Real-time processing method for endoscopic image blood vessel enhancement and microcirculation evaluation

Abstract

The invention provides a real-time processing method for endoscopic image blood vessel enhancement and microcirculation evaluation, which comprises the following steps of: processing an original endoscopic image, and locally enhancing a brightness channel through a contrast-limited adaptive histogram equalization technology; self-adaptive frequency domain-space domain decomposition is realized based on local texture complexity analysis; performing multi-scale Hessian matrix blood vessel detection on the low-frequency component; applying a directional Gabor filter bank to the high-frequency component; spatial-temporal feature fusion is realized through multi-resolution pyramid optical flow calculation; synchronously completing blood vessel probability prediction, blood vessel diameter estimation and blood flow direction prediction by using a lightweight multi-task deep learning network; the blood flow velocity is analyzed and calculated based on a speckle mode, the perfusion density is subjected to accelerated statistics through an integrogram, and vascular morphological parameters are extracted by adopting an improved skeleton algorithm. According to the invention, an enhanced blood vessel visualization effect and a real-time microcirculation quantitative evaluation function can be provided, and the overall improvement of the endoscope image processing quality and efficiency is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a real-time processing method for endoscopic image blood vessel enhancement and microcirculation evaluation, which can be widely applied to various endoscopic examination systems such as digestive tract endoscopy, bronchoscopy, laparoscopy, arthroscopy, etc., to provide enhanced blood vessel visualization and real-time microcirculation quantitative evaluation functions for clinicians, and is mainly applied to early lesion screening, surgical navigation, tissue activity evaluation, etc. clinical scenarios. BACKGROUND

[0002] Endoscopy technology, as an important means of modern medicine, has been widely used in digestive tract, respiratory tract, urinary system and other clinical fields. With the development of optical technology and image sensor technology, the imaging quality of endoscopes has been continuously improved, but there are still many challenges. Especially when observing the microvascular structure of the tissue and evaluating the microcirculation state, due to the special imaging environment of the endoscope (such as uneven lighting, tissue surface reflection, motion artifacts, etc.), the original image often cannot clearly show the details of the blood vessels, affecting the accuracy of clinical diagnosis.

[0003] Microcirculation refers to the blood circulation between arterioles, capillaries and venules, and is an important place for material exchange between tissue cells and blood. Microcirculation abnormalities are closely related to the occurrence and development of many diseases, such as neovascularization of malignant tumors, vascular changes in inflammatory diseases, and insufficient perfusion in ischemic diseases. Therefore, real-time and accurate evaluation of the microcirculation state of the tissue is of great significance for early diagnosis and treatment effect evaluation of diseases.

[0004] Existing endoscopic image blood vessel enhancement techniques mainly include: Filter-based enhancement methods use Hessian matrix analysis or Gabor filters to enhance tubular structures. This type of method identifies and enhances blood vessels by analyzing the local structural features of the image.

[0005] Machine learning-based methods use traditional machine learning algorithms such as support vector machines and random forests for blood vessel segmentation and enhancement.

[0006] Deep learning-based methods: use convolutional neural networks to achieve end-to-end blood vessel segmentation.

[0007] Narrow-band imaging technology (NBI) enhances blood vessel contrast by using specific wavelengths of light, but requires special hardware support.

[0008] Problems and shortcomings of existing technologies: Slow processing speed: Although existing deep learning methods have good results, they have high computational complexity and cannot meet the real-time requirements of endoscopic surgery (usually requiring more than 30 fps). Most methods take more than 100 milliseconds to process, making real-time processing impossible.

[0009] Poor multi-scale blood vessel enhancement effect: Traditional filtering methods are often designed for specific scales of blood vessels and cannot effectively enhance both large blood vessel trunks and small capillaries. When the parameters are adjusted to suit large blood vessels, small blood vessels are easily overlooked; conversely, excessive noise is generated.

[0010] Poor temporal stability: Most existing methods use single-frame independent processing, resulting in flickering and discontinuity in the video sequence, affecting the doctor's visual experience and judgment.

[0011] Lack of quantitative evaluation function: Most methods only provide visual enhancement effects and lack quantitative analysis functions for microcirculation parameters such as blood flow velocity, perfusion density, and blood vessel morphology, making it difficult to provide objective evaluation indicators for clinical practice.

[0012] Low utilization of computing resources: Traditional methods use the same processing strategy for all image regions, resulting in wasted computing resources in simple regions and insufficient processing in complex regions.

[0013] Susceptible to motion artifacts: During endoscopy, physiological movements such as organ peristalsis, respiratory motion, and heartbeat, as well as camera movements caused by the doctor's operation, cannot be avoided. Existing methods are difficult to effectively handle the effects of these movements.

[0014] Single function: Existing systems usually treat image enhancement and parameter analysis as independent processing steps, lack overall optimization, resulting in low processing efficiency, and the functions cannot promote each other. SUMMARY

[0015] To address the shortcomings of existing technologies, the present application provides a real-time processing method for endoscopic image blood vessel enhancement and microcirculation evaluation, which specifically addresses the following technical problems: 1. Solving the problem of slow processing speed of existing methods and inability to meet the real-time requirements of clinical practice, achieving real-time processing capability of more than 30 fps to meet the real-time requirements of clinical practice; 2. Solving the problem of traditional methods that cannot effectively enhance blood vessels of different scales at the same time, achieving full-scale enhancement from the main blood vessel to the capillary; 3. Solving the problem of temporal instability caused by single-frame processing, providing time-coherent and visually stable enhancement effects; 4. Solving the problem of lack of quantitative evaluation function, providing real-time blood flow velocity, perfusion density, and blood vessel morphology microcirculation parameters; 5. Solve the problem of unreasonable utilization of computing resources, and improve overall processing efficiency through adaptive resource scheduling; 6. Solve the problem of mutual independence of image enhancement and parameter analysis, construct an integrated processing flow, and realize functional cooperation of image enhancement and parameter analysis.

[0016] The present application realizes the above-mentioned purposes through the following technical solutions: A real-time processing method for endoscopic image blood vessel enhancement and microcirculation evaluation, comprising: The original endoscopic image is processed by using a CIE LAB color space, the brightness channel and the chrominance channel are separated, the local enhancement of the brightness channel is performed through a contrast-limited adaptive histogram equalization technology, and the discontinuity of the sub-region boundary is eliminated through a bilinear interpolation technology to generate a basic enhanced image; Based on local texture complexity analysis, adaptive frequency-space decomposition is realized, and the basic enhanced image is decomposed into a low-frequency component and a high-frequency component; The low-frequency component is subjected to multi-scale Hessian matrix blood vessel detection, the enhanced main blood vessel structure is calculated through eigenvalue analysis and response strength, and the high-frequency component is subjected to directional Gabor filter group application combined with neighborhood direction consistency constraint to enhance the microvessel continuity; Temporal and spatial feature fusion is realized through multi-resolution pyramid optical flow calculation to improve the time sequence stability; A lightweight multi-task deep learning network based on feature sharing is used to synchronously complete blood vessel probability prediction, blood vessel diameter estimation and blood flow direction prediction; Adaptive GPU computing resource scheduling is realized based on image content complexity evaluation to optimize the computing efficiency; Based on speckle pattern analysis, blood flow velocity is calculated, integral graph acceleration is used to statistically calculate perfusion density, an improved skeleton algorithm is used to extract blood vessel morphological parameters, and multi-dimensional evaluation indexes are synchronously output.

[0017] According to the real-time processing method for endoscopic image blood vessel enhancement and microcirculation evaluation provided by the present application, the original endoscopic image is processed by using a CIE LAB color space, comprising: The RGB color value of the original endoscopic image is converted into a CIE LAB color space through a color space conversion matrix to obtain a brightness channel and two chrominance channels; The converted brightness channel image is extracted, divided into a regular grid of 8x8, and the gray scale histogram of each sub-region is independently calculated to count the number of pixels of each sub-region, and a 256-level gray scale histogram is generated; A contrast limit is implemented on the histogram of each sub-region, the limit threshold is dynamically adjusted according to the size of the sub-region, the histogram part exceeding the threshold is uniformly clipped and redistributed to all gray levels, noise is avoided from being excessively amplified, the equalization is applied to the clipped histogram of each sub-region, and a sub-region image after local enhancement is generated; The transition pixel value of the adjacent sub-region boundary is calculated through a bilinear interpolation technology, block-like artifacts caused by independent processing are eliminated, and the formula is as follows:

[0018] wherein, is a distance-based weight coefficient, is a pixel value of the adjacent sub-region at coordinates (x, y); In the brightness channel enhancement process, the original values of the two chroma channels are kept unchanged, and only the brightness channel is modified. The enhanced brightness channel and the original chroma channel are recombined and converted back to the RGB color space to generate an endoscope image that takes into account brightness contrast and color fidelity.

[0019] According to the real-time processing method for endoscope image blood vessel enhancement and microcirculation evaluation provided by the application, the adaptive frequency-space decomposition is realized based on local texture complexity analysis, and the method comprises the following steps: A 5*5 square sliding window is constructed with each pixel point in the image as the center to cover the neighborhood pixels; The mean value mu and the local variance V of the pixel gray values in the window are calculated to reflect the degree of regional gray variation, and the Sobel operator is used to calculate the gradient values of the pixels in the window in the horizontal direction and the vertical direction to calculate the gradient amplitude G to reflect the edge strength; According to the local variance V and the gradient amplitude G, the comprehensive texture complexity C is calculated by weighted summation, which is expressed by the following formula: C = w1×V + w2×G Wherein, w1 and w2 are preset weight coefficients.

[0020] According to the real-time processing method for endoscope image blood vessel enhancement and microcirculation evaluation provided by the application, the decomposition method is dynamically selected according to the numerical range of the comprehensive texture complexity C: When C < T_smooth, it is determined that the region is smooth, the fast guided filter is used, the original image is used as the guide image, the low-frequency and high-frequency components are decomposed by optimizing the energy function, and the low-frequency component is extracted; When C ≥ T_smooth, it is determined that the region is rich in texture, the high-frequency residual calculation method is adopted, the high-frequency component is extracted through the difference between the original image and the low-pass filtering result; The radius parameter r of the guided filter is dynamically adjusted according to the local complexity: r = r_base × (1 + k×C), where r_base is the base radius and k is the adjustment coefficient; The low-frequency and high-frequency components are output. The low-frequency components are passed to the multi-scale Hessian matrix for blood vessel detection, and the high-frequency components are passed to the directional Gabor filter bank.

[0021] The present invention provides a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment, which performs multi-scale Hessian matrix vascular detection on low-frequency components, including: A scale space of 5 scales is constructed for the low-frequency component image, with the scale parameter σ taking values ​​of {1.0, 1.5, 2.0, 2.5, 3.0} respectively; A Gaussian kernel function is used to smooth the low-frequency components, generating smooth images at various scales. Iσ ( x , y ):

[0022] in, I σ ( x , y ) is the blurred image at coordinates ( x , y The pixel value at () I ( x , y ) is the original image in coordinates ( x , y The pixel value at () G σ ( x , y ) is the Gaussian function in coordinates ( x , y The value at ) σ It is the standard deviation, which controls the degree of fuzziness; Smoothed images at each scale σ I σ ( x , y ), calculate a 2×2 Hessian matrix at each pixel (x,y). H ;

[0023] in, These are the second-order partial derivatives of the image in the x and y directions, respectively, which are approximated by convolution using the Sobel operator; For Hessian matrices HEigenvalue decomposition yields two eigenvalues ​​λ1 and λ2 (|λ1| ≤ |λ2|):

[0024] According to the real-time processing method for endoscopic image vascular enhancement and microcirculation assessment provided by the present invention, the presence of blood vessels is determined based on the relationship of feature values, and a pixel is determined to be a blood vessel when the following conditions are met: When λ1≈0 and λ2<0, it indicates the presence of blood vessels; For pixels that meet the criteria, calculate the blood vessel response intensity R:

[0025] Here, |λ2| reflects the contrast of blood vessels, and the exp term ensures that only tubular structures (λ1 is close to 0) have a strong response; Vascular response intensity at 5 scales R ( σi (i=1,2,...,5) performs pixel-by-pixel comparisons and takes the maximum value as the final response intensity. R final :

[0026] By employing a multi-scale fusion mechanism, the response of both large and small blood vessels is enhanced simultaneously, achieving full-scale vascular coverage; based on the final response intensity... R final Generate a blood vessel enhancement map, where pixel values ​​are related to... R final Proportional; the enhanced blood vessel image is passed to the binarization segmentation module for blood vessel contour extraction and quantitative analysis.

[0027] The present invention provides a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment, which applies a directional Gabor filter bank to high-frequency components, including: Design a Gabor filter bank with eight equally spaced directions, covering the range from 0° to 157.5°, with a directional spacing of 22.5°. The center frequency of each filter is... f 0 is set to 0.25 (normalized frequency), Gaussian envelope standard deviation σ Set to 1.2, generate a real part filter kernel with 8 directions through inverse Fourier transform. Gθi ( x , y (i=1,2,...,8), the kernel size is set to 15×15 pixels to balance directional selectivity and computational efficiency; For high-frequency component images I high ( x , yThe response maps for each direction are obtained by convolving the signal with Gabor filter kernels in eight directions. R θi ( x , y ); For each pixel (x, y), calculate the local consistency of the response in each direction within its 3×3 neighborhood; calculate the response of the pixels in the neighborhood in each direction. θi mean response ; Select a set of candidate directions that meet the following conditions:

[0028] in, Set it to 0.7 to control the consistency threshold; If the candidate direction set is not empty, select the direction with the strongest response as the optimal direction; otherwise, retain the original direction with the largest response. Based on the optimal direction, from the response diagrams of each direction R θi ( x , y Extract the response value of the corresponding pixel from the ) to generate the optimal orientation response map. R opt ( x , y ); Non-maximum suppression is applied to the optimal directional response map, retaining only local maxima to enhance the continuity of microvessels and suppress noise response of non-vascular structures; Optimal directional response map after nonmaximum suppression R opt ( x , y The results are then weighted and fused with the multi-scale Hessian enhancement results of the low-frequency components. The output is a fused enhanced vascular image, in which the microvascular structure is significantly enhanced by directional Gabor filtering while maintaining the integrity of the main blood vessels.

[0029] According to the present invention, a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment is provided, wherein the multi-resolution pyramid structure design includes: Layer 0: The original resolution input image is used directly as the base layer, preserving all detail information; Layer 1: Downsample the image of layer 0 by 1 / 2; before downsampling, use a 5×5 Gaussian kernel (for smoothing filtering to eliminate high-frequency noise and suppress aliasing). Layer 2: The image from Layer 1 is downsampled by 1 / 2 again; similarly, a 5×5 Gaussian kernel smoothing is applied before downsampling to ensure that the main structural information of the image is completely preserved; In each image layer, the Scharr operator is used instead of the traditional Sobel operator to calculate the spatial derivative. Ix , Iy To improve the accuracy of gradient estimation, the Scharr operator enhances sensitivity to edge directions and reduces gradient direction bias by optimizing the convolution kernel coefficients.

[0030] The real-time processing method for endoscopic image vascular enhancement and microcirculation assessment provided by the present invention includes a multi-resolution pyramid optical flow estimation strategy from coarse to fine: Top-level initialization: Initial optical flow estimation is performed at the second layer, utilizing its low-resolution characteristics to transform large displacement motions into small displacement problems; Layer-by-layer propagation and upsampling: The optical flow estimation result of the current layer is upsampled by 2 times through bilinear interpolation and used as the initial optical flow value of the next layer to realize the layer-by-layer refinement of motion information; Iterative optimization: In each image layer, an overdetermined system of equations is established based on a 9×9 local window, and the optical flow increment is solved by weighted least squares method; Convergence control: Each layer is fixed to perform 5 iterations of optimization to ensure that the optical flow estimation converges fully; the weight matrix is ​​dynamically updated during the iteration process to gradually suppress noise interference.

[0031] The present invention provides a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment, which achieves adaptive GPU computing resource scheduling based on image content complexity assessment, including: The input image is divided into non-overlapping 32×32 pixel processing blocks, with each block serving as an independent computational unit. Complexity is evaluated for each processing block, with evaluation metrics including local gradient magnitude, edge density, or texture richness. Based on the evaluation results, the blocks are divided into three complexity levels: high, medium, and low. Three independent CUDA streams are created, corresponding to high, medium, and low complexity processing blocks respectively, to achieve parallel task isolation; GPU computing resources are dynamically allocated according to the complexity level, with more thread blocks and shared memory allocated to high complexity blocks, and lightweight kernel configuration used for low complexity blocks; Within each CUDA stream, asynchronous memory copying is used to transfer the processing block data from the host to the device memory while simultaneously starting the GPU kernel to process the transferred block. Design a shared memory access pattern to ensure that threads within a thread block access shared memory in a merged manner; for high-complexity blocks, increase the shared memory cache capacity to store intermediate calculation results and reduce the overhead of repeated global memory reads; After each CUDA stream is processed, the resulting blocks are fused together in their original positions to form a complete output image.

[0032] Therefore, compared with the prior art, the real-time processing method for endoscopic image vascular enhancement and microcirculation assessment proposed in this invention has the following beneficial effects: 1. Basic image quality improvement effect This invention employs color space conversion combined with contrast-limited adaptive histogram equalization technology to overcome the technical problems of uneven brightness distribution and low contrast in endoscopic images caused by uneven illumination and tissue reflection characteristics. This achieves the technical effect of improving the uniformity of image brightness distribution while maintaining the integrity of color information. By converting the RGB space to the LAB space and performing local adaptive enhancement only on the brightness channel, this invention avoids color distortion, thereby enhancing the contrast between blood vessels and surrounding tissues and laying the image foundation for accurate identification and analysis of subsequent vascular structures.

[0033] 2. Precise enhancement effect on vascular structure This invention employs an adaptive frequency-spatial decomposition mechanism based on local texture complexity, combined with a targeted multi-scale blood vessel detection algorithm. This overcomes the technical problem that traditional single-scale methods cannot simultaneously address both large blood vessel trunks and micro-blood vessel tips, thereby achieving the technical effect of effectively enhancing blood vessel structures at different scales.

[0034] Specifically, this invention uses an adaptive decomposition strategy to intelligently separate an image into a low-frequency component containing the main vascular structure and a high-frequency component containing detailed information. For the low-frequency component, a multi-scale tubular structure detection method based on the Hessian matrix is ​​used to enhance the continuity and integrity of the main vascular trunk. For the high-frequency component, a Gabor filter bank with direction selectivity combined with direction consistency constraints is used to enhance the visibility of microvessels. The organic combination of the two enhancement mechanisms realizes the enhancement of the vascular network structure from the trunk to the tip.

[0035] This invention overcomes the structural breakage and discontinuity problems caused by low signal-to-noise ratio in microvessels by introducing a neighborhood orientation consistency constraint mechanism during Gabor filtering, thereby achieving the technical effect of maintaining the integrity of the microvessel topology.

[0036] 3. Enhanced timing stability This invention employs a spatiotemporal feature fusion mechanism based on multi-resolution pyramid optical flow calculation, overcoming the technical problems of unstable and flickering blood vessel enhancement results in the temporal sequence caused by independent processing of single frames. This achieves the technical effect of maintaining temporal continuity and visual stability in the enhancement results. By using a pyramid structure to achieve coarse-to-fine motion estimation, combined with an adaptive weight allocation strategy based on motion reliability assessment, it fully utilizes the stable information of historical frames to enhance the processing results of the current frame while avoiding motion blur and artifacts in fast-moving regions.

[0037] 4. Multi-task collaborative processing effect This invention employs a lightweight, multi-task deep learning network architecture based on feature sharing. This overcomes the technical problems of traditional methods, which require multiple independent models to complete different tasks, leading to redundant computational resource consumption and isolated information between tasks. Consequently, it achieves the technical effect of simultaneously completing three tasks—vessel probability prediction, vessel diameter estimation, and blood flow direction prediction—through a single network. The shared feature extraction layer allows different tasks to mutually learn feature representations. Vessel probability information guides the accuracy of diameter estimation, while diameter information assists in determining blood flow direction, forming a collaborative enhancement mechanism between tasks.

[0038] 5. Real-time processing performance optimization effect This invention employs an adaptive computing resource scheduling mechanism based on image content complexity assessment. This overcomes the resource utilization inefficiency caused by traditional uniform processing strategies that apply the same computational intensity to all image regions, thus achieving the technical effect of optimizing overall processing efficiency while ensuring processing quality. By intelligently dividing the image into processing blocks of different complexity levels and dynamically allocating corresponding GPU computing resources and processing precision according to the actual processing needs of each block, optimized matching of computing resources and processing tasks is achieved, improving the overall processing capability of the system.

[0039] 6. Quantitative assessment of microcirculation effects This invention employs a comprehensive evaluation technique that integrates speckle pattern analysis, rapid integral plot statistics, and improved skeleton extraction. This overcomes the technical challenge of traditional methods in obtaining real-time quantitative microcirculation parameters during endoscopic examinations, thus achieving the simultaneous output of hemodynamic and vascular morphology parameters. The invention obtains blood flow velocity distribution through a robust velocity estimation algorithm, calculates perfusion density using an efficient regional statistical method, and extracts vascular morphology features through precise skeleton analysis, providing clinicians with multi-dimensional information for assessing microcirculation status.

[0040] 7. Overall system synergy effect This invention overcomes the technical problems of existing technologies where each processing step is independent and lacks overall optimization by constructing a complete processing pipeline from image preprocessing, adaptive decomposition, multi-scale enhancement, spatiotemporal fusion, multi-task learning to parameter extraction. This achieves the technical effect of integrated image enhancement and microcirculation assessment, with each module mutually reinforcing the others. The enhanced vascular structure provides a clearer analytical basis for parameter calculation, and the intermediate results obtained during parameter calculation guide the optimization and adjustment of the enhancement strategy. The organic combination of data flow and control flow among the processing modules forms a collaborative working mechanism, realizing an overall improvement in the quality and efficiency of endoscopic image processing.

[0041] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0042] Figure 1 This is a flowchart of an embodiment of a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment according to the present invention.

[0043] Figure 2 This is a schematic diagram illustrating the architecture of an embodiment of a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment according to the present invention.

[0044] Figure 3 This is a flowchart illustrating the algorithm execution process, decision branches, and loop structure in an embodiment of a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment according to the present invention.

[0045] Figure 4 This is a network architecture diagram of a lightweight multi-task deep learning network in an embodiment of a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment according to the present invention.

[0046] Figure 5 This is a schematic diagram illustrating the calculation of microcirculation parameters in an embodiment of a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment according to the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0048] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0049] See Figures 1 to 5 This embodiment provides a real-time processing method for endoscopic image vascular enhancement and microcirculation assessment, including: Step S1: The original endoscopic image is processed using the CIE LAB color space to separate the luminance channel and chrominance channel. The luminance channel is locally enhanced using contrast-limited adaptive histogram equalization. The discontinuity of the sub-region boundaries is eliminated by bilinear interpolation to generate the basic enhanced image. Step S2: Based on local texture complexity analysis, adaptive frequency-spatial decomposition is implemented to decompose the basic enhanced image into low-frequency components and high-frequency components. Step S3: Perform multi-scale Hessian matrix vessel detection on the low-frequency components, and enhance the main vessel structure through eigenvalue analysis and response intensity calculation; apply a directional Gabor filter bank to the high-frequency components, and enhance the continuity of microvessels by combining neighborhood direction consistency constraints. Step S4: Spatiotemporal feature fusion is achieved through multi-resolution pyramid optical flow calculation to improve temporal stability; Step S5: Using a lightweight multi-task deep learning network based on feature sharing, the probability prediction of blood vessels, the estimation of blood vessel diameter, and the prediction of blood flow direction are completed simultaneously. Step S6: Implement adaptive GPU computing resource scheduling based on image content complexity assessment to optimize computing efficiency; Step S7: Calculate blood flow velocity based on speckle pattern analysis, accelerate the statistical analysis of perfusion density using integral plots, extract vascular morphology parameters using an improved skeleton algorithm, and simultaneously output multidimensional evaluation indicators.

[0050] In step S1 above, the original endoscopic image is processed using the CIE LAB color space, including: The RGB color values ​​of the original endoscopic image are converted to the CIE LAB color space using a color space conversion matrix to obtain the luminance channel and two chrominance channels. Extract the converted luminance channel image, divide it into an 8×8 regular grid, calculate the gray-level histogram independently for each sub-region, count the number of pixels at each gray level in each sub-region, and generate a 256-level gray-level histogram. Apply contrast limiting to the histogram of each sub-region, with a limit threshold. T clip Dynamically adjust based on sub-region size: T clip = N pixels / N bins • k in, N pixels This represents the total number of pixels in the sub-region. N bins The gray level of the histogram is (256).k This is the contrast limiting factor (typical value 0.01~0.1).

[0051] Will exceed the threshold T clip The histogram portion is uniformly cropped and redistributed to all gray levels to avoid excessive noise amplification. The cropped histogram is then applied to each sub-region for equalization to generate a locally enhanced sub-region image. The transition pixel values ​​at the boundaries of adjacent sub-regions are calculated using bilinear interpolation to eliminate blocky artifacts caused by independent processing. The formula is as follows:

[0052] in, For distance-based weighting coefficients, The pixel value of the adjacent sub-region at coordinates (x, y); In the process of enhancing the luminance channel, the original values ​​of the two chroma channels are kept unchanged, and only the luminance channel is modified. The enhanced luminance channel is recombined with the original chrominance channel and converted back to the RGB color space to generate an endoscopic image that balances luminance contrast and color fidelity.

[0053] As can be seen, in step S1, the CIE LAB (Commission Internationale de l'EclairageLab*) color space is used for image processing to effectively separate luminance and chromaticity information. The LAB color space has perceptual uniformity, which is more in line with the characteristics of human vision. Accurate preservation of color information is ensured through a precise color space transformation matrix and nonlinear mapping function. Furthermore, the image is divided into an 8×8 regular grid, and histograms are independently calculated and contrast limiting processing is applied to each sub-region. Bilinear interpolation technology is used to eliminate discontinuities at sub-region boundaries, achieving a smooth enhancement effect. The contrast limiting threshold is adaptively adjusted according to the size of the sub-region to prevent excessive noise amplification.

[0054] In step S2 above, adaptive frequency-spatial decomposition is achieved based on local texture complexity analysis, including: Construct a 5×5 square sliding window centered on each pixel in the image, covering its neighboring pixels; Calculate the mean μ and local variance V of the pixel grayscale values ​​within the window to reflect the degree of grayscale variation in the region:

[0055]

[0056] The Sobel operator is used to calculate the gradient values of pixels within the window in the horizontal and vertical directions respectively, and the gradient magnitude G is calculated to reflect the edge strength:

[0057]

[0058]

[0059] Based on the local variance V and the gradient magnitude G, the comprehensive texture complexity C is calculated through weighted summation, expressed by the following formula: C = w1×V + w2×G where w1 and w2 are preset weight coefficients (typical values are both 0.5).

[0060] According to the numerical range of the comprehensive texture complexity C, the decomposition method is dynamically selected: When C < T_smooth, it is determined as a smooth area, and fast guided filtering is used. With the original image as the guidance map, the low-frequency and high-frequency components are decomposed by optimizing the energy function to extract the low-frequency component; When C ≥ T_smooth, it is determined as a texture-rich area, and the high-frequency residual calculation method is adopted. The high-frequency component is extracted through the difference between the original image and the low-pass filtering result; The radius parameter r of the guided filtering is dynamically adjusted according to the local complexity: r = r_base × (1 + k×C), where r_base is the base radius and k is the adjustment coefficient. This adaptive mechanism ensures that the optimal frequency domain decomposition effect can be obtained in different texture areas.

[0061] The decomposed low-frequency component (including the overall structure information of the image) and high-frequency component (including blood vessel edges and fine texture information) are output. The low-frequency component is passed to the multi-scale Hessian matrix blood vessel detection, and the high-frequency component is passed to the directional Gabor filter bank to achieve targeted enhancement.

[0062] In the above step S3, the multi-scale Hessian matrix blood vessel detection is performed on the low-frequency component, including: A scale space of 5 scales is constructed for the low-frequency component image, and the scale parameters σ of each scale are successively taken as {1.0, 1.5, 2.0, 2.5, 3.0}; The low-frequency component is smoothed by using the Gaussian kernel function to generate the smoothed images at each scale Iσ ( x , y )

[0063] where Iσ ( x , y ) is the blurred image at coordinates ( x , y The pixel value at () I ( x , y ) is the original image in coordinates ( x , y The pixel value at () G σ ( x , y ) is the Gaussian function in coordinates ( x , y The value at ) σ It is the standard deviation, which controls the degree of fuzziness; Smoothed images at each scale σ I σ ( x , y ), calculate a 2×2 Hessian matrix at each pixel (x,y). H ;

[0064] in, These are the second-order partial derivatives of the image in the x and y directions, respectively, which are approximated by convolution using the Sobel operator; For Hessian matrices H Eigenvalue decomposition yields two eigenvalues ​​λ1 and λ2 (|λ1| ≤ |λ2|):

[0065] The presence of blood vessels is determined based on the relationship between feature values. A pixel is identified as a blood vessel if the following conditions are met: When λ1≈0 and λ2<0, it indicates the presence of blood vessels; For pixels that meet the criteria, calculate the blood vessel response intensity R:

[0066] The response function cleverly combines information from two eigenvalues: |λ2| reflects the contrast of blood vessels, and the exp term ensures that only tubular structures (λ1 is close to 0) have a strong response. Vascular response intensity at 5 scales R ( σi (i=1,2,...,5) performs pixel-by-pixel comparisons and takes the maximum value as the final response intensity. R final :

[0067] By employing a multi-scale fusion mechanism, the response of both large and small blood vessels is enhanced simultaneously, achieving full-scale vascular coverage; based on the final response intensity... R final Generate a blood vessel enhancement map, where pixel values ​​are related to... R final The enhanced vascular image is proportional to the blood vessel size; it is then passed to the binarization segmentation module for vascular contour extraction and quantitative analysis. This multi-scale detection and selective enhancement mechanism achieves full coverage detection from large vascular trunks to microvascular tips, overcoming the limitations of traditional single-scale methods.

[0068] In step S3 above, a directional Gabor filter bank is applied to the high-frequency components, including: Design a Gabor filter bank with eight equally spaced directions, covering the range from 0° to 157.5°, with a directional spacing of 22.5°. The center frequency of each filter is... f 0 is set to 0.25 (normalized frequency), Gaussian envelope standard deviation σ Set to 1.2, generate a real part filter kernel with 8 directions through inverse Fourier transform. Gθi ( x , y (i=1,2,...,8), the kernel size is set to 15×15 pixels to balance directional selectivity and computational efficiency; For high-frequency component images I high ( x , y The response maps for each direction are obtained by convolving the signal with Gabor filter kernels in eight directions. R θi ( x , y ); For each pixel (x, y), calculate the local consistency of the response in each direction within its 3×3 neighborhood; calculate the response of the pixels in the neighborhood in each direction. θi mean response ; Select a set of candidate directions that meet the following conditions:

[0069] in, Set it to 0.7 to control the consistency threshold; If the candidate direction set is not empty, select the direction with the strongest response as the optimal direction; otherwise, retain the original direction with the largest response. Based on the optimal direction, from the response diagrams of each direction R θi ( x , yExtract the response value of the corresponding pixel from the ) to generate the optimal orientation response map. R opt ( x , y ); Non-maximum suppression is applied to the optimal directional response map, retaining only local maxima to enhance the continuity of microvessels and suppress noise response of non-vascular structures; Optimal directional response map after nonmaximum suppression R opt ( x , y The results are then weighted and fused with the multi-scale Hessian enhancement results of the low-frequency components. The output is a fused enhanced vascular image, in which the microvascular structure is significantly enhanced by directional Gabor filtering while maintaining the integrity of the main blood vessels.

[0070] In step S4 above, a three-layer Gaussian pyramid is constructed to estimate optical flow from coarse to fine. The Scharr operator is used to calculate the spatial derivative to improve accuracy, and an overdetermined system of equations is established within a 9×9 window. The system is solved using weighted least squares, with weights based on gradient reliability. Five iterations are performed per layer to ensure convergence.

[0071] Specifically, the structural design of the multi-resolution pyramid includes: Layer 0 (Original Layer): The original resolution input image is used directly as the base layer, preserving all detail information; Layer 1 (intermediate layer): The image of layer 0 is downsampled by 1 / 2, reducing the resolution by 50%; before downsampling, a smoothing filter with a 5×5 Gaussian kernel (standard deviation σ=1.0) is used to eliminate high-frequency noise and suppress aliasing. Layer 2 (Top Layer): The image of Layer 1 is downsampled by 1 / 2 again, reducing the resolution to 25% of the original; similarly, a 5×5 Gaussian kernel (σ=1.0) is applied for smoothing before downsampling to ensure that the main structural information of the image is completely preserved; In each image layer, the Scharr operator is used instead of the traditional Sobel operator to calculate the spatial derivative. Ix , Iy To improve the accuracy of gradient estimation, the Scharr operator enhances its sensitivity to edge directions and reduces gradient direction bias by optimizing the convolution kernel coefficients (horizontal direction: [-3 0 3; -10 0 10; -30 3], vertical direction: [-3 -10 -3; 0 0 0; 3 10 3]).

[0072] In this embodiment, the coarse-to-fine optical flow estimation strategy using a multi-resolution pyramid includes: Top-level initialization: Initial optical flow estimation is performed in the second layer (the coarsest layer). Its low-resolution characteristics are used to transform large displacement motion into a small displacement problem, reducing the estimation difficulty. Layer-by-layer propagation and upsampling: The optical flow estimation result of the current layer is upsampled by 2 times through bilinear interpolation and used as the initial optical flow value of the next layer (higher resolution layer), so as to realize the layer-by-layer refinement of motion information; Iterative optimization: In each image layer, an overdetermined system of equations is established based on a 9×9 local window, and the optical flow increment is solved by weighted least squares method; the weights are dynamically allocated according to the gradient reliability, and higher weights are given to high gradient regions (edges) to enhance constraints; Convergence control: Five iterations of optimization are performed at each layer to ensure that the optical flow estimation converges fully; the weight matrix is ​​dynamically updated during the iteration process to gradually suppress noise interference; Through a three-tiered pyramid design, the top layer captures global coarse-grained motion, the middle layer supplements medium-scale details, and the bottom layer corrects local micro-displacements, achieving a gradual approximation from large-scale motion to fine-scale structure. It is evident that large displacement motions are effectively captured in the coarse layer, while detailed motions are accurately estimated in the fine layer, achieving a balance between accuracy and efficiency in motion estimation.

[0073] In step S5 above, this embodiment uses depthwise separable convolutions to reduce the number of parameters and batch normalization to accelerate training convergence. The network includes a shared feature extraction layer and three task-specific output heads. 1×1 convolutions are used to achieve flexible combination of feature channels, keeping the total number of parameters below 10K.

[0074] Specifically, lightweight multitasking networks can be implemented in the following ways: Depth-separable convolutional structure: Depth-separable convolutional layers are used to replace traditional convolutional layers, decomposing standard convolution into two steps: depth convolution and point convolution. Depth convolution performs spatial filtering independently for each input channel, while point convolution achieves cross-channel information fusion through a 1×1 convolutional kernel, significantly reducing the number of parameters. Batch normalization accelerates training: A batch normalization module is inserted after each convolutional layer to eliminate internal covariate bias by standardizing the mean and variance of the input data. In practice, the feature maps of each batch of training data are subjected to mean subtraction and standard deviation scaling, and learnable scaling and offset parameters are introduced to enable the network to adapt to data with different distributions. Shared feature extraction layer design: Construct a shared backbone network containing 5 consecutive depthwise separable convolutional blocks. Each convolutional block contains depthwise convolution, batch normalization, ReLU activation function and point convolution in sequence. The first 3 convolutional blocks use 3×3 convolutional kernels to extract low-level texture features, and the last 2 convolutional blocks use 5×5 convolutional kernels to enhance the multi-scale feature expression capability. Multi-task output head configuration: Three task branches are derived in parallel from the final feature map of the shared feature extraction layer: Blood vessel probability prediction head: The spatial dimension is compressed by a global average pooling layer, followed by two fully connected layers (with 64 and 1 neurons respectively), and the pixel-level probability of blood vessel presence is output. Blood vessel diameter estimation head: A 1×1 convolutional layer is used to generate a single-channel feature map, the blood vessel centerline is extracted by non-maximum suppression, and the blood vessel diameter at each center point is calculated by combining the distance transform algorithm; Blood flow direction prediction head: The gradient magnitude and direction of the feature map are calculated by applying an 8-direction Sobel operator group, and then normalized to the prediction probability of 8 directions by the softmax function, and the main blood flow direction is output. Parameter optimization control: The total number of parameters is constrained by limiting the number of output channels of each convolutional block. The number of channels in the shared feature extraction layer is set to 16, 32, 64, 32, and 16 respectively, and the number of channels in the intermediate layers of the three task heads is set to 8. After quantization verification, the total number of parameters is controlled at 9,876, which meets the requirement of being within 10K.

[0075] In step S6 above, adaptive GPU computing resource scheduling is implemented based on image content complexity assessment, including: The input image is divided into non-overlapping 32×32 pixel processing blocks, with each block serving as an independent computational unit. Complexity is evaluated for each processing block, with evaluation metrics including local gradient magnitude, edge density, or texture richness. Based on the evaluation results, the blocks are divided into three complexity levels: high, medium, and low. Create three independent CUDA streams, corresponding to high, medium, and low complexity processing blocks respectively, to achieve parallel task isolation; GPU computing resources are dynamically allocated based on the complexity level. High-complexity blocks are allocated more thread blocks and shared memory, while low-complexity blocks use a lightweight kernel configuration. Within each CUDA stream, asynchronous memory copying is used to transfer the processing block data from the host to the device memory while simultaneously launching the GPU kernel to process the transferred block. By using the CUDA stream synchronization mechanism, overlapping execution of data transmission and computation is achieved, hiding I / O latency and improving overall throughput; In the GPU kernel, processing block data is loaded from global memory to shared memory, reducing the number of global memory accesses; The shared memory access mode is designed to ensure that threads within a thread block access shared memory in a merged manner, thus avoiding bank conflicts. For highly complex blocks, increase the capacity of the shared memory cache to store intermediate computation results and reduce the overhead of repeated global memory reads; Monitor the execution progress of each CUDA stream in real time. If the processing speed of a certain stream is significantly slower than that of other streams, dynamically adjust the subsequent task allocation strategy. The complexity of incomplete blocks is reassessed, allowing for cross-stream rescheduling to balance the load; After each CUDA stream is processed, the resulting blocks are merged into a complete output image according to their original positions; The actual processing time and resource utilization of each complexity level are statistically analyzed to generate a scheduling efficiency report for subsequent task parameter optimization.

[0076] In step S7 above, the specific implementation methods for synchronous extraction of microcirculation parameters include: Blood flow velocity calculation based on speckle pattern analysis: By analyzing the correlation between speckle motion in consecutive frames, the light intensity fluctuation characteristics caused by blood cell flow in microvessels on the tissue surface are captured. A sliding time window is used to statistically analyze the speckle contrast changes in local areas to generate a blood flow velocity distribution map. The window size is set to 5×5 pixels and the time span is 3 consecutive frames to balance spatial resolution and temporal sensitivity. A motion direction compensation mechanism is introduced to correct the pseudo velocity caused by lens shake or organ movement through optical flow, thereby improving the accuracy of blood flow velocity calculation.

[0077] Statistics on perfusion density accelerated by integral plots: By pre-constructing an integral map of the enhanced blood vessel image, and quickly calculating the pixel sum of any rectangular region through a lookup table, the complexity of perfusion density statistics is reduced from O(n log n). 2 The perfusion density is defined as the ratio of the sum of pixel intensity within the blood vessel region to the region area. Millisecond-level region statistics are achieved using integral images. The blood vessel region is segmented using dynamic thresholds, and the thresholds are adaptively adjusted according to local contrast to ensure statistical robustness under different lighting conditions.

[0078] Improved skeleton algorithm for extracting vascular morphology parameters: An improved Zhang-Suen skeletonization algorithm is applied to enhanced blood vessel images. It retains the vessel centerline by iteratively deleting boundary pixels, while avoiding the topological breakage problem of traditional algorithms. An orientation consistency constraint is introduced during skeletonization to prioritize the retention of pixels extending along the main direction of the vessel and suppress spur noise at the branch ends. Based on the skeleton map, morphological parameters of the blood vessel are calculated, including vessel length (total number of skeleton pixels), number of branches (number of skeleton nodes), curvature (rate of change of skeleton orientation), and diameter (average distance between the vessel edges on both sides of the skeleton).

[0079] Simultaneous output of multi-dimensional evaluation indicators: The blood flow velocity distribution map, perfusion density thermogram, and vascular morphology parameters are superimposed on the enhanced endoscopic image to generate a comprehensive assessment screen containing quantitative information. The output format is a multi-channel PNG image, where the RGB channels encode the vascular enhancement results, blood flow velocity pseudo-color map, and perfusion density map, respectively, and the Alpha channel stores the vascular morphology skeleton. Real-time display and storage are supported, and doctors can switch to view the visualization results of single parameters or combinations of parameters through the interactive interface.

[0080] In summary, this embodiment overcomes the technical problems of existing technologies where each processing step is independent and lacks overall optimization by constructing a complete processing pipeline from image preprocessing, adaptive decomposition, multi-scale enhancement, spatiotemporal fusion, multi-task learning to parameter extraction. This achieves the technical effect of integrated image enhancement and microcirculation assessment, with each module mutually reinforcing the others. The enhanced vascular structure provides a clearer analytical basis for parameter calculation, and the intermediate results obtained during parameter calculation guide the optimization and adjustment of the enhancement strategy. The organic combination of data flow and control flow among the processing modules forms a collaborative working mechanism, realizing an overall improvement in the quality and efficiency of endoscopic image processing.

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A real-time processing method for endoscopic image vascular enhancement and microcirculation assessment, characterized in that, include: The original endoscopic image was processed using the CIE LAB color space, separating the luminance and chrominance channels. The luminance channel was locally enhanced using a contrast-limited adaptive histogram equalization technique. Bilinear interpolation was used to eliminate discontinuities at sub-region boundaries, generating a base enhanced image. Adaptive frequency-spatial decomposition is achieved based on local texture complexity analysis, which decomposes the base enhanced image into low-frequency components and high-frequency components. Multi-scale Hessian matrix vessel detection is performed on low-frequency components, and the main vessel structure is enhanced through eigenvalue analysis and response intensity calculation; directional Gabor filter banks are applied to high-frequency components, and microvascular continuity is enhanced by neighborhood direction consistency constraints. Spatiotemporal feature fusion is achieved through multi-resolution pyramid optical flow computation to improve temporal stability; A lightweight multi-task deep learning network based on feature sharing is used to simultaneously perform blood vessel probability prediction, blood vessel diameter estimation, and blood flow direction prediction. Adaptive GPU computing resource scheduling based on image content complexity assessment optimizes computational efficiency. Blood flow velocity is calculated based on speckle pattern analysis, perfusion density is accelerated by integral plotting, vascular morphology parameters are extracted using an improved skeleton algorithm, and multidimensional evaluation indicators are output simultaneously.

2. The method according to claim 1, characterized in that, The raw endoscopic images were processed using the CIE LAB color space, including: The RGB color values ​​of the original endoscopic image are converted to the CIE LAB color space using a color space conversion matrix to obtain the luminance channel and two chrominance channels. Extract the converted luminance channel image, divide it into an 8×8 regular grid, calculate the gray-level histogram independently for each sub-region, count the number of pixels at each gray level in each sub-region, and generate a 256-level gray-level histogram. Contrast limiting is applied to the histogram of each sub-region. The limiting threshold is dynamically adjusted according to the size of the sub-region. Histogram portions exceeding the threshold are uniformly cropped and redistributed to all gray levels to avoid excessive noise amplification. The cropped histogram is applied to each sub-region for equalization to generate a locally enhanced sub-region image. The transition pixel values ​​at the boundaries of adjacent sub-regions are calculated using bilinear interpolation to eliminate blocky artifacts caused by independent processing. The formula is as follows: in, For distance-based weighting coefficients, The pixel value of the adjacent sub-region at coordinates (x, y); In the process of enhancing the luminance channel, the original values ​​of the two chroma channels are kept unchanged, and only the luminance channel is modified. The enhanced luminance channel is recombined with the original chrominance channel and converted back to the RGB color space to generate an endoscopic image that balances luminance contrast and color fidelity.

3. The method according to claim 1, characterized in that, The adaptive frequency-spatial decomposition based on local texture complexity analysis includes: Construct a 5×5 square sliding window centered on each pixel in the image, covering its neighboring pixels; The mean μ and local variance V of the pixel gray values ​​within the window are calculated to reflect the degree of gray value change in the region. The Sobel operator is used to calculate the gradient values ​​of the pixels within the window in the horizontal and vertical directions, and the gradient magnitude G is calculated to reflect the edge intensity. Based on the local variance V and gradient magnitude G, the overall texture complexity C is calculated by weighted summation, expressed as the following formula: C = w1×V + w2×G Among them, w1 and w2 are preset weight coefficients.

4. The method according to claim 3, characterized in that: Based on the numerical range of the overall texture complexity C, the decomposition method is dynamically selected: When C < T_smooth, it is determined to be a smooth region. Fast guided filtering is used, with the original image as the guide map. The low-frequency and high-frequency components are decomposed by optimizing the energy function to extract the low-frequency component. When C ≥ T_smooth, it is determined to be a region with rich texture. The high-frequency residual calculation method is used to extract the high-frequency components by the difference between the original image and the low-pass filter result. The radius parameter r of the guided filter is dynamically adjusted according to the local complexity: r = r_base × (1 + k×C), where r_base is the base radius and k is the adjustment coefficient; The low-frequency and high-frequency components are output. The low-frequency components are passed to the multi-scale Hessian matrix for blood vessel detection, and the high-frequency components are passed to the directional Gabor filter bank.

5. The method according to claim 1, characterized in that, Multi-scale Hessian matrix vessel detection was performed on the low-frequency components, including: A scale space of 5 scales is constructed for the low-frequency component image, with the scale parameter σ taking values ​​of {1.0, 1.5, 2.0, 2.5, 3.0} respectively; A Gaussian kernel function is used to smooth the low-frequency components, generating smooth images at various scales. Iσ ( x , y ): in, I σ ( x , y ) is the blurred image at coordinates ( x , y The pixel value at () I ( x , y ) is the original image in coordinates ( x , y The pixel value at () G σ ( x , y ) is the Gaussian function in coordinates ( x , y The value at ) σ It is the standard deviation, which controls the degree of fuzziness; Smoothed images at each scale σ I σ ( x , y ), calculate a 2×2 Hessian matrix at each pixel (x,y). H ; in, These are the second-order partial derivatives of the image in the x and y directions, respectively, which are approximated by convolution using the Sobel operator; For Hessian matrices H Eigenvalue decomposition yields two eigenvalues ​​λ1 and λ2 (|λ1| ≤ |λ2|):

6. The method according to claim 4, characterized in that: The presence of blood vessels is determined based on the relationship between feature values. A pixel is identified as a blood vessel if the following conditions are met: When λ1≈0 and λ2<0, it indicates the presence of blood vessels; For pixels that meet the criteria, calculate the blood vessel response intensity R: Here, |λ2| reflects the contrast of blood vessels, and the exp term ensures that only tubular structures (λ1 is close to 0) have a strong response; Vascular response intensity at 5 scales R ( σi (i=1,2,...,5) performs pixel-by-pixel comparisons and takes the maximum value as the final response intensity. R final : By employing a multi-scale fusion mechanism, the response of both large and small blood vessels is enhanced simultaneously, achieving full-scale vascular coverage; based on the final response intensity... R final Generate a blood vessel enhancement map, where pixel values ​​are related to... R final Proportional; the enhanced blood vessel image is passed to the binarization segmentation module for blood vessel contour extraction and quantitative analysis.

7. The method according to claim 1, characterized in that, Applying directional Gabor filter banks to high-frequency components includes: Design a Gabor filter bank with eight equally spaced directions, covering the range from 0° to 157.5°, with a directional spacing of 22.5°. The center frequency of each filter is... f 0 is set to 0.25 (normalized frequency), Gaussian envelope standard deviation σ Set to 1.2, generate a real part filter kernel with 8 directions through inverse Fourier transform. Gθi ( x , y (i=1,2,...,8), the kernel size is set to 15×15 pixels to balance directional selectivity and computational efficiency; For high-frequency component images I high ( x , y The response maps for each direction are obtained by convolving the signal with Gabor filter kernels in eight directions. R θi ( x , y ); For each pixel (x, y), calculate the local consistency of the response in each direction within its 3×3 neighborhood; calculate the response of the pixels in the neighborhood in each direction. θi mean response ; Select a set of candidate directions that meet the following conditions: in, Set it to 0.7 to control the consistency threshold; If the candidate direction set is not empty, select the direction with the strongest response as the optimal direction; otherwise, retain the original direction with the largest response. Based on the optimal direction, from the response diagrams of each direction R θi ( x , y Extract the response value of the corresponding pixel from the ) to generate the optimal orientation response map. R opt ( x , y ); Non-maximum suppression is applied to the optimal directional response map, retaining only local maxima to enhance the continuity of microvessels and suppress noise response of non-vascular structures; Optimal directional response map after nonmaximum suppression R opt ( x , y The results are then weighted and fused with the multi-scale Hessian enhancement results of the low-frequency components. The output is a fused enhanced vascular image, in which the microvascular structure is significantly enhanced by directional Gabor filtering while maintaining the integrity of the main blood vessels.

8. The method according to any one of claims 1 to 7, characterized in that, The structural design of the multi-resolution pyramid includes: Layer 0: The original resolution input image is used directly as the base layer, preserving all detail information; Layer 1: Downsample the image of layer 0 by 1 / 2; before downsampling, use a 5×5 Gaussian kernel (for smoothing filtering to eliminate high-frequency noise and suppress aliasing). Layer 2: The image from Layer 1 is downsampled by 1 / 2 again; similarly, a 5×5 Gaussian kernel smoothing is applied before downsampling to ensure that the main structural information of the image is completely preserved; In each image layer, the Scharr operator is used instead of the traditional Sobel operator to calculate the spatial derivative. Ix , Iy To improve the accuracy of gradient estimation, the Scharr operator enhances sensitivity to edge directions and reduces gradient direction bias by optimizing the convolution kernel coefficients.

9. The method according to claim 8, characterized in that, The coarse-to-fine optical flow estimation strategy using a multi-resolution pyramid includes: Top-level initialization: Initial optical flow estimation is performed at the second layer, utilizing its low-resolution characteristics to transform large displacement motions into small displacement problems; Layer-by-layer propagation and upsampling: The optical flow estimation result of the current layer is upsampled by 2 times through bilinear interpolation and used as the initial optical flow value of the next layer to realize the layer-by-layer refinement of motion information; Iterative optimization: In each image layer, an overdetermined system of equations is established based on a 9×9 local window, and the optical flow increment is solved by weighted least squares method; Convergence control: Each layer is fixed to perform 5 iterations of optimization to ensure that the optical flow estimation converges fully; the weight matrix is ​​dynamically updated during the iteration process to gradually suppress noise interference.

10. The method according to any one of claims 1 to 7, characterized in that, Adaptive GPU computing resource scheduling based on image content complexity assessment includes: The input image is divided into non-overlapping 32×32 pixel processing blocks, with each block serving as an independent computational unit. Complexity is evaluated for each processing block, with evaluation metrics including local gradient magnitude, edge density, or texture richness. Based on the evaluation results, the blocks are divided into three complexity levels: high, medium, and low. Three independent CUDA streams are created, corresponding to high, medium, and low complexity processing blocks respectively, to achieve parallel task isolation; GPU computing resources are dynamically allocated according to the complexity level, with more thread blocks and shared memory allocated to high complexity blocks, and lightweight kernel configuration used for low complexity blocks; Within each CUDA stream, asynchronous memory copying is used to transfer the processing block data from the host to the device memory while simultaneously starting the GPU kernel to process the transferred block. Design a shared memory access pattern to ensure that threads within a thread block access shared memory in a merged manner; for high-complexity blocks, increase the shared memory cache capacity to store intermediate calculation results and reduce the overhead of repeated global memory reads; After each CUDA stream is processed, the resulting blocks are fused together in their original positions to form a complete output image.

Images