Automatic segmentation and quantification method of microcirculatory vessels based on enhanced feature fusion

Abstract

The present invention discloses a method for automatic segmentation and quantification of microcirculatory blood vessels based on enhanced feature fusion, the method comprising the following steps: step one, data preparation and preprocessing; step two, enhanced feature fusion network model design; step three, automatic segmentation of microcirculatory blood vessels; step four, quantification of blood vessel diameter based on segmentation results. The enhanced feature fusion network model proposed in the present invention achieves accurate segmentation of microvessels through the fusion and enhancement of multi-scale features. Through effective data enhancement, the model exhibits good generalization ability on a variety of clinical image data and can adapt to image data with different lighting conditions, contrasts and resolutions. This makes the model highly practical and reliable in practical applications. The present invention can accurately measure the diameter of microvessels by quantifying blood vessel diameters through segmentation results, providing strong support for the evaluation of clinical microcirculatory function.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology and relates to a microcirculation assessment method, and specifically to an automatic segmentation and quantification method of microcirculation blood vessels based on enhanced feature fusion. Background Art

[0002] Microcirculatory function is crucial for maintaining tissue health. Microcirculatory dysfunction is particularly prevalent in acute conditions such as sepsis and is closely associated with disease severity and patient prognosis. Traditional microcirculatory assessment methods rely primarily on direct microscopic observation, which requires high operator skill and is susceptible to subjective influences, making standardization and automation difficult. Furthermore, these methods often fail to provide rapid and accurate quantitative data, limiting their application in clinical decision support. In recent years, deep learning technology has made significant progress in medical image analysis, particularly in image segmentation, lesion detection, and tissue feature quantification. The introduction of deep learning models offers new possibilities for microvascular assessment. However, in microvascular image segmentation and quantification, there remains the challenge of effectively integrating multi-scale features and accurately extracting microscopic details. Summary of the Invention

[0003] To address the existing challenges of microvascular image segmentation and quantification, such as insufficient feature fusion, low segmentation accuracy, and difficulty handling complex tissue structures, the present invention provides an automated microvascular segmentation and quantification method based on enhanced feature fusion. This method employs an improved deep learning network model for the segmentation and quantification of sublingual microcirculatory vessels. The model employs residual connections in the encoder convolution and decoder deconvolution operations, ensuring the integrity of information transfer and gradient stability. Furthermore, the model employs enhanced skip connections, resampling the next-level encoder features using a self-attention mechanism and fusing them with the encoder and decoder features of the current level. This effectively enhances the model's ability to capture detailed microvascular features. Furthermore, the model performs a loss calculation on each decoder feature map relative to the target segmentation map, ensuring segmentation accuracy and consistency. This design significantly improves image segmentation precision and microvascular quantification accuracy, providing a more efficient and reliable tool for automated clinical assessment of microcirculatory function, particularly when dealing with complex tissue structures and poor signal quality.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A method for automatic segmentation and quantification of microcirculatory vessels based on enhanced feature fusion, comprising the following steps:

[0006] Step 1: Data preparation and preprocessing

[0007] Step 1: Collect microcirculatory vascular images under different lighting conditions, contrasts, and resolutions, where the microcirculatory images include static and video images of microcirculation under different health conditions and pathological conditions;

[0008] Step 1 and 2: Extract the region of interest (ROI) of the microcirculatory blood vessel image, select and mark the microvascular area in the image, and crop these areas to form a new ROI image (an image containing the microvascular target area);

[0009] Step 13: Create a corresponding label image for each ROI image and accurately mark the microcirculatory blood vessels;

[0010] Step 14: Save the annotation results as a binary label image, where the microvascular area is marked in white with a pixel value of 1, and the background area is marked in black with a pixel value of 0. The generated label image corresponds one-to-one with the original ROI image to form a complete training sample pair;

[0011] Step 15: Data set images are processed with data augmentation operations before being input into the enhanced feature fusion network model;

[0012] Step 2: Strengthen the design of feature fusion network model

[0013] The enhanced feature fusion network consists of an encoder and a decoder, where:

[0014] In the encoding stage, the input image first passes through a series of residual convolution modules for feature extraction. Each layer of residual convolution module performs a pooling operation after convolution, gradually reducing the spatial resolution of the image and extracting high-level feature information;

[0015] In the decoding stage, the spatial resolution of the image is gradually restored. The decoder features of each level are upsampled through deconvolution operations to restore low-resolution features to high resolution.

[0016] In the upsampling feature fusion process, the design of enhanced jump connection is adopted. Specifically, the encoder features of the next level are deconvolved and upsampled through the self-attention mechanism module, and are fused with the encoder features and decoder features of the current level. After feature fusion, the encoder features x1 and the decoder features x2 are fused through the channel spatial attention module, and then feature extraction is performed through the channel attention mechanism. The feature map is then added with the decoder features x3 upsampled by the self-attention mechanism module. Finally, after spatial attention feature extraction, the feature map x4 is output. The fused feature map x4 is processed by a double-layer convolution module with residual connection. At the same time, the decoder features at all levels are upsampled by pixel reorganization, feature fusion is performed in the channel direction, and features are extracted through a double-layer convolution module. The segmentation result is output through the output layer.

[0017] Step 3: Automatic segmentation of microcirculatory vessels

[0018] Step 31: During the training phase, the preprocessed ROI image and its corresponding binary label image are input and trained using the enhanced feature fusion network model. The network parameters are continuously adjusted through the optimizer to minimize the error between the predicted results and the true label image. The joint loss function L, which combines Dice loss and cross entropy loss, is used for monitoring to ensure that the model can effectively learn microvascular features.

[0019] Step 32: During the testing phase, the microcirculation ROI image is input and automatically segmented using the enhanced feature fusion network model. The model accurately identifies and segments the microvascular area in the image through feature extraction and multi-scale feature fusion, and outputs a binary segmentation result.

[0020] Step 4: Quantification of blood vessel diameter based on segmentation results

[0021] Step 41: Obtain a binary vascular image from the segmentation result output by the enhanced feature fusion network model, and perform skeletonization on the binary vascular image to obtain the centerline of the vascular;

[0022] Step 42: Apply distance transform to the binarized vascular image to generate a distance transform map. In the skeletonized image, extract the distance values ​​on the distance transform map along the center line and multiply these distances by 2 to obtain the local diameter.

[0023] Step 43: By performing statistical analysis on the local diameters, the diameter distribution of blood vessels is obtained, thereby achieving accurate quantification of the diameters of microcirculatory blood vessels.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] 1. The enhanced feature fusion network model proposed in this paper achieves accurate microvascular segmentation by fusing and enhancing multi-scale features. The model can accurately identify and segment microvessels even in complex backgrounds and low-contrast images, demonstrating excellent robustness and adaptability.

[0026] 2. Through effective data augmentation, the model of the present invention demonstrates good generalization capabilities across diverse clinical image data and is adaptable to image data with varying lighting conditions, contrast, and resolution. This makes the model highly practical and reliable in practical applications.

[0027] 3. The present invention quantifies the blood vessel diameter through segmentation results, which can accurately measure the diameter of microvessels and provide strong support for the evaluation of clinical microcirculatory function.

[0028] 4. By introducing residual connections, self-attention mechanism and enhanced skip connection design, the present invention can effectively improve the segmentation accuracy and quantification accuracy of microvascular images. It not only performs well under conditions of limited viewing angle and poor signal quality, but also significantly improves the automation level of microcirculatory function assessment, providing reliable technical support for clinical diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the network model for enhanced feature fusion proposed in the present invention;

[0030] Figure 2 This is an example of the application effect of the method proposed in the present invention in the automatic segmentation of sublingual microcirculation blood vessel images. DETAILED DESCRIPTION

[0031] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0032] The present invention provides a method for automatic segmentation and quantification of microcirculatory vessels based on enhanced feature fusion, the method comprising the following steps:

[0033] Step 1: Data preparation and preprocessing

[0034] To ensure the effectiveness of model training and validation, the present invention conducted detailed microcirculatory image data preparation and preprocessing. First, a large number of microcirculatory static and video images were collected from various sources. These data were primarily derived from high-resolution microscopes and microcirculatory cameras, covering microcirculatory images under various health conditions and pathological conditions. This diverse data included images with varying lighting conditions, contrast, and resolution, ensuring the model's robust generalization capabilities.

[0035] To improve the learning efficiency and segmentation accuracy of the model, ROI (region of interest) extraction is performed on the image. Since the original image usually contains a large amount of irrelevant information, the goal of ROI extraction is to focus on the key areas of the image containing microcirculatory blood vessels, thereby removing unnecessary background information. The present invention uses professional image processing software to manually select and annotate the microvascular regions in the image, and then crop these areas to form a new ROI image. During the cropping process, it is ensured that the selected ROI area contains the complete microvascular structure and is of appropriate size to facilitate subsequent model training.

[0036] In order to conduct supervised deep learning training, a corresponding label image is created for each ROI image, and professional labeling tools such as LabelMe are used to accurately label the microcirculatory vessels. Specifically, each ROI image is first imported into LabelMe, and then the polygon tool (PolygonTool) is used to manually outline the outline of each microvessel to ensure the accuracy and consistency of the labeling. After the labeling is completed, the labeling result is saved as a binary label image, in which the microvascular area is marked as white (pixel value 1) and the background area is marked as black (pixel value 0). Finally, the generated label image corresponds one-to-one with the original ROI image to form a complete training sample pair.

[0037] The annotated dataset was divided into training, validation, and test sets in a ratio of 8:1:1. The training set was used for model parameter learning and optimization, ensuring that the training set covered all microcirculatory vessel types and variations to enhance the model's learning capabilities. The validation set was used to evaluate model performance during training and guide parameter adjustment to prevent overfitting. After model training and validation, the test set was used for final evaluation of the model, testing its generalization and accuracy in real-world applications. The data in the test set was also new to the model during training, ensuring a fair evaluation.

[0038] Before being fed into the network model, the dataset images underwent a series of data augmentation operations, including contrast enhancement and filtering. These preprocessing steps aim to improve image quality, enhance the contrast between blood vessels and the background, and reduce the impact of noise, thus laying the foundation for subsequent feature extraction. The processed input images were sized at 256×256 pixels, ensuring that the model could capture sufficient detail while maintaining computational efficiency.

[0039] Step 2: Strengthen the design of feature fusion network model

[0040] In this paper, a deep learning network model with enhanced feature fusion is designed to achieve accurate segmentation and quantification of microcirculatory vessels. This model is based on the U-Net architecture and combines multi-scale feature enhancement processing to improve the accuracy and robustness of microvascular segmentation.

[0041] The architecture of the enhanced feature fusion network model is as follows Figure 1 During the encoding phase, the input image first passes through a series of residual convolutional modules (Resblock) for feature extraction. Resblock uses two layers of convolution with residual connections to ensure effective feature transfer and gradient stability. Each layer of Resblock convolution is followed by a pooling operation (Down), which gradually reduces the spatial resolution of the image and extracts high-level feature information.

[0042] During the decoding phase, the spatial resolution of the image is gradually restored. The decoder features at each level are upsampled (up) through deconvolution operations, restoring low-resolution features to high resolution. During the upsampled feature fusion process, a reinforced skip connection design is employed. Specifically, the encoder features of the next level are deconvolved and upsampled by a self-attention module (module A1) and then fused with the encoder and decoder features of the current level. To enhance cross-dimensional interactions across channels and space and compensate for the loss of spatial information caused by downsampling, a channel-wise spatial attention module (module A2) is used after feature fusion. In this module, the encoder features x1 and decoder features x2 are fused channel by channel, and feature extraction is performed using the channel-wise attention mechanism. The feature maps are then added to the decoder features x3 upsampled by module A1. Finally, after spatial attention feature extraction, the output feature map x4 is generated. To further enhance the expressive power of the features, the fused feature map x4 is processed through a two-layer convolutional module (Resblock) with residual connections, ensuring the model's accuracy in capturing complex vascular structures. At the same time, the features of the decoders at all levels are upsampled (Up) by pixel reorganization, and feature fusion is performed in the channel direction. The features are extracted by the Resblock module, and the segmentation results are output by the output layer 1×1 convolution layer (1×1Conv).

[0043] By enhancing the feature fusion design, the network model proposed in this paper performs well in the task of microcirculation vessel segmentation. It not only effectively improves the accuracy of segmentation, but also greatly improves the accuracy of microvascular quantification, especially in complex backgrounds and low-contrast images, showing strong robustness and adaptability.

[0044] Step 3: Automatic segmentation of microcirculatory vessels

[0045] Step 3.1: During the training phase, the preprocessed ROI image and its corresponding binary label image are input and trained using the enhanced feature fusion network model. The optimizer continuously adjusts the network parameters to minimize the error between the predicted results and the true label image. Evaluation metrics such as the loss function and the Dice coefficient are used to monitor and ensure that the model effectively learns microvascular features.

[0046] Step 32: During the testing phase, the microcirculatory ROI image is input and automatically segmented using the encoder-decoder reinforced feature fusion network model. Through feature extraction and multi-scale feature fusion, the model accurately identifies and segments the microvascular region in the image, outputting a binary segmentation result.

[0047] Step 3: Quantitatively evaluate the microvascular segmentation results generated by the model. Image segmentation metrics, including the Dice coefficient, average intersection-over-union ratio, and pixel accuracy, are used to assess segmentation accuracy and ensure high consistency between the segmented results and the actual microvascular regions. Furthermore, qualitative analysis of the segmentation results is performed, and the model's practical application is verified through visualization.

[0048] Step 4: Quantification of blood vessel diameter based on segmentation results

[0049] First, a binary vascular image is generated from the segmentation results output by the deep learning model, where the vascular portion is white (value 1) and the background portion is black (value 0). Then, the binary image is skeletonized to obtain the vascular centerline. This centerline is the minimum path that maintains the vascular topology and preserves the vascular topology.

[0050] Next, a distance transform is applied to the binarized image. The distance transform calculates the distance to the nearest background pixel for each pixel in the image, generating a distance map. In the skeletonized image, the distance values ​​from the distance transform map are extracted along the centerline and multiplied by two to obtain the local diameter (since the distance from the centerline to the edge of the vessel is equivalent to the radius).

[0051] Finally, by performing statistical analysis on these local diameters (such as calculating the average, median, etc.), the diameter distribution of blood vessels can be obtained, thereby achieving accurate quantification of the diameter of microcirculatory blood vessels.

[0052] Step 4: Experimental Verification and Application

[0053] In order to verify the effectiveness of the enhanced feature fusion network model designed in the present invention, the model performance was evaluated on clinical sublingual microcirculation detection data.

[0054] In model training, Dice loss and cross entropy loss are combined, and the joint loss function L is expressed as:

[0055] L=αL c +βL d

[0056] Among them, α is the cross entropy loss L c The weight of β is the Dice loss L d In the embodiment, α and β are set to 0.4 and 0.6 respectively. The Adam optimizer is used to optimize and adjust the network parameters.

[0057] A comprehensive performance evaluation of the model was conducted on the previously divided test set. The images in the test set came from different sources and conditions, covering microcirculatory images under complex backgrounds, different contrasts, and various pathological conditions. The segmentation results of each image were obtained by reasoning on the test set. In order to evaluate the accuracy and consistency of the segmentation, commonly used image segmentation evaluation indicators were used, including the Dice coefficient (DSC), mean intersection over union (mIoU), and pixel accuracy (PixelAccuracy, PA). The deep learning model with enhanced feature fusion proposed in the present invention has a Dice coefficient of 0.87, an average intersection over union of 0.81, and a pixel accuracy of 0.89 in the segmentation of sublingual microcirculatory vessels. The experimental results show that the model proposed in the present invention performs well under various test conditions, especially when processing complex backgrounds and low-contrast images, the segmentation results are highly accurate and consistent.

[0058] Figure 2 The visualization results of the model's automatic segmentation of sublingual microcirculatory blood vessels are presented. Experimental results show that the model of the present invention can still accurately identify and segment microvessels in complex backgrounds and low-contrast images.

[0059] During the quantification phase, vascular diameters were measured and analyzed based on the model output. The diameter of each vessel was calculated using the method described in step 3, and a diameter distribution map was generated. The stability and reliability of these quantification results under different experimental conditions were further verified. The experimental results demonstrate that the proposed model has high accuracy in the quantitative analysis of microvascular diameters, with an acceptable error compared to manually labeled results, providing strong support for clinical application.

Claims

1. A method for automatic segmentation and quantification of microcirculatory vessels based on enhanced feature fusion, characterized by The method comprises the following steps: Step 1: Data preparation and preprocessing Step 1: Collect microcirculatory vascular images under different lighting conditions, contrasts, and resolutions, where the microcirculatory images include static and video images of microcirculation under different health conditions and pathological conditions; Step 1 and 2: Extract ROI from the microcirculation blood vessel image, select and mark the microvascular area in the image, and crop these areas to form a new ROI image; Step 13: Create a corresponding label image for each ROI image and accurately mark the microcirculatory blood vessels; Step 14: Save the annotation results as a binary label image, where the microvascular area is marked in white with a pixel value of 1, and the background area is marked in black with a pixel value of 0. The generated label image corresponds one-to-one with the original ROI image to form a complete training sample pair; Step 15: Data set images are processed with data augmentation operations before being input into the enhanced feature fusion network model; Step 2: Strengthen the design of feature fusion network model The enhanced feature fusion network consists of an encoder and a decoder, where: In the encoding stage, the input image first passes through a series of residual convolution modules for feature extraction. Each layer of residual convolution module performs a pooling operation after convolution, gradually reducing the spatial resolution of the image and extracting high-level feature information; In the decoding stage, the spatial resolution of the image is gradually restored. The decoder features of each level are upsampled through deconvolution operations to restore low-resolution features to high resolution. In the upsampling feature fusion process, the design of enhanced jump connection is adopted. Specifically, the encoder features of the next level are deconvolved and upsampled through the self-attention mechanism module, and are fused with the encoder features and decoder features of the current level. After feature fusion, the encoder features x1 and the decoder features x2 are fused through the channel spatial attention module, and then feature extraction is performed through the channel attention mechanism. The feature map is then added with the decoder features x3 upsampled by the self-attention mechanism module. Finally, after spatial attention feature extraction, the feature map x4 is output. The fused feature map x4 is processed by a double-layer convolution module with residual connection. At the same time, the decoder features at all levels are upsampled by pixel reorganization, feature fusion is performed in the channel direction, and features are extracted through a double-layer convolution module. The segmentation result is output through the output layer. Step 3: Automatic segmentation of microcirculatory vessels Step 31: During the training phase, the preprocessed ROI image and its corresponding binary label image are input and trained using the enhanced feature fusion network model. The network parameters are continuously adjusted through the optimizer to minimize the error between the predicted results and the true label image. The joint loss function L, which combines Dice loss and cross entropy loss, is used for monitoring to ensure that the model can effectively learn microvascular features. Step 32: During the testing phase, the microcirculation ROI image is input and automatically segmented using the enhanced feature fusion network model. The model accurately identifies and segments the microvascular area in the image through feature extraction and multi-scale feature fusion, and outputs a binary segmentation result. Step 4: Quantification of blood vessel diameter based on segmentation results Step 41: Obtain a binary vascular image from the segmentation result output by the enhanced feature fusion network model, and perform skeletonization on the binary vascular image to obtain the centerline of the vascular; Step 42: Apply distance transform to the binarized vascular image to generate a distance transform map. In the skeletonized image, extract the distance values ​​on the distance transform map along the center line and multiply these distances by 2 to obtain the local diameter. Step 43: By performing statistical analysis on the local diameters, the diameter distribution of blood vessels is obtained, thereby achieving accurate quantification of the diameters of microcirculatory blood vessels.

2. The microcirculation blood vessel automatic segmentation and quantification method based on enhanced feature fusion according to claim 1 is characterized in that The joint loss function L is expressed as: L=αL c +βL d Among them, α is the cross entropy loss L c The weight of β is the Dice loss L d The weight of .

Images