Myopia image deep learning recognition model training method

Abstract

The invention discloses a myopia image deep learning recognition model training method, particularly relates to the technical field of medical image processing and deep learning, and is used for solving the problem that an existing deep learning model lacks anatomical structure priori knowledge guidance in myopia eye bottom image analysis. The method comprises the following steps: acquiring a myopia eye bottom image and anatomical structure priori knowledge data, extracting a multi-scale feature map by using a deep learning model, analyzing the geometric morphology of a key anatomical component based on standard spatial relationship information, and generating a spatial constraint loss item; according to the method, key anatomical path topology coherence is evaluated based on topology connection information, topology constraint loss items are generated, a loss item fusion strategy is dynamically adjusted according to a training stage, finally, a model is iteratively trained to convergence through a gradient back propagation algorithm, and organic combination of medical priori knowledge and a deep learning model is realized. And the clinical rationality and reliability of model output are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing and deep learning, more particularly, the present application relates to a myopic image deep learning recognition model training method. BACKGROUND

[0002] Medical image analysis technology based on deep learning has been widely applied in the field of ophthalmic auxiliary diagnosis, especially for the identification of myopia-related fundus lesions. The existing technology usually adopts an end-to-end training mode, uses a large amount of labeled data to drive a convolutional neural network to automatically learn image features, so as to complete classification, segmentation and other tasks, and relies on the statistical rules hidden in the data to construct feature representation and mapping relationship.

[0003] However, in the training process of the existing deep learning model, the learning goal is completely driven by the data loss function, and there is a lack of effective fusion of domain prior knowledge. In the analysis of myopic fundus images, the key anatomical structures and their spatial relationships have clear medical significance, but the purely data-driven training method cannot guarantee that the model learns the feature representation that meets the anatomical constraints, resulting in that the model output result may violate the medical prior knowledge, affecting its clinical applicability and reliability. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a myopic image deep learning recognition model training method to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: The myopic image deep learning recognition model training method comprises: S1, obtaining a plurality of myopic fundus images and corresponding anatomical structure prior knowledge data, the anatomical structure prior knowledge data comprising standard spatial relationship information and topological connection information of key anatomical components; S2, using a deep learning model to perform multi-scale feature extraction on the myopic fundus images to obtain multi-scale feature maps; S3, based on the standard spatial relationship information, analyzing the geometric morphology composed of the key anatomical components from the multi-scale feature maps, and obtaining a spatial constraint loss term by verifying the degree of conformity between the geometric morphology and the standard geometric morphology; S4, based on the topological connection information, evaluating the topological coherence of the key anatomical path represented by the multi-scale feature maps in the feature space to obtain a topological constraint loss term; S5, dynamically adjusting the fusion strategy of the spatial constraint loss term and the topological constraint loss term according to the model training stage, and sequentially executing a dominant training stage dominated by the spatial constraint loss term and a fine-tuning training stage dominated by the topological constraint loss term; S6. Based on the fusion strategy, the deep learning model is iteratively trained using the gradient backpropagation algorithm, and the model parameters are adjusted until convergence.

[0006] Furthermore, multiple myopic fundus images and corresponding prior anatomical knowledge data are acquired. This prior anatomical knowledge data includes standard spatial relationship information and topological connectivity information of key anatomical components, including: Acquire multiple fundus images of nearsighted individuals; Obtain standard spatial relationship information of key anatomical components based on medical anatomical atlases; Topological connection information of key anatomical components is obtained based on medical anatomical atlases; The medical anatomy atlas contains predefined anatomical structure data.

[0007] Furthermore, a deep learning model is used to extract multi-scale features from the myopic fundus images, resulting in multi-scale feature maps, including: Low-level features of myopic fundus images are extracted using the convolutional layers of a deep learning model. Low-level features are input into different layers of a deep learning model for multi-scale feature fusion. The deep learning model outputs multi-scale feature maps corresponding to features at different scales from different levels.

[0008] Furthermore, based on standard spatial relationship information, the geometric shapes of key anatomical components are parsed from the multi-scale feature map. The spatial constraint loss term is obtained by verifying the degree of conformity between the geometric shape and the standard geometric shape, including: Extracting spatial location information of key anatomical components from multi-scale feature maps; Construct the geometric shape of key anatomical components based on spatial location information; Computational geometry differs from standard geometry; Spatial constraint loss terms are generated based on shape differences.

[0009] Furthermore, the calculation of the shape difference between the geometric shape and the standard geometric shape includes: calculating the average Euclidean distance between key points as the shape difference by comparing the key point coordinates of the geometric shape with those of the standard geometric shape; wherein the key point coordinates are obtained by regression from the multi-scale feature map, and the key point coordinates of the standard geometric shape are defined based on standard spatial relationship information.

[0010] Furthermore, based on topological connectivity information, the topological coherence of key anatomical paths represented by multi-scale feature maps is evaluated in the feature space to obtain a topological constraint loss term, including: Construct an adjacency matrix for key anatomical paths based on topological connectivity information; Extract feature vectors of each node in the key anatomical path from the multi-scale feature map; Calculate the topological coherence measure of key anatomical paths based on adjacency matrix and eigenvectors. Generate a topology constraint loss term based on the topology coherence metric.

[0011] Furthermore, the calculation of the topological coherence measure of the key anatomical path based on the adjacency matrix and eigenvectors includes: determining the connection relationship of nodes in the key anatomical path according to the adjacency matrix, calculating the cosine similarity between the eigenvectors of adjacent nodes, and taking the average of the cosine similarity of all adjacent nodes on the path as the topological coherence measure; wherein the eigenvectors are extracted from the multi-scale feature map.

[0012] Furthermore, based on the dynamic adjustment strategy of the fusion of spatial constraint loss term and topological constraint loss term during the model training phase, a dominant training phase led by the spatial constraint loss term and a fine-tuning training phase led by the topological constraint loss term are executed sequentially, including: During the dominant training phase, the weight of the spatial constraint loss term is set to be greater than the weight of the topological constraint loss term; During the fine-tuning training phase, the weight of the topological constraint loss term is set to be greater than the weight of the spatial constraint loss term; Switch from the dominant training phase to the fine-tuning training phase based on the number of training iterations.

[0013] Furthermore, based on the fusion strategy, the deep learning model is iteratively trained using the gradient backpropagation algorithm, adjusting the model parameters until convergence, including: According to the fusion strategy, the spatial constraint loss term and the topological constraint loss term are combined into a comprehensive loss function with corresponding weights. Calculate the gradient of the comprehensive loss function with respect to the parameters of the deep learning model; Update the parameters of the deep learning model based on the gradient; Repeat the parameter update process until the deep learning model training converges.

[0014] Furthermore, calculating the gradient of the comprehensive loss function with respect to the parameters of the deep learning model includes: calculating the partial derivatives of the comprehensive loss function with respect to the parameters of each layer of the deep learning model using the chain rule to obtain the gradient; wherein the comprehensive loss function is a weighted combination of spatial constraint loss term and topological constraint loss term according to the fusion strategy.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By systematically embedding prior medical knowledge into a deep learning framework, the model's ability to perceive anatomical structures in myopic fundus image analysis is effectively improved. A dual supervision mechanism of spatial and topological constraints is constructed, simultaneously optimizing geometric accuracy and topological coherence during feature learning. The spatial constraint loss term ensures that the spatial distribution of key anatomical components conforms to medical standards, while the topological constraint loss term maintains the rationality of anatomical path connections. The model not only learns statistical features but also grasps anatomical patterns consistent with clinical experience. Through a dynamically adjusted two-stage training strategy, the model can learn spatial localization and topological relationships in stages, ultimately obtaining feature representations with both pixel-level accuracy and structural rationality.

[0016] 2. Significantly improved the structural preservation ability and clinical applicability of deep learning models in ophthalmic image analysis. By transforming anatomical prior knowledge into differentiable constraints, the model naturally incorporates medical norms during end-to-end training. Its output anatomical structure localization results maintain the sensitivity advantage of deep learning, making the model more robust to changes in image quality and pathological interference, ensuring medical rationality while maintaining high accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart of the training method for the deep learning recognition model of near-view images according to 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: Figure 1 The present invention provides a method for training a deep learning recognition model for near-view images, comprising: S1. Acquire multiple myopic fundus images and corresponding prior knowledge data of anatomical structures. The prior knowledge data of anatomical structures includes standard spatial relationship information and topological connection information of key anatomical components. S2. Use a deep learning model to extract multi-scale features from myopic fundus images to obtain multi-scale feature maps; S3. Based on standard spatial relationship information, the geometric shape of key anatomical components is parsed from the multi-scale feature map, and the spatial constraint loss term is obtained by verifying the degree of conformity between the geometric shape and the standard geometric shape. S4. Based on topological connectivity information, evaluate the topological coherence of key anatomical paths represented by multi-scale feature maps in the feature space and obtain the topological constraint loss term. S5. Dynamically adjust the fusion strategy of spatial constraint loss term and topological constraint loss term according to the model training stage, and execute the dominant training stage dominated by spatial constraint loss term and the fine-tuning training stage dominated by topological constraint loss term in sequence. S6. Based on the fusion strategy, the deep learning model is iteratively trained using the gradient backpropagation algorithm, and the model parameters are adjusted until convergence.

[0020] S1. Acquire multiple myopic fundus images and corresponding prior anatomical structure knowledge data. The prior anatomical structure knowledge data includes standard spatial relationship information and topological connection information of key anatomical components. Specifically, the implementation is as follows: Multiple myopic fundus images were acquired from publicly available medical imaging datasets, such as retinal disease databases or ophthalmology hospital clinical imaging libraries. These datasets contain color fundus images at various degrees of myopia, with image resolutions of, for example, 2048 x 2048 pixels. Each image includes a complete visual field of key anatomical structures such as the optic disc, macula, and vascular arch. Uniform imaging standards were maintained during image acquisition, employing standardized color correction and white balance processing. For example, gamma values ​​were adjusted within the range of 1.8 to 2.2 to ensure image quality met the input requirements of the deep learning model. Specifically, for example, 5000 myopic fundus images were randomly selected from the dataset as training samples. These samples covered patients of different genders, ages, and degrees of myopia, ranging from -1.0D to -10.0D, to ensure the diversity of the training data. Image preprocessing included size normalization and noise filtering, such as using Gaussian filtering to remove image noise, ensuring the consistency of the input data. Lossless compression, such as PNG format, was used for image storage to avoid data loss, while metadata records the acquisition device and timestamp for easy traceability and management. During implementation, image quality assessment was also considered. For example, image sharpness and contrast metrics were calculated to ensure that only images meeting quality standards were selected. The sharpness threshold was set to an edge gradient value greater than 50. This threshold was based on historical image quality analysis data. The gradient magnitude of edge pixels in the image was calculated and compared with the sharpness threshold; if the average gradient magnitude was greater than the sharpness threshold, the image was considered sharp enough. The contrast threshold was set to a grayscale standard deviation greater than 30. This threshold was based on visual perception experimental data. The standard deviation of image grayscale values ​​was calculated and compared with the contrast threshold; if the standard deviation was greater than the contrast threshold, the contrast was considered sufficient. All threshold settings were validated through multiple experiments, such as using cross-validation to adjust the threshold range, to ensure adaptability to different acquisition conditions.

[0021] Obtaining standard spatial relationship information for key anatomical components based on medical anatomical atlases is achieved by parsing standard ophthalmic anatomical atlases. For example, standard coordinate data for key anatomical components such as the optic disc center, fovea centralis, and vascular arch branch points are extracted from digital anatomical atlases. Specifically, spatial relationship parameters are extracted from the atlas, including a straight-line distance range of 4.5 to 5.5 mm from the optic disc center to the fovea centralis, and a distance range of 3.2 to 4.1 mm from the optic disc center to the farthest point of the vascular arch. These parameters are set based on clinical measurement statistics, for example, by collecting the mean and subtracting the standard deviation from 100 normal eye samples. The mean is set to 5.0 mm, and the standard deviation to 0.5 mm, thus setting the distance range to 4.5 to 5.5 mm. These parameters constitute standard spatial relationship information, stored as vectors in a configuration file for subsequent steps. During implementation, differences in ocular anatomy among different ethnic groups were considered. For example, the fluctuation range of standard values ​​was adjusted to adapt to population characteristics. The optic disc-macular distance (ODD) in Asian populations may have an increased offset of 0.2 mm. This offset was set based on multicenter clinical study data and determined by comparing the mean differences between samples from different ethnic groups. The spatial relationship information was also set with reference to statistical results in medical literature, such as using regression analysis to determine the correlation between distance and axial length, ensuring that the parameters were within a clinically reasonable range. All spatial relationship data were standardized in format before storage, for example, using JSON format for easy subsequent parsing and use. Parameter comparison was achieved by calculating the deviation of actual measurements from the standard range; for example, if the optic disc-macular distance falls within 4.5 to 5.5 mm, it is considered to meet the standard.

[0022] Obtaining topological connectivity information for key anatomical components based on medical anatomical atlases is achieved by analyzing the connection relationships between various parts of the anatomical structure. For example, based on the course of retinal vessels, a tree-like topological structure is established with the optic disc as the center and the vascular arches as branches. This structure clearly defines the connection relationships between the optic disc and the superior branch vessels, the optic disc and the inferior branch vessels, and the macula and the vascular arches. Topological connectivity information is represented in the form of an adjacency matrix, where a value of 1 indicates a direct connection between two anatomical components, and 0 indicates no direct connection. For example, the connection value between the optic disc and the superior branch vessels is 1, while the connection value between the optic disc and the macula is 0. During implementation, the range of vascular branch angles is also considered. For example, the angle between major vascular branches is limited to between 45 and 90 degrees. This range is set based on ophthalmological expert consensus and clinical observation data, with the minimum value of 45 degrees and the maximum value of 90 degrees determined by collecting expert questionnaires and measuring the angle distribution in clinical images. The acquisition of topological connectivity information also includes path consistency checks, such as calculating the consistency of distances between adjacent nodes in the path to ensure that the connection logic conforms to anatomical patterns. All topological data is stored in matrix format, such as CSV files, for easy subsequent calculations and processing. Connection verification rules are also set, such as requiring each node to have at least one connection point to avoid isolated structures. Angle comparisons are achieved by measuring the actual angles of blood vessel branches and comparing them to a range; for example, angles within 45 to 90 degrees are considered topologically coherent.

[0023] The medical anatomical atlas contains predefined anatomical structure data compiled by integrating multiple authoritative ophthalmic anatomy works, such as atlases of fundus anatomy and professional literature on the retinal vascular system. These atlases detail the standard dimensions, relative positions, and connections of various retinal anatomical structures. During implementation, the textual descriptions and diagrams in the literature were converted into digital parameters. For example, by scanning charts and extracting coordinate points, a standard dataset containing a coordinate reference system, scale standards, and connection rules was established. This dataset also includes the range of anatomical variations in patients of different age groups. For instance, the macula's position relative to the optic disc center is 0.3 to 0.5 mm smaller in adolescent patients than in adult patients. This offset is based on age group statistics, and the range is determined by measuring the mean offset of samples from different age groups. All anatomical structure data has been independently validated by three attending ophthalmologists. For example, clinical accuracy was ensured by comparing with clinical cases. Validation criteria included a positional error threshold of less than 0.1 mm. This threshold was set based on measurement accuracy requirements. Validation was considered successful if the distance between the labeled position and the actual position was calculated and compared with the positional error threshold; if the distance was less than the threshold, the validation was considered successful. The data update mechanism is based on new clinical studies, such as annual reviews, to ensure synchronization with the latest medical knowledge, while version control records change history for easy traceability and adjustment.

[0024] In the specific implementation process, multiple myopic fundus images are correlated with prior anatomical knowledge data. For example, each fundus image is associated with a configuration file containing standard spatial relationship information and topological connectivity information corresponding to that image. Image hash values ​​are used as unique identifiers during implementation, for example, using the MD5 algorithm to generate hash values, ensuring accurate matching between images and prior data. For special cases, such as fundus images of highly myopic patients, the standard spatial relationship information is adaptively adjusted according to the axial length. For example, the adjustment coefficient is obtained through regression analysis of clinical measurement data; for every 1 mm increase in axial length, the optic disc-macular distance increases by 0.15 mm. This coefficient is based on a linear regression model, and the slope is determined by fitting data points of axial length and distance. All prior knowledge data undergoes consistency checks before use. For example, correlation coefficients from different sources are calculated to ensure consistency in dimensions and precision. The correlation coefficient threshold is set to be greater than 0.9, based on statistical significance requirements. The Pearson correlation coefficient is calculated and compared with the threshold; if the correlation coefficient is greater than the threshold, the data is considered consistent. Data management also includes backup and recovery mechanisms, such as regularly storing data in multiple locations to prevent data loss, while access control ensures that only authorized personnel can modify the data, maintaining its integrity and security. The entire acquisition process emphasizes repeatability, for example, using scripts to automate data extraction and verification, reducing human error and improving implementation efficiency. Threshold comparisons are all implemented programmatically, for example, by embedding conditional statements in the verification scripts to ensure that all checks are executed automatically.

[0025] S2. Use a deep learning model to extract multi-scale features from myopic fundus images to obtain multi-scale feature maps. The specific implementation is as follows: When extracting low-level features from myopic fundus images using convolutional layers of a deep learning model, a convolutional neural network architecture is used to process the input myopic fundus images. These images are derived from multiple myopic fundus images obtained in the preceding steps, with a resolution of, for example, 2048×2048 pixels. Each convolutional layer contains multiple convolutional kernels, with a kernel size of, for example, 3×3 pixels and a stride of 1 pixel. Zero padding is used to maintain the feature map size. The number of convolutional kernels gradually increases with network depth; for example, the first convolutional layer uses 64 kernels, and the second uses 128 kernels. Each convolutional layer is followed by an activation function, such as the ReLU function, to introduce a non-linear transformation and avoid the vanishing gradient problem. Low-level feature extraction is achieved through convolution operations, i.e., a sliding window calculation is performed on the input image. The pixel values ​​within each window are multiplied by the convolutional kernel weights, summed, and then a bias term is added before the feature map is output through the activation function. The convolutional kernel weights are initialized using the Xavier method, with the bias initialized to zero. During training, the weights are updated using the gradient descent algorithm. Feature map size calculation is based on input size, kernel size, stride, and padding parameters. For example, if the input size is 2048×2048 pixels, the output feature map size after convolution will be 2048×2048 pixels. Low-level features contain basic visual information such as edges, texture, and color, which serves as input for subsequent multi-scale feature fusion.

[0026] When inputting low-level features into different layers of a deep learning model for multi-scale feature fusion, the deep learning model employs an encoder structure containing multiple layers, such as shallow, mid-level, and deep layers, each corresponding to a different level of feature abstraction. Shallow-level feature maps have higher spatial resolution, such as 512×512 pixels, containing more detailed information; deep-level feature maps have lower spatial resolution, such as 64×64 pixels, containing more semantic information. Multi-scale feature fusion is achieved through a feature pyramid structure, which involves upsampling deep feature maps to the size of shallow feature maps and then fusing them. The upsampling method uses bilinear interpolation; for example, upsampling a 64×64 pixel feature map to 512×512 pixels, bilinear interpolation calculates the new pixel value based on a weighted average of adjacent pixel values. The fusion operation uses element-wise addition or channel concatenation; for example, the upsampled deep feature map and the shallow feature map are concatenated along the channel dimension to form a multi-channel fused feature. The fusion weights are dynamically adjusted based on feature importance. For example, the weights of each feature map are calculated using an attention mechanism, with the attention weights adaptively generated based on the feature map content, implemented through fully connected layers and the Softmax function. During the fusion process, the number of channels in the feature maps is uniformly adjusted; for example, using 1×1 convolutions reduces the number of channels to 256 to reduce computational complexity. Multi-scale feature fusion ensures the complementarity of information from different levels, enhancing the model's ability to perceive multi-scale anatomical structures.

[0027] When a deep learning model outputs multi-scale feature maps corresponding to different scales of features from different layers, the output feature maps come from different fused layers, with each layer outputting one feature map. These feature maps have different spatial resolutions and number of channels. For example, shallow-level output feature maps have a resolution of 512×512 pixels and 256 channels; mid-level output feature maps have a resolution of 256×256 pixels and 512 channels; and deep-level output feature maps have a resolution of 128×128 pixels and 1024 channels. The generation of multi-scale feature maps is achieved through model forward propagation, where the input image passes through convolutional layers, pooling layers, and fusion layers, and feature maps are extracted from a specified layer as output. The pooling layer uses max pooling, with a pooling window size of, for example, 2×2 pixels and a stride of 2 pixels, to reduce the feature map size while preserving salient features. The scale of the output feature maps is selected based on the requirements of subsequent tasks; for example, high-resolution feature maps are used to locate key anatomical parts, while low-resolution feature maps are used to classify overall structures. Multi-scale feature maps are stored in a multi-dimensional array format, such as using floating-point tensors, to facilitate geometric analysis and topological coherence evaluation in subsequent steps. Feature map normalization employs batch normalization, standardizing the mean and variance of each feature map channel to ensure numerical stability. The entire multi-scale feature extraction process optimizes parameters through iterative training, ultimately outputting stable and usable multi-scale feature maps for subsequent analysis.

[0028] S3. Based on standard spatial relationship information, the geometric shape of key anatomical components is parsed from the multi-scale feature map. The spatial constraint loss term is obtained by verifying the degree of conformity between the geometric shape and the standard geometric shape. The specific implementation is as follows: When extracting spatial location information of key anatomical components from multi-scale feature maps, a regression network structure is used to process the multi-scale feature maps from the preceding steps. This regression network contains fully connected layers and convolutional layers. The number of neurons in the fully connected layer is determined based on the number of key anatomical components. For example, when key anatomical components include the optic disc center, the fovea centralis, and the branch points of the vascular arch, the output dimension of the fully connected layer is set to 6, corresponding to the two-dimensional coordinate values ​​of the three key points. Spatial location information extraction is achieved through regression operations. The multi-scale feature map is input into the regression network, and after forward propagation, the coordinate values ​​of each key anatomical component are output, represented by pixel positions. For example, the coordinates of the optic disc center are output as 512 pixels on the x-axis and 512 pixels on the y-axis. The weights of the regression network are initialized using the He method, and the bias term is initialized to zero. During training, the network parameters are updated using the gradient backpropagation algorithm. Data normalization is applied during the extraction process, for example, the output coordinate values ​​are linearly transformed to the range of 0 to 1 to ensure numerical stability. The extracted spatial location information is stored in vector format, for example, using a floating-point array to record the coordinate sequence of each key point in sequence, which is convenient for use in subsequent geometric morphology construction steps.

[0029] When constructing the geometric shape of key anatomical components based on spatial location information, a point set connection method is used to combine the extracted key point coordinates into a complete geometric shape. Geometric shape construction is based on the spatial distribution relationship between key anatomical components. For example, connecting the three key points—optic disc center, fovea centralis, and vascular arch branch points—in sequence forms a triangular structure, which represents the spatial layout characteristics of the main anatomical components of the fundus. The construction process includes two steps: coordinate system transformation and shape generation. Coordinate system transformation converts the key point coordinates in the pixel coordinate system to the actual physical scale coordinate system. For example, based on the ratio between the image resolution of 2048×2048 pixels and the actual size of 20×20 mm, a transformation coefficient of 0.01 mm per pixel is set. Shape generation is achieved by connecting adjacent key points with straight lines. For example, connecting the optic disc center with the fovea centralis, the optic disc center with the vascular arch branch points, and the fovea centralis with the vascular arch branch points in sequence forms a closed triangular geometric shape. The geometric shape is represented using a polygonal model. For example, a vertex sequence is used to define the shape boundary, and the vertex coordinates are derived from the key point coordinates in the spatial location information. Smoothing techniques are applied during the construction process, such as using cubic spline interpolation to optimize the shape contour, ensuring geometric continuity and anatomical rationality. The generated geometry is stored in a specific data structure, such as using a list of point sets to record the coordinates of all vertices, for subsequent shape difference calculations.

[0030] When calculating the shape difference between the geometric morphology and the standard geometric morphology, the Euclidean distance metric is used to compare the deviations of the two geometric morphologies in keypoint coordinates. The shape difference calculation is based on keypoint coordinates obtained from regression of multi-scale feature maps and standard keypoint coordinates defined based on standard spatial relationship information. For example, if the standard spatial relationship information specifies that the distance from the optic disc center to the fovea is 5.0 mm, then the keypoint coordinates of the standard geometric morphology are calculated based on this distance and relative positional relationships. Euclidean distance is calculated by measuring the straight-line distance between corresponding keypoints. For example, the straight-line distance between the optic disc center coordinates in the geometric morphology and the optic disc center coordinates in the standard geometric morphology is calculated using the square root of the sum of the squares of the differences in each dimension between the two coordinate points. The shape difference is taken as the average of the Euclidean distances of all keypoints. For example, if the calculated Euclidean distances for three keypoints are 0.2 mm, 0.3 mm, and 0.1 mm, then the shape difference value is 0.2 mm. Scale normalization is applied during the calculation process, for example, dividing the Euclidean distance value by the physical length of the image diagonal. The normalization coefficient is determined based on the ratio of the image size to the actual size; for example, 2048 pixels correspond to 20 millimeters. The shape difference calculation result is stored as a floating-point scalar value, representing the degree of deviation between the two geometric shapes, which is used for the subsequent generation of the spatial constraint loss term.

[0031] When generating spatial constraint loss terms based on shape differences, a loss function is used to convert shape differences into constraints required for model training. Spatial constraint loss term generation is achieved through linear transformations, such as multiplying the shape difference by a weighting coefficient to obtain the loss value. The weighting coefficient is dynamically adjusted according to the training phase; for example, the weighting coefficient is set to 1.0 in the initial dominant training phase and 0.5 in the later fine-tuning phase. The loss function is designed as a mean squared error form; for example, the spatial constraint loss term equals the square of the shape difference value, amplifying the penalty for larger deviations through squaring. An upper limit for the shape difference threshold is set during generation, for example, setting it to 1.0 mm. When the shape difference exceeds this threshold, the spatial constraint loss term is fixed to a preset maximum value to avoid instability during training. The output of the spatial constraint loss term is a scalar value, ranging from 0 to 1, facilitating weighted fusion with the topological constraint loss term in subsequent steps. Loss term generation is based on the deviation of the model's predicted geometry from the standard geometry; for example, the larger the shape difference value, the higher the spatial constraint loss term value, thereby prompting the model to optimize the accuracy of spatial location prediction. Spatial constraint loss terms are stored as part of the training parameters, for example, by recording the change in loss value for each iteration in the training log, which is used to monitor the training process and model convergence.

[0032] S4. Based on topological connectivity information, evaluate the topological coherence of key anatomical paths represented by multi-scale feature maps in the feature space to obtain the topological constraint loss term. The specific implementation is as follows: When constructing the adjacency matrix of key anatomical paths based on topological connectivity information, the topological connectivity information comes from the anatomical structure data obtained from medical anatomical atlases in the preceding steps. Adjacency matrix construction is achieved through matrix initialization. The matrix dimension is determined based on the number of nodes in the key anatomical path. For example, when the key anatomical path includes optic disc nodes, macula nodes, and vascular arch nodes, the adjacency matrix is ​​a 3x3 square matrix. Matrix element values ​​are assigned based on the connection relationships recorded in the topological connectivity information. For example, when two nodes are directly connected, the corresponding matrix element value is 1; when there is no direct connection, the corresponding matrix element value is 0. The adjacency matrix is ​​stored in a two-dimensional array format for easy calculation of subsequent topological coherence metrics. During construction, a connection verification threshold is set, which is set as the number of connected components threshold. For example, when the number of connected components is greater than 1, the topological structure is considered discontinuous. This threshold is set based on the physiological structural characteristics of the retinal vascular network and is determined by statistically analyzing the distribution of connected components of the vascular network in normal fundus images.

[0033] When extracting feature vectors from nodes in key anatomical paths from a multi-scale feature map, the multi-scale feature map is derived from the multi-scale feature map extracted by the deep learning model in the preceding steps. Feature vector extraction is achieved through spatial pooling operations, such as global average pooling of the anatomical region corresponding to each node on the multi-scale feature map. The feature vector dimension is determined based on the number of channels in the multi-scale feature map; for example, when the number of channels in the multi-scale feature map is 256, the extracted feature vector dimension is 256. Feature vector normalization uses L2 normalization to ensure that the vector magnitude is 1. A feature quality threshold is set during the extraction process, which is set as the lower limit of the feature vector magnitude. For example, if the feature vector magnitude is less than 0.1, the feature quality is considered unqualified. This threshold is based on feature distribution statistics and is determined by calculating the percentile of the feature vector magnitude in the training set. Feature quality comparison is achieved by calculating the feature vector magnitude and comparing it with the feature quality threshold. If the magnitude is less than the feature quality threshold, the features are re-extracted.

[0034] When calculating the topological coherence metric for key anatomical paths based on adjacency matrices and eigenvectors, the topological coherence metric is achieved by calculating the cosine similarity between the eigenvectors of adjacent nodes. Based on the connection relationships determined by the adjacency matrix, all pairs of adjacent nodes are traversed and their cosine similarities are calculated. The cosine similarity calculation is performed using vector dot product and modulus operations; for example, the dot product of two eigenvectors is the sum of the products of the numerical values ​​of each dimension, and the modulus is the square root of the sum of the squares of the numerical values ​​of each dimension. The topological coherence metric is taken as the arithmetic mean of the cosine similarities of all pairs of adjacent nodes. An effective similarity threshold is set during the calculation process; this threshold is set as the lower limit of cosine similarity. For example, a connection is considered abnormal when the cosine similarity is below 0.3. This threshold is set based on node feature correlation analysis and determined by statistically analyzing the distribution pattern of adjacent node feature similarities in normal samples. Similarity comparison is achieved by comparing the calculated cosine similarity value with the effective similarity threshold. If the cosine similarity is lower than the effective similarity threshold, it is marked as an abnormal connection.

[0035] When generating the topological constraint loss term based on the topological coherence metric, the generation of the topological constraint loss term is achieved through a linear transformation. The loss function is designed to be inversely proportional; for example, the topological constraint loss term equals 1 minus the topological coherence metric. A lower limit threshold for the topological coherence metric is set, which is set to 0.6. When the topological coherence metric falls below this threshold, the loss term is weighted and amplified. The weights of the loss term are dynamically adjusted according to the training phase; for example, the weight coefficient is set to 2.0 during the fine-tuning training phase and to 1.0 during the dominant training phase. The weight coefficient setting is based on training effect evaluation, determined by analyzing the impact of the loss term on model convergence at different training phases. A loss truncation threshold is set during the loss term generation process, which is set to 2.0. When the topological constraint loss term exceeds this value, it is truncated to 2.0. This threshold is set based on training stability requirements and determined through experimental observation of the impact of the loss value on gradient updates. Threshold comparisons are achieved through program condition judgments. For example, conditional statements can be embedded in the training code. When the topology coherence metric is lower than the lower limit of the topology coherence metric threshold, a weighting coefficient is applied. When the topology constraint loss term exceeds the loss truncation threshold, a truncation operation is performed.

[0036] S5. Based on the model training phase, dynamically adjust the fusion strategy of spatial constraint loss term and topological constraint loss term, and sequentially execute the dominant training phase dominated by spatial constraint loss term and the fine-tuning training phase dominated by topological constraint loss term. The specific implementation is as follows: When the weight of the spatial constraint loss term is set to be greater than that of the topological constraint loss term during the dominant training phase, this weight setting is based on the training strategy of prioritizing the establishment of accurate spatial localization capabilities in the early stages of model training. For example, the weight coefficient of the spatial constraint loss term is set to 2.0, and the weight coefficient of the topological constraint loss term is set to 0.5. The specific values ​​of these two weight coefficients are determined through hyperparameter optimization methods. That is, during model training, a grid search strategy is used to test multiple weight combinations within a predetermined range, and the weight value that achieves the optimal spatial localization accuracy on the validation set is selected. The weight coefficients are applied through a weighted summation operation, i.e., the total loss function equals the spatial constraint loss term multiplied by its weight coefficient plus the topological constraint loss term multiplied by its weight coefficient. The duration of the dominant training phase is controlled by a training iteration threshold, which is set to, for example, 10,000 iterations. This threshold is determined based on the analysis of the model's loss convergence curve on the validation set. When the decrease in the value of the spatial constraint loss term on the validation set is less than 0.001 over 100 consecutive iterations, the model is considered to have reached preliminary convergence. During the primary training phase, the model's performance on the validation set is checked every 100 iterations. If the rate of decrease of the spatial constraint loss term is lower than a preset rate of decrease threshold, the weight coefficients are dynamically adjusted. The rate of decrease threshold is set to, for example, 0.1 decreases every thousand iterations.

[0037] During the fine-tuning training phase, when the weight of the topology constraint loss term is set to be greater than that of the spatial constraint loss term, the weight adjustment is based on the training objective of strengthening the topology preservation ability after the model has established basic spatial localization capabilities. For example, the weight coefficient of the topology constraint loss term is adjusted to 1.5, and the weight coefficient of the spatial constraint loss term is adjusted to 0.8. This weight adjustment is achieved through linear interpolation, that is, within 10 iterations, the weight coefficients gradually transition from those in the dominant training phase to those in the fine-tuning training phase in an arithmetic progression. The duration of the fine-tuning training phase is controlled by a threshold for the number of fine-tuning training iterations, which is set to, for example, 5000 iterations. This threshold is determined based on the convergence characteristics of the topology constraint loss term on the validation set. Fine-tuning is considered complete when the change in the topology constraint loss term is less than 0.0005 over 50 consecutive iterations. During the fine-tuning training phase, the improvement in the topological coherence metric is evaluated every 50 iterations. If the improvement is less than the topological coherence improvement threshold, the weight coefficient of the topological constraint loss term is appropriately increased. The topological coherence improvement threshold is set to, for example, an increase of 0.05 per thousand iterations.

[0038] When switching from the dominant training phase to the fine-tuning training phase based on the number of training iterations, the phase switch is achieved by comparing the current training iteration count with a phase switch iteration count threshold. This threshold is set, for example, to 10,000 iterations. This threshold is determined based on a comprehensive analysis of the model's performance on both the training and validation sets. The phase switch is triggered when the dominant training phase reaches the set number of iterations and the spatial constraint loss term on the validation set reaches a stable state. The phase switch process includes saving the current model parameters, resetting the optimizer state, and updating the loss function weight coefficients. Switching condition checks are implemented through training monitoring, recording the training state every 100 iterations. The phase switch process is automatically executed when the current iteration count is greater than or equal to the phase switch iteration count threshold. After the phase switch, the learning rate is adjusted to 0.1 times the initial learning rate to accommodate the parameter fine-tuning requirements of the fine-tuning training phase. The entire phase switch process ensures a smooth transition in model training, avoiding drastic fluctuations in the loss value that could affect the model's convergence stability. During training, a maximum iteration threshold is set as the training termination condition. The maximum iteration threshold is set to, for example, 20,000 times. This threshold is determined based on training resource constraints and model convergence experience. When this number of iterations is reached, the training process is terminated regardless of whether the model has fully converged.

[0039] S6. Based on the fusion strategy, the deep learning model is iteratively trained using the gradient backpropagation algorithm, adjusting the model parameters until convergence. The specific implementation is as follows: When combining spatial constraint loss terms and topological constraint loss terms into a comprehensive loss function according to the fusion strategy, the fusion strategy employs the dynamic weight adjustment method defined in the preceding steps. This dynamic weight adjustment method automatically adjusts the weights of the loss terms based on changes during model training. The comprehensive loss function is constructed through a weighted linear combination. Specifically, in each training iteration, the spatial constraint loss term is multiplied by its real-time weight coefficient, and the topological constraint loss term is multiplied by its real-time weight coefficient. The two weighted results are then summed to obtain the comprehensive loss function value. The weight coefficients are dynamically adjusted according to the training phase. For example, during the dominant training phase, the weight coefficient for the spatial constraint loss term is set to, for example, 2.0, and the weight coefficient for the topological constraint loss term is set to, for example, 0.5; during the fine-tuning training phase, the weight coefficient for the spatial constraint loss term is adjusted to, for example, 0.8, and the weight coefficient for the topological constraint loss term is adjusted to, for example, 1.5. The specific values ​​of these weight coefficients are determined through a hyperparameter search method. A grid search strategy is used to test multiple weight combinations within a predetermined range. For example, the weight coefficient search range is set to 0.1 to 3.0, and the step size is set to 0.1. The optimal weight value is selected by evaluating the model performance on the validation set. During the calculation of the comprehensive loss function, a valid range check for the loss value is implemented. When the comprehensive loss function value exceeds a reasonable range, numerical truncation is performed. The upper limit of the reasonable range is set, for example, 10.0, and the lower limit is set, for example, 0.0. These thresholds are determined based on the statistical distribution of loss values ​​during training, and reasonable boundaries are set by analyzing the distribution of loss values ​​in historical training records. The loss value range check is implemented through a comparison operation: the comprehensive loss function value is calculated and compared with the upper and lower limits of the reasonable range. If the comprehensive loss function value is greater than the upper limit of the reasonable range, it is set as the upper limit; if the comprehensive loss function value is less than the lower limit, it is set as the lower limit.

[0040] When calculating the gradient of the comprehensive loss function with respect to the parameters of the deep learning model, gradient calculation is implemented using the backpropagation algorithm. Specifically, starting from the output of the comprehensive loss function, the partial derivatives of the parameters of each layer are calculated sequentially along the reverse path of the computation graph. For convolutional layer parameters, the partial derivatives of the comprehensive loss function with respect to the kernel weights and biases are calculated; for fully connected layer parameters, the partial derivatives of the comprehensive loss function with respect to the weight matrix and bias vector are calculated. The partial derivative calculation adopts the chain rule, that is, it is passed layer by layer through the product of the gradient of the previous layer and the local gradient of the current layer. For example, for convolutional layers, the local gradient includes the partial derivatives of the input feature map and the kernel weights. Gradient clipping is applied during gradient calculation. When the gradient norm exceeds the gradient norm threshold, the gradient is scaled. The gradient norm threshold is set to, for example, 1.0. This threshold is determined based on training stability requirements and is set by observing the distribution of gradient magnitudes during training. The gradient values ​​are stored as tensors with the same dimensions as the model parameters for easy use in subsequent parameter updates. The specific steps of gradient calculation include forward propagation to cache intermediate calculation results, backpropagation to calculate gradients at each layer, and gradient aggregation to ensure the accuracy and efficiency of gradient calculation.

[0041] When updating deep learning model parameters based on gradients, the parameter update is implemented using an optimization algorithm. Specifically, the Adam optimizer is used, which combines momentum and an adaptive learning rate adjustment strategy. The parameter update formula includes the calculation of first-order moment estimates and second-order moment estimates. The exponential decay rate of the first-order moment estimate is set to, for example, 0.9, and the exponential decay rate of the second-order moment estimate is set to, for example, 0.999. These decay rate parameters are determined based on the optimizer's standard configuration and experimental verification. The learning rate is dynamically adjusted according to the training phase. During the dominant training phase, the learning rate is set to, for example, 0.001, and during the fine-tuning training phase, the learning rate is adjusted to, for example, 0.0001. The specific value of the learning rate is determined by a learning rate scheduling strategy, such as using cosine annealing to gradually adjust the learning rate during training. The parameter update steps include calculating the gradient, updating the moment estimates, correcting biases, and applying the parameter update. For example, in each iteration, the current gradient is calculated first, then the first-order and second-order moment estimates are updated, and finally, the parameter update amount is calculated based on the corrected moment estimates. After each parameter update, a validity check is performed on the model parameters. When an outlier is found, the parameter is reset. The outlier threshold is set, for example, a parameter absolute value greater than 100.0. This threshold is determined based on an analysis of the reasonable range of model parameter values. The parameter validity check is performed by iterating through all parameters and comparing them with the outlier threshold. If a parameter value exceeds the outlier threshold, it is reset to its initial value.

[0042] The parameter update process is repeated until the deep learning model training converges. Convergence is determined based on the change of the comprehensive loss function on the validation set. A maximum iteration threshold is set as a mandatory termination condition during training, for example, 20,000 iterations. This threshold is determined based on training resource constraints and model convergence experience. An early stopping strategy is used for convergence determination: the model is considered convergent when the comprehensive loss function fails to improve on the validation set for several consecutive iterations. The early stopping patience value is set, for example, 1,000 iterations. The improvement threshold is set, for example, when the decrease in the comprehensive loss function is less than 0.0001. These thresholds are set by analyzing the training curves and the performance changes on the validation set. Each training iteration includes four steps: forward propagation calculation, comprehensive loss function calculation, gradient backpropagation, and parameter update. For example, in each iteration, forward propagation is performed first to obtain the prediction result, then the comprehensive loss function is calculated, followed by gradient backpropagation, and finally, the model parameters are updated. During training, the change curves of the comprehensive loss function, spatial constraint loss term, and topological constraint loss term are recorded to monitor the training status and analyze model behavior. For example, the loss value is saved to a log file every 100 iterations. The final trained model is stored as a model parameter file for subsequent inference tasks. The model storage format uses, for example, Protocol Buffers to ensure compatibility and efficiency. The entire training process is controlled by automated scripts, such as using Python programming languages ​​to implement training loops and conditional checks, ensuring the repeatability and stability of the training process.

[0043] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0044] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0045] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0046] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0047] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0048] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0049] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0050] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0052] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 for training a deep learning recognition model for near-view images, characterized in that, include: S1. Acquire multiple myopic fundus images and corresponding prior knowledge data of anatomical structures. The prior knowledge data of anatomical structures includes standard spatial relationship information and topological connection information of key anatomical components. S2. Use a deep learning model to extract multi-scale features from myopic fundus images to obtain multi-scale feature maps; S3. Based on standard spatial relationship information, the geometric shape of key anatomical components is parsed from the multi-scale feature map, and the spatial constraint loss term is obtained by verifying the degree of conformity between the geometric shape and the standard geometric shape. S4. Based on topological connectivity information, evaluate the topological coherence of key anatomical paths represented by multi-scale feature maps in the feature space and obtain the topological constraint loss term. S5. Dynamically adjust the fusion strategy of spatial constraint loss term and topological constraint loss term according to the model training stage, and execute the dominant training stage dominated by spatial constraint loss term and the fine-tuning training stage dominated by topological constraint loss term in sequence. S6. Based on the fusion strategy, the deep learning model is iteratively trained using the gradient backpropagation algorithm, and the model parameters are adjusted until convergence.

2. The training method for a deep learning recognition model of near-view images according to claim 1, characterized in that, Multiple myopic fundus images and corresponding prior anatomical knowledge data were acquired. This prior anatomical knowledge data includes standard spatial relationship information and topological connectivity information of key anatomical components, including: Acquire multiple fundus images of nearsighted individuals; Obtain standard spatial relationship information of key anatomical components based on medical anatomical atlases; Topological connection information of key anatomical components is obtained based on medical anatomical atlases; The medical anatomy atlas contains predefined anatomical structure data.

3. The method for training a deep learning recognition model for near-view images according to claim 1, characterized in that, Deep learning models are used to extract multi-scale features from myopic fundus images, resulting in multi-scale feature maps, including: Low-level features of myopic fundus images are extracted using the convolutional layers of a deep learning model. Low-level features are input into different layers of a deep learning model for multi-scale feature fusion. The deep learning model outputs multi-scale feature maps corresponding to features at different scales from different levels.

4. The method for training a deep learning recognition model for near-view images according to claim 1, characterized in that, Based on standard spatial relationship information, the geometric shape of key anatomical components is parsed from multi-scale feature maps. The spatial constraint loss term is obtained by verifying the degree of conformity between the geometric shape and the standard geometric shape, including: Extracting spatial location information of key anatomical components from multi-scale feature maps; Construct the geometric shape of key anatomical components based on spatial location information; Computational geometry differs from standard geometry; Spatial constraint loss terms are generated based on shape differences.

5. The method for training a deep learning recognition model for near-view images according to claim 4, characterized in that, The shape difference between the computational geometry and the standard geometry includes: calculating the average Euclidean distance between key points as the shape difference by comparing the key point coordinates of the geometry with those of the standard geometry; where the key point coordinates are obtained by regression from the multi-scale feature map, and the key point coordinates of the standard geometry are defined based on standard spatial relationship information.

6. The method for training a deep learning recognition model for near-view images according to claim 1, characterized in that, Based on topological connectivity information, the topological coherence of key anatomical paths represented by multi-scale feature maps is evaluated in the feature space, yielding a topological constraint loss term, including: Construct an adjacency matrix for key anatomical paths based on topological connectivity information; Extract feature vectors of each node in the key anatomical path from the multi-scale feature map; Calculate the topological coherence measure of key anatomical paths based on adjacency matrix and eigenvectors. Generate a topology constraint loss term based on the topology coherence metric.

7. The method for training a deep learning recognition model for near-view images according to claim 6, characterized in that, The calculation of topological coherence measure for key anatomical paths based on adjacency matrix and eigenvectors includes: determining the connection relationship of nodes in key anatomical paths according to adjacency matrix, calculating the cosine similarity between eigenvectors of adjacent nodes, and taking the average of the cosine similarity of all adjacent nodes on the path as the topological coherence measure; wherein the eigenvectors are extracted from multi-scale feature maps.

8. The method for training a deep learning recognition model for near-view images according to claim 1, characterized in that, Based on the dynamic adjustment strategy of the fusion of spatial constraint loss and topological constraint loss during the model training phase, a primary training phase dominated by the spatial constraint loss and a fine-tuning training phase dominated by the topological constraint loss are executed sequentially, including: During the dominant training phase, the weight of the spatial constraint loss term is set to be greater than the weight of the topological constraint loss term; During the fine-tuning training phase, the weight of the topological constraint loss term is set to be greater than the weight of the spatial constraint loss term; Switch from the dominant training phase to the fine-tuning training phase based on the number of training iterations.

9. The method for training a deep learning recognition model for near-view images according to claim 1, characterized in that, Based on the fusion strategy, the deep learning model is iteratively trained using the gradient backpropagation algorithm, adjusting the model parameters until convergence, including: According to the fusion strategy, the spatial constraint loss term and the topological constraint loss term are combined into a comprehensive loss function with corresponding weights. Calculate the gradient of the comprehensive loss function with respect to the parameters of the deep learning model; Update the parameters of the deep learning model based on the gradient; Repeat the parameter update process until the deep learning model training converges.

10. The method for training a deep learning recognition model for near-view images according to claim 9, characterized in that, Calculating the gradient of the comprehensive loss function with respect to the parameters of the deep learning model involves: calculating the partial derivatives of the comprehensive loss function with respect to the parameters of each layer of the deep learning model using the chain rule to obtain the gradient; wherein the comprehensive loss function is a weighted combination of spatial constraint loss term and topological constraint loss term according to the fusion strategy.

Images