Multi-scale image processing method, system, medium and equipment

Abstract

The invention relates to a multi-scale image processing method and system, a medium and equipment, and belongs to the field of image processing and computer vision. The method comprises the following steps: data set construction: selecting a medical image public data set containing normal and lesion marks, and dividing the medical image public data set into a training set and a test set; preprocessing data: sequentially executing multi-channel fusion graying, local contrast enhancement and enhanced image mechanical correction; data enhancement: executing random set transformation and adaptive block sampling on the training set; building a multi-scale dynamic space extraction model: building a multi-scale dynamic space extraction module, a space and channel attention jump connection module and a multi-stage feature dense connection module based on U-Net; and training and segmentation: inputting the preprocessed and enhanced training set image into a multi-scale dynamic space extraction model for training, and inputting the test set into the model to obtain an image processing result. The lesion region segmentation precision can be improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing and computer vision, and specifically relates to a multi-scale image processing method, system, medium and device. Background Technology

[0002] Retinal fundus images serve as the core carrier for non-invasively acquiring information about human vascular structures. The morphological features of retinal vessels (such as vessel curvature, diameter, branch density, and the integrity of the tree-like structure) contained within them are key indicators for pathological analysis. Accurate retinal vessel segmentation determines the reliability of subsequent pathological feature extraction, thus becoming a core research direction in the field of fundus image processing. Current retinal vessel processing faces multiple technical challenges: First, fundus images suffer from uneven contrast, brightness fluctuations, and noise interference. Furthermore, the vessels exhibit a complex tree-like branching structure, making manual segmentation time-consuming, labor-intensive, and susceptible to the subjective experience of physicians, making it difficult to meet the needs of large-scale clinical screening. Second, while existing deep learning methods based on U-Net have shown outstanding performance in medical image segmentation, they still suffer from insufficient receptive field, inadequate network depth, poor segmentation quality of small vessels, and low segmentation sensitivity due to extreme imbalance between positive and negative samples, making them unsuitable for the demands of high-precision segmentation.

[0003] Existing retinal vessel processing technologies are mainly divided into two categories: traditional digital image processing methods and deep learning methods. Among them, the technology closest to this invention is the deep learning method based on U-Net and its improved architecture. This type of technology has the following core drawbacks: Insufficient feature extraction capability: Traditional U-Net uses fixed-size convolutional kernels, making it difficult to capture multi-scale vascular features simultaneously, and it does not model long-range dependencies between pixels, resulting in insufficient representation of vascular structure continuity and easy vascular rupture; Low feature transmission efficiency: The skip connections of the original U-Net directly concatenate the encoder and decoder features without filtering effective information, which easily introduces redundant background interference, especially in lesion areas, further reducing feature transmission accuracy and affecting segmentation results; Weak model generalization ability: In view of the common problems of small sample size and imbalance between healthy and lesion samples in datasets, the existing model does not fully integrate multi-stage encoded features, which easily leads to overfitting and significantly insufficient sensitivity for vascular recognition in lesion fundus images. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a multi-scale image processing method, system, medium, and device.

[0005] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a multi-scale image processing method, comprising the following steps: S1. Dataset Construction: Select a publicly available medical image dataset containing normal and lesion annotations, divide it into a training set and a test set, and use the results of manual segmentation by experts as the gold standard; Furthermore, the publicly available dataset is either the DRIVE dataset or the STARE dataset.

[0006] Furthermore, the DRIVE dataset adopts the official data partitioning method of the DRIVE dataset, the STARE dataset uses the entire dataset, and both use the manual partitioning result of the first expert as the gold standard.

[0007] S2. Data preprocessing: Sequentially perform multi-channel fusion grayscale conversion, local contrast enhancement, and enhanced image mechanical correction; S3, Data Augmentation: Perform random set transformation and adaptive block sampling on the training set; Furthermore, the DRIVE dataset uses adaptive block sampling with a total of 100,000 patches, and the number of samples per image is evenly distributed according to the number of images in the training set.

[0008] S4. Construction of Multi-Scale Dynamic Space Extraction Model: Based on U-Net, a multi-scale dynamic space extraction module, a spatial and channel attention skip connection module, and a multi-stage feature dense connection module are constructed. Furthermore, the multi-scale dynamic space extraction module includes: S411. Initialization: The input feature map is processed by three 1×1 convolutions to generate three sets of intermediate feature maps, and learnable parameters are introduced to weight and adjust the three sets of intermediate feature maps. S412, Multi-scale dynamic convolution: Design three parallel branches to perform equivalent 3×3 convolution, equivalent 5×5 convolution and 1×1 convolution respectively, and then fuse them with attention weighting; S413. Convolution and Self-Attention Co-fusion: The intermediate feature map generates self-attention features, which are then fused with the multi-scale convolution feature channels before being output.

[0009] Furthermore, the spatial and channel attention jump connection module includes: S421, Spatial Attention Module: Using the feature maps of the upper and lower encoders as gating signals, a weight map is generated to weight the current feature pixel by pixel, and the spatially refined feature is output. S422, Channel Attention Module: It splices spatially refined features and decoder upsampled features, generates channel weights through global average pooling and fully connected layers, weights each channel, and outputs channel-dimensional refined features.

[0010] Furthermore, the multi-stage feature-dense connection module includes: S431. Feature preprocessing: The features output from each stage of the encoder are pooled to unify the resolution and combined with the high-level semantic features of the bottleneck layer to form the input feature set. S432, Dense Fusion: After pooling in all previous stages, the features and basic feature channels are concatenated and then fused by 1×1 convolution.

[0011] S5. Training and Segmentation: Input the preprocessed and enhanced training set images into the multi-scale dynamic space extraction model for training, and input the test set into the model to obtain the image processing results.

[0012] Secondly, the present invention provides a multi-scale medical image processing system, which performs a multi-scale image processing method, including: Dataset construction module: used to select publicly available medical image datasets containing normal and lesion annotations, divide them into training and testing sets, and use the results of expert manual segmentation as the gold standard; Data preprocessing module: used to sequentially perform multi-channel fusion grayscale conversion, local contrast enhancement, and enhanced image mechanical correction; Data augmentation module: used to perform random set transformations and adaptive block sampling on the training set; Multi-scale dynamic space extraction model construction module: Based on U-Net, a multi-scale dynamic space extraction module, a spatial and channel attention skip connection module, and a multi-stage feature dense connection module are constructed. Training module: Used to input preprocessed and enhanced training set images into the multi-scale dynamic space extraction model for training; Segmentation module: Used to input the test image into the trained model and obtain the image processing result.

[0013] Thirdly, the present invention provides an electronic device, comprising: Processor; memory, wherein executable instructions of the processor are stored; The processor is configured to execute the instructions to implement a multi-scale image processing method.

[0014] Fourthly, the present invention also provides a computer-readable storage medium for storing a program, which, when executed by the processor, implements the steps of a multi-scale image processing method.

[0015] The advantages of this invention are: The multi-scale dynamic spatial extraction model of this invention simultaneously captures local details and global structure of the target region through multi-scale dynamic convolution and self-attention, effectively solving the problem of blood vessel rupture. It alleviates overfitting caused by small sample size and imbalanced distribution in medical datasets by fusing cross-stage features through a multi-stage feature dense connection module. Furthermore, it filters effective features through spatial and channel attention jump connection modules, reducing redundant background interference and maintaining high segmentation accuracy, especially in lesion areas (such as exudation and hemorrhage). The method of this invention is adaptable to multi-resolution and multi-type medical images (such as retinal fundus images), meeting the needs of clinical batch screening and pathological analysis, and providing reliable support for disease diagnosis. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a network architecture diagram of the multi-scale dynamic space extraction model of the present invention; Figure 2 This is a diagram illustrating the data preprocessing effect of the present invention; Figure 3 This is a structural diagram of the multi-scale dynamic space extraction module of the present invention; Figure 4 This is a schematic diagram of the spatial and channel attention jump connection module of the present invention; Figure 5 This is a comparison diagram of the network segmentation effect of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 In this embodiment, the present invention provides a multi-scale image processing method, the specific steps of which include: S1, Dataset Construction We selected the DRIVE and STARE datasets, which are widely recognized in the field of retinal vessel segmentation, to conduct model training and performance testing, respectively, to ensure the comparability and reliability of the experimental results.

[0020] The DRIVE dataset contains 40 retinal fundus images with a resolution of 565×584 dpi, including 7 images of mild fundus lesions and 33 images of normal fundus. In the experimental design, in order to follow the standard protocol of this dataset and ensure comparability with existing studies, this study strictly adopted the official data splitting method of the DRIVE dataset (20 training images and 20 test images), and used the manual annotation results of the first expert as the gold standard.

[0021] The STARE dataset contains 20 annotated color retinal images at a resolution of 700×605 dpi, 10 of which are from healthy individuals and 10 contain retinal lesions of different types and degrees. All 20 annotated images from this dataset were used in the experiment, with the manual segmentation result of the first expert serving as the gold standard.

[0022] S2, Data Preprocessing The preprocessing workflow employs a three-step process: "grayscale conversion - contrast enhancement - nonlinear correction," and the specific steps are as follows: 1. A multi-channel fusion grayscale strategy is adopted to replace the traditional single G-channel grayscale method. After grayscaleing the RGB three channels separately, they are weighted and fused as follows to fully utilize the information of each channel and enhance the three-dimensionality and distinguishability of the blood vessel edge contour; the formula is as follows: , in, This represents the pixel values ​​of the merged grayscale image; Represents the pixel value of the R channel in an RGB image; This represents the pixel value of the G channel in an RGB image; This represents the pixel value of the B channel in an RGB image; 2. To address the issue of uneven contrast in fundus images, CLAHE was used for local contrast enhancement. By limiting the number of local pixel values, over-enhancement was avoided, highlighting vascular details while suppressing image noise, thus laying the foundation for subsequent feature extraction. 3. Perform a Gamma nonlinear transformation (Gamma=1.1) on the image after CLAHE processing; first, normalize the pixel values ​​to the [0,1] interval, and then adjust the brightness and contrast through power function mapping to further enhance the grayscale difference between the micro-vessels and the background and improve the recognizability of the blood vessel area.

[0023] Preprocessed image, such as Figure 2As shown, from left to right, the images are the original retinal fundus image, the grayscale image, the CLAHE equalized image, and the Gamma-transformed image. The original fundus image contains an orange-red background, dendritic vessels (the main trunk is dark red and thick, with gradually thinning branches), and slight dotted noise, with local brightness fluctuations. The grayscale image fuses the grayscale values ​​of the RGB three channels to eliminate channel redundancy, but the overall contrast is low, and some small vessels are still blurred. The CLAHE equalized image limits over-enhancement to suppress noise. The Gamma-transformed image eliminates brightness fluctuations and shows strong vessel continuity.

[0024] S3 Image Data Enhancement 1. Random geometric transformation: Random horizontal / vertical flipping is performed on the training set images, with a probability of 0.5 for each, to simulate the natural differences in the shooting angle of clinical fundus images and enhance the robustness of the model to changes in image orientation.

[0025] 2. Adaptive Patch Sampling: To address the image resolution differences between the DRIVE and STARE datasets, a preset total number of patches is used (100,000 for the DRIVE dataset). The number of samples per image is evenly distributed based on the number of images in the training set (e.g., 5,000 patches per image for 20 images in the DRIVE training set). By randomly extracting fixed-size patches, a single image is transformed into multiple local samples. This increases the amount of training data while allowing the model to focus on local vascular details and reduce background redundancy.

[0026] S4 Multi-Scale Dynamic Space Extraction Model Construction Based on the U-Net framework, a three-tiered collaborative architecture of "encoder-bottleneck layer-decoder" is constructed, focusing on four core stages: "feature extraction, information transfer, feature fusion, and loss supervision." Specific module designs are as follows: Figure 1 As shown.

[0027] 1) Multi-scale dynamic space extraction module like Figure 3 Specifically, it includes: (1) Initialization Input feature map Intermediate feature maps are generated using three 1×1 convolutions. To avoid degradation of initial features, learnable parameters are introduced. (Initial value set to 1), the intermediate feature maps are dynamically adjusted to ensure that the features entering subsequent calculations have at least basic representational capabilities, laying a stable foundation for multi-scale feature extraction and global information modeling. The three sets of intermediate feature maps are dynamically weighted and adjusted: , in, The feature map is used as the input model. The feature map after performing the convolution operation. The adjusted feature map is a learnable parameter. ( (where is the number of intermediate channels), which serves as the input for subsequent multi-scale convolutions.

[0028] (2) Multi-scale dynamic convolution module Three parallel convolutional branches are designed to perform multi-scale dynamic convolution, adapting to the morphological characteristics of blood vessels that are "uneven in thickness and complex in branching": The first branch uses an equivalent 3×3 convolution to capture the directional features of medium-diameter blood vessels: First, for Perform up, down, left, and right offset feature shifting operations to generate shift feature maps. Then, after concatenating the shifted feature maps along the channel dimension, a 1×1 convolution is performed to compress the channels, resulting in an equivalent 3×3 convolution result: , in, For the K shift feature maps in the first branch, This represents the result of the first branch. For height; Width; Indicates the number of output channels in a single branch. For splicing operations, This is a convolution operation; The second branch uses an equivalent 3×3 convolution to cover the entire vascular structure, and it is a direct response to the output of the first branch. Perform feature shifting, concatenation, and convolution operations again: , in, yes The feature map after shifting, This indicates the result of the second branch; The third branch uses 1×1 convolution to capture fine-grained local features: , in, This indicates the result of the third branch.

[0029] Introducing attention module to generate coefficients (Weights sum to 1), the convolutional kernel weights applied to the 3 branches, the weighted branch features are: , in, This represents the feature map of the first branch after attention weighting; This represents the feature map of the second branch after attention weighting; This represents the feature map of the third branch after attention weighting; the features of the three weighted branches are concatenated along the channel dimension, and then compressed to the target number of channels using a 1×1 convolution. To obtain multi-scale local features : .

[0030] (3) Co-fusion of convolution and self-attention First, through weighted adjustment of intermediate feature maps Calculate the query (Q), key (K), and weighted value (V) to generate self-attention features. Establish long-distance dependencies between pixels through self-attention to complete the global structural association of blood vessels. , Then, calculate the self-attention weights: , in, Indicates the self-attention weights; The dimension is Q / K (to avoid gradient vanishing); Softmax represents the activation function, ensuring that the sum of the weights is 1; Next, self-attention features are generated. : , Finally, multi-scale local features With self-attention characteristics The channels are stacked and fused using 1×1 convolutions to obtain the final output features. : , in, For splicing operations, This represents the number of input channels for the next layer of the encoder.

[0031] In this embodiment, the multi-scale dynamic space extraction module is arranged in each stage of the encoder. Let the input features of the l-th encoder stage be... The output features are obtained after processing by this module. .in, On the one hand, the features are downsampled or stride-convolutioned and used as input features for the (l+1)th encoder stage. On the other hand, the skip connection features cached for the l-th stage are input to the spatial and channel attention skip connection module when the decoder reaches the l-th stage corresponding to its resolution.

[0032] 2) Spatial & Channel Attention Jump Connection To address the problem of redundant information interference in traditional skip connections, such as Figure 4 As shown, a two-stage feature purification path of "spatial attention filtering - channel attention optimization" is designed to enhance the effective feature transfer from the encoder to the decoder: (1) Spatial attention module a. Gating signal processing: Select the feature map of the upper encoder layer of the current skip connection layer. With lower layer encoder feature map As a gating signal (containing feature context information); through downsampling operation Resolution adjusted to the current layer feature map The resolution is adjusted by upsampling. Resolution unified to the current layer feature map The resolution is the same, avoiding feature bias caused by random gating; that is: , ,in, This indicates a downsampling operation. Indicates an upsampling operation; This represents the feature map of the upper-layer encoder after downsampling; This represents the feature map of the lower-layer encoder after upsampling; The U-Net encoder of the present invention comprises four (4) encoder blocks, each corresponding to one of the four scales: Encoder block 1: Full resolution; Encoder block 2: 1 / 2 resolution; Encoder block 3: 1 / 4 resolution; Encoder block 4: 1 / 8 resolution; Each encoder block contains: multi-scale dynamic space extraction module, batch normalization & ReLU activation.

[0033] The decoder correspondingly contains four (4) decoder blocks, which are symmetrical to the four layers of the encoder: Decoder Block 1: Connected to Encoder Layer 4; Decoder Block 2: Connected to Encoder Layer 3; Decoder Block 3: Connected to Encoder Layer 2; Decoder Block 4: Connected to Encoder Layer 1; Each decoder block includes: Spatial Channel Attention Skip Connection Module, Batch Normalization & ReLU Activation.

[0034] b. Attention coefficient generation: and After concatenation along the channel dimension, and compression of the channel number via 1×1 convolution, it is then combined with... Perform additive attention operations, sequentially generating a spatial attention weight map S through ReLU activation and the Sigmoid function: , in, Represents the gating feature map; For ReLU functions; For the Sigmoid function; S∈ The weight range is (0,1); This indicates additive attention computation; c. Feature-weighted optimization: The attention weight map S is then compared with... Pixel-by-pixel multiplication outputs features with reduced spatial dimensions. This lays the foundation for subsequent channel integration. ; This indicates a pixel-by-pixel multiplication operation.

[0035] (2) Channel attention module Will Output features of the upper-layer decoder Superimposed along the channel dimension, we obtain superimposed features. The formula is expressed as: The SE module is introduced to perform feature filtering along the channel dimension, dynamically enhancing blood vessel-related channels. a. Use global average pooling to combine features The spatial dimension is compressed to 1×1, global information from each channel is aggregated, and channel statistical descriptors are generated. : , in, These represent the pixel positions of the feature map in the height (H) and width (W) directions, respectively. , ( for (Number of channels); b. Attention coefficient learning: The input is a fully connected layer with ReLU activation for dimensionality reduction, and then the channel attention coefficients are output via a sigmoid function. : , in, , These are two different parameters of the fully connected layer. The dimensionality reduction coefficient; c. Channel feature weighting: ... and Channel-by-channel multiplication, outputting the purified features of each channel. : , in, This indicates a channel-by-channel multiplication operation.

[0036] 3) Multi-stage feature-dense connection module To address the problem of semantic separation of layered features in encoders, this module, referencing the DenseNet concept, constructs a multi-stage feature progressive fusion cross-layer gradient transfer mechanism to achieve feature complementarity between "low-level edges, mid-level textures, and high-level semantics," as detailed below: First, the feature maps output from different stages of the encoder are... High-level semantic features at the bottom layer of the bottleneck layer As input, features are arranged in descending order of resolution and ascending order of semantics to establish a hierarchical relationship: , in, for After pooling, when i=0, For the original input image, ensure the alignment of feature resolutions at multiple stages, laying the foundation for cross-layer fusion; For the first Stage characteristics, unified resolution through pooling, in this invention , Then, referring to the design of DenseNet, the feature fusion path is designed, the first... Layer features Perform convolutional fusion with features from all preceding stages: , Here, Concat represents the concatenation operation, enabling information complementarity across stages; the Conv operation... Convolution compresses the number of channels, balancing computational cost and feature representation, and generates fused features. This enhances the efficiency of feature reuse.

[0037] S5 inputs the preprocessed and enhanced training set images into the multi-scale dynamic space extraction model for training, and inputs the test set into the model to obtain the image processing results.

[0038] Example 2 In this embodiment, as Figure 5 As shown, this paper selects an image from the test set of the DRIVE dataset and puts it into the model of this invention for processing. The processed result, the binarized result of the image, and the gold standard manually divided by experts are combined into a single image for a brief comparison.

[0039] From the perspective of vascular continuity, the segmentation results of the method of this invention are significantly better than those of the gold standard manually annotated by experts than those of traditional binarization. Binarized segmentation images exhibit obvious vascular breakpoints, especially at branching points, where relying on fixed grayscale thresholds makes it difficult to capture the global structural relationships of the vessels, easily resulting in a "break" phenomenon. In contrast, the method of this invention, through a multi-scale dynamic spatial extraction module to simultaneously capture local details and global structures of the vessels, and through spatial-channel attention jump connections to accurately transmit effective features, results in better connectivity between the main trunk and branches of the vessels in the segmentation results, with the overall morphology more closely resembling the tree-like structure manually annotated by experts.

[0040] The differences among the three methods are even more pronounced in the challenging microvascular and densely vascular regions of fundus images. Traditional binarization methods are prone to losing vascular terminals, or failing to distinguish between low-contrast microvascular and background noise, due to their inability to differentiate between them. This results in a significant gap between the results and the detailed integrity of manually annotated images by experts. In contrast, the method of this invention uses multi-stage feature dense connection to fuse "low-level details, mid-level texture, and high-level semantics," and multi-stage output fusion loss function to focus on optimizing microvascular pixels. This method can accurately preserve the branching structure of vascular terminals and achieve clear differentiation of blood vessels in dense regions. The segmentation results show a significantly higher degree of agreement with manually annotated images by experts in terms of detail compared to binarization methods.

[0041] The experimental results of the method of the present invention and existing methods on the DRIVE and STARE datasets are compared in Tables 1 and 2: Table 1. Performance comparison of various networks on the DRIVE dataset. The sensitivity of this invention reaches 0.8137, which is 2.49% higher than the attention U-Net method and 2.97% higher than the fully convolutional autoencoder method. The core reason is that the multi-scale dynamic space extraction module captures the features of small blood vessels through multi-scale dynamic convolution, while the multi-stage fusion loss function alleviates the imbalance between positive and negative samples (blood vessel pixels account for only 7%), reducing the missed detection of small blood vessels. Leading overall performance: The area under the curve (AUC) (0.9812) and F1-value (0.8326) are both the best. In particular, the F1-value is 1.25% higher than the second-best bidirectional convolutional long short-term memory network attention U-Net method, proving that the model is better at balancing "accurate segmentation and reduced missed detections".

[0042] Table 2: Performance comparison of various networks on the STARE dataset The sensitivity of the method in this invention is 9.15% higher than that of the attention U-Net method. The core reason is that the multi-stage feature dense connection module integrates "low-level edge-medium-high-level semantic" features, which enhances the robustness of lesion area recognition.

[0043] Example 3 This embodiment provides a multi-scale medical image processing system that performs a multi-scale image processing method, including: Dataset construction module: used to select publicly available medical image datasets containing normal and lesion annotations, divide them into training and testing sets, and use the results of expert manual segmentation as the gold standard; Data preprocessing module: used to sequentially perform multi-channel fusion grayscale conversion, local contrast enhancement, and enhanced image mechanical correction; Data augmentation module: used to perform random set transformations and adaptive block sampling on the training set; Multi-scale dynamic space extraction model construction module: Based on U-Net, a multi-scale dynamic space extraction module, a spatial and channel attention skip connection module, and a multi-stage feature dense connection module are constructed. Training module: Used to input preprocessed and enhanced training set images into the multi-scale dynamic space extraction model for training; Segmentation module: Used to input the test image into the trained model and obtain the image processing result.

[0044] Example 4 This embodiment provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the multi-scale image processing method described in Embodiment 1.

[0045] Example 5 This embodiment provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the steps of the multi-scale image processing method described in Embodiment 1.

[0046] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of multi-scale image processing, characterized by, The method comprises the following steps: S1, dataset construction: selecting a medical image public dataset containing normal and lesion annotations, dividing into a training set and a test set, and taking expert manual segmentation results as a gold standard; S2, data preprocessing: sequentially performing multi-channel fusion graying, local contrast enhancement, and enhanced image mechanical correction; S3, data augmentation: performing random set transformation and adaptive block sampling on the training set; S4, multi-scale dynamic spatial extraction model construction: constructing a multi-scale dynamic spatial extraction module, a spatial and channel attention jump connection module, and a multi-stage feature dense connection module based on U-Net; S5, training and segmentation: inputting the preprocessed and enhanced training set images into the multi-scale dynamic spatial extraction model for training, inputting the test set into the model, and obtaining image processing results.

2. The method of claim 1, wherein, The multi-scale dynamic spatial extraction module comprises: S411, initialization: inputting a feature map to generate three groups of intermediate feature maps through three 1x1 convolutions, and introducing learnable parameters to weight and adjust the three groups of intermediate feature maps; S412, multi-scale dynamic convolution: designing three parallel branches to perform equivalent 3x3 convolution, equivalent 5x5 convolution, and 1x1 convolution respectively, and performing attention weighted fusion; S413, convolution and self-attention collaborative fusion: generating self-attention features from the intermediate feature maps, and outputting after channel fusion of the multi-scale convolution features.

3. The method of claim 2, wherein, The spatial and channel attention jump connection module comprises: S421, spatial attention module: taking the upper layer encoder feature map and the lower layer encoder feature map as a gating signal, generating a weight map to weight the current feature pixel by pixel, and outputting a spatial purification feature; S422, channel attention module: concatenating the spatial purification feature and the decoder up-sampling feature, generating a channel weight through global average pooling and a fully connected layer, weighting each channel, and outputting a channel dimension purification feature.

4. The method of claim 3, wherein, The multi-stage feature dense connection module comprises: S431, feature preprocessing: pooling the features output by each stage of the encoder to unify the resolution, and constructing an input feature set with the bottleneck layer high-level semantic features; S432, dense fusion: concatenating the pooled features of all previous stages and the basic features, and performing 1x1 convolution fusion.

5. The method of claim 4, wherein, The public dataset is a DRIVE dataset or a STARE dataset.

6. The method of claim 5, wherein, The DRIVE dataset adopts the official data division method of the DRIVE dataset, and the STARE dataset is used in full, and the expert manual segmentation result is taken as the gold standard.

7. The method of claim 1, wherein, The total number of adaptive block sampling patches of the DRIVE dataset is 100000, and the number of single image samples is evenly distributed according to the number of training set images.

8. A multi-scale medical image processing system, performing a multi-scale image processing method as claimed in any one of claims 1-7, characterized in that, It comprises: A dataset construction module for selecting a medical image public dataset containing normal and lesion annotations, dividing into a training set and a test set, and taking expert manual segmentation results as a gold standard; A data preprocessing module for sequentially performing multi-channel fusion graying, local contrast enhancement, and enhanced image mechanical correction; A data augmentation module for performing random set transformation and adaptive block sampling on the training set; The multi-scale dynamic space extraction model construction module comprises a multi-scale dynamic space extraction module, a space and channel attention jump connection module, and a multi-stage feature dense connection module, which are constructed based on a U-Net; The training module is configured to input the preprocessed and enhanced training set images into the multi-scale dynamic space extraction model for training; The segmentation module is configured to input a test image into the trained model to obtain an image processing result.

9. An electronic device, comprising: The method comprises the following steps: a processor; a memory having executable instructions of the processor stored therein; The processor is configured to execute the instructions to implement the multi-scale image processing method of any one of claims 1-7.

10. A computer readable storage medium for storing a program, characterized in that, The program is executed by the processor to implement the steps of the multi-scale image processing method of any one of claims 1-7.

Images