured to automatically generate a weight map by analyzing input features to enhance attention to the channels and spatial regions of amplitude and frequency information in blood oxygen fluctuation features, and to suppress the response of noise and non-vascular tissue regions. The vascular-specific characteristics meet the conditions that the fluctuation frequency matches the human hemodynamic cycle and the amplitude is higher than the preset threshold of the surrounding non-vascular tissue.

6. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 5, characterized in that, The attention mechanism module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule is used to learn and assign greater weights to feature channels that are highly correlated with frequency and amplitude features. The spatial attention submodule is used to highlight the spatial location of blood vessels on the feature map.

7. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 5, characterized in that, The condition for matching the fluctuation frequency with the human hemodynamic cycle is: introduce a frequency constraint term into the network's loss function, and the frequency constraint term penalizes the prediction results of pixels whose dominant frequency components deviate from the preset physiological heart rate range.

8. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 5, characterized in that, The condition for an amplitude higher than the preset threshold of the surrounding non-vascular tissue is as follows: In the last layer of the improved three-dimensional convolutional neural network, post-processing logic is used to compare the predicted amplitude of each pixel with the average amplitude of the non-vascular background region in the neighborhood. Only when the amplitude of a pixel is significantly higher than the background amplitude is it determined to be a vascular point.

9. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 1, characterized in that, The training process of the improved convolutional neural network adopts a multi-task learning strategy, and the output includes at least: a binary classification segmentation map for determining whether a pixel is a blood vessel; an amplitude map for regressing the blood oxygen fluctuation amplitude value of the pixel; and a frequency map for regressing the dominant fluctuation frequency of the pixel.

10. The label-free real-time vascular tracking method driven by dynamic blood oxygenation characteristics according to claim 1, characterized in that, The specific content of S5 is as follows: The established multi-scale vascular contour prediction model is a convolutional neural network with an encoder-decoder structure. The encoder extracts and compresses the features of the input feature map step by step through multi-layer convolution and pooling operations to obtain high-level features containing global context information. The decoder gradually recovers spatial details through multi-layer upsampling and skip connection operations, and finally outputs a blood vessel contour probability map with the same size as the input image, where the value of each pixel represents the probability of belonging to a blood vessel.

Images