Voxel recognition method, system and device for choroidal neovascularization and medium

Abstract

The invention relates to a voxel recognition method, system and device for choroidal neovascularization and a medium. The method comprises the following steps: S1, acquiring a choroidal neovascularization data set; s2, a three-dimensional structure data set is constructed and preprocessed; stacking the two-dimensional cross section image sequences of different light spots into a three-dimensional structure; s3, constructing a model; s4, performing model training; s5, applying the model; and identifying and predicting the choroidal neovascularization data set by using the trained voxel segmentation model, and outputting a prediction result of the choroidal neovascularization for subsequent medical statistical analysis. According to the method, voxel recognition and analysis statistics time of choroidal neovascularization is remarkably shortened, the recognition precision of the model is improved, the resource cost of manpower and material resources is greatly reduced in large-scale clinical tests, the morphology and position of the lesion are analyzed more comprehensively, and misjudgment and area omission caused by human factors are reduced.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a voxel identification method, system, device and medium for choroidal neovascularization. Background Art

[0002] Choroidal neovascularization (CNV) refers to abnormal formation of new blood vessels in the choroid layer at the back of the eye. These new blood vessels usually grow through the retinal pigment epithelium (RPE) and invade the underlying retina, which may cause decreased vision, dark spots in the center of the visual field, and even blindness. CNV is a common complication of many eye diseases, especially age-related macular degeneration (AMD).

[0003] However, the diagnostic difficulties of CNV include at least the following points:

[0004] (1) Subtle changes are difficult to detect: CNV may only appear as minor structural changes in the early stages, which are very subtle and difficult to detect with the naked eye. This requires doctors to have a high degree of vigilance and experience to identify these early changes.

[0005] (2) Image noise and artifacts: When using imaging technology, the image may be affected by noise and artifacts. These interference factors may mask the actual lesions and make the identification of CNV more difficult.

[0006] (3) Complex anatomical structure: The anatomical structure of the retina and choroid is complex, and CNV may be hidden in these complex structures, increasing the difficulty of identification and positioning.

[0007] (4) Experience dependence: Identification of CNV often requires doctors to have rich clinical experience and professional knowledge. For novice doctors, lack of experience may lead to misdiagnosis or missed diagnosis.

[0008] (5) Diverse manifestations: CNVs have diverse morphologies and locations, which may manifest as different imaging features. This requires doctors to be able to comprehensively analyze various imaging features to accurately diagnose CNVs.

[0009] First, the traditional method relies on animal models for histopathological staining and slicing to obtain images for analysis. This method not only consumes a lot of manpower and material resources, but also takes a long time and cannot meet the needs of rapid diagnosis. In addition, the accuracy of the traditional method is limited by the resolution of the staining and slicing technology and the subjective judgment of the operator, resulting in limited repeatability and accuracy of the results.

[0010] Secondly, the traditional image processing method for identifying choroidal neovascularization relies on the analyst manually setting the color threshold through specific software (e.g., ImageJ, HALO) to segment each slice in the image sequence. After completing the segmentation of the entire sequence, the analyst needs to manually set the sequence slice interval and calculate the volume using the software. There are two main problems with this method: (1) The operation process is cumbersome, especially when processing a large number of images, such as 20 eye image sequences, each sequence contains an average of 15 images, and the analyst needs to adjust the threshold for 300 images one by one to identify the new choroidal vascular area, and then manually set the sequence slice interval and calculate the volume; (2) Small and unconnected CNV areas are easily missed, especially in slices at the edge of choroidal neovascularization, which are easily missed due to their small size and discontinuity.

[0011] Although the image segmentation method based on the 2D deep learning model can better and quickly identify the CNV area in each image and calculate the 3D volume using calculus methods, this method still has two problems: (1) the overall recognition accuracy is insufficient and cannot meet the clinical needs for high-precision diagnosis; (2) the accuracy is particularly inaccurate in the CNV recognition of slices related to the edge of the new choroidal blood vessels. This may be because the traditional method mainly relies on two-dimensional image recognition and ignores the information and features of another dimension, resulting in a decrease in recognition accuracy.

[0012] In summary, existing technologies have shortcomings in terms of accuracy, medical costs, efficiency and ease of operation of CNV volume identification. Summary of the invention

[0013] In view of the above-mentioned defects of the prior art, a first aspect of the present invention provides a voxel identification method for choroidal neovascularization, which comprises the following steps:

[0014] Step S1: Acquire a choroidal neovascularization dataset; perform optical coherence tomography on the eyes of experimental animals to acquire a two-dimensional cross-sectional image sequence of different eye spots of the experimental animals as the choroidal neovascularization dataset;

[0015] Step S2: constructing a three-dimensional structure data set and preprocessing it; stacking the two-dimensional cross-sectional image sequence of each different light spot into a three-dimensional structure to obtain a complete three-dimensional structure data set; dividing the three-dimensional data set into a training set and a test set;

[0016] Step S3: Model construction; constructing a voxel segmentation model based on a three-dimensional convolutional neural network; wherein the structure of the voxel segmentation model includes an input layer, a three-layer encoder structure, a three-layer decoder structure and an output layer that are linked in sequence; wherein the three-layer encoder structure includes a first encoder, a second encoder, and a third encoder that are linked in sequence, and the three-layer decoder structure includes a first decoder, a second decoder, and a third decoder that are linked in sequence, and the third encoder and the third decoder are connected through a bottleneck layer, and the first encoder and the first decoder, the second encoder and the second decoder, and the third encoder and the third decoder are all connected by skip connections respectively;

[0017] Step S4: model training; randomly selecting the image of the three-dimensional structure from the training set as a validation set, training based on the voxel segmentation model on the training set and the validation set, and using the test set to evaluate the generalization ability of the voxel segmentation model;

[0018] Step S5: model application; using the trained voxel segmentation model to identify and predict the choroidal neovascularization dataset, outputting the prediction results of the choroidal neovascularization for subsequent medical statistical analysis, wherein the medical statistical analysis at least includes calculating the area and volume of the choroidal neovascularization.

[0019] In the voxel identification method for choroidal neovascularization as described above, optionally, the step S2 specifically includes: stacking each image in the two-dimensional cross-sectional image sequence in a time series to obtain a three-dimensional Numpy array to effectively capture the structural information of each light spot that varies with depth.

[0020] In the voxel identification method for choroidal neovascularization as described above, optionally, each image in the two-dimensional cross-sectional image sequence is a grayscale image of size [h, w] pixels, wherein the height dimension h=330 and the width dimension w=770.

[0021] In the voxel identification method for choroidal neovascularization as described above, optionally, step S3 specifically includes the following steps:

[0022] Step S3.1: The three-dimensional structure data set is input into the input layer of the voxel segmentation model in the form of three-dimensional voxel bricks, and the size of each voxel brick is d×h×w×c, wherein the depth dimension d represents the size of the three-dimensional voxel brick along the depth direction, the height dimension h represents the size of the three-dimensional voxel brick along the height direction, the width dimension w represents the size of the three-dimensional voxel brick along the width direction, and the channel number c represents the number of channels of the three-dimensional voxel brick;

[0023] Step S3.2: The three-layer encoder structure maps the three-dimensional voxel bricks to the vector space, performs feature extraction and downsampling to capture the high-level semantic information of the image; the three-layer decoder structure identifies each voxel of the input through the extracted feature vector, and restores the spatial details and resolution of the image through feature reconstruction and upsampling; the bottleneck layer combines the high-level features extracted by the third encoder with the feature map in the third decoder; the jump connection directly transfers the features of the three-layer encoder structure to the corresponding stages of the three-layer decoder structure to fuse feature information of different scales;

[0024] Step S3.3: The output layer outputs a voxel category prediction map of the same size as the three-dimensional voxel brick of the input layer.

[0025] In the voxel recognition method for choroidal neovascularization as described above, optionally, each encoder and decoder of the three-layer encoder structure and the three-layer decoder structure adopts a sequentially ordered combination of GroupNorm3d+Conv3d+ReLU substructures, wherein the GroupNorm3d divides the input three-dimensional data of the three-dimensional voxel bricks into multiple groups, and normalizes the data in each group, and the normalized data is convolved with the Conv3d convolution kernel to extract three-dimensional image features, and after the convolution operation, the ReLU activation function is applied to convert linear features into nonlinear features. features; in each layer of the encoder, the GroupNorm3d is used to normalize the feature map output by each convolutional layer, the Conv3d is used to extract features in three-dimensional space, and the ReLU activation function is applied to the convolutional layer to enable the network to learn complex feature representations; in each layer of the decoder, the GroupNorm is used to normalize the feature map output by each convolutional layer, the Conv3d is used to gradually restore the spatial resolution of the feature map and reduce the number of channels of the feature map to match the dimension of the original input three-dimensional data, and the ReLU function is used to reconstruct the details and structure of the input data and restore the complexity of the input data.

[0026] In the voxel identification method for choroidal neovascularization as described above, optionally, the section indicators for evaluating the voxel segmentation model in step S4 include at least precision, recall rate and Dice coefficient, and the optimal model training parameters are selected by reverse deduction through the values ​​of the section indicators.

[0027] In the voxel recognition method for choroidal neovascularization as described above, optionally, the voxel segmentation model is an end-to-end 3D recognition algorithm.

[0028] In order to achieve the above object, the second aspect of the present invention provides a voxel recognition system for choroidal neovascularization, which comprises:

[0029] An optical coherence tomography device is used to obtain cross-sectional images of ocular spots of experimental animals as a choroidal neovascularization dataset;

[0030] A server, wherein the digital recognition platform of the server receives the choroidal neovascularization dataset from the optical coherence tomography device, identifies and predicts the choroidal neovascularization dataset, and outputs the prediction result of the choroidal neovascularization for subsequent medical statistical analysis, wherein the medical statistical analysis at least includes calculating the area and volume of the choroidal neovascularization.

[0031] In order to achieve the above-mentioned purpose, the third aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor runs the program, the steps of the voxel identification method for choroidal neovascularization as described in any one of the first aspects above are implemented.

[0032] In order to achieve the above-mentioned purpose, the fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions or a computer program, and when the computer-executable instructions or the computer program are processed and executed, the voxel identification method for choroidal neovascularization as described in any one of the first aspects above is implemented.

[0033] The voxel recognition method, system, device and medium for choroidal neovascularization provided by the present invention can significantly reduce the voxel recognition and analysis statistical time of choroidal neovascularization, greatly reduce the resource costs of manpower and material resources in large-scale clinical trials, identify subtle structural changes through an end-to-end 3D algorithm model, more comprehensively analyze the pathological morphology and location of choroidal neovascularization, improve the detection and recognition capabilities of early choroidal neovascularization, and reduce misjudgments caused by human factors and regional omissions caused by visual fatigue or lack of concentration.

[0034] In general, the voxel recognition method and system for choroidal neovascularization proposed in the present invention reduces the reliance on experience and improves the recognition accuracy of the model and the efficiency of clinical diagnosis. By combining deep learning and image processing technology, a 3D model is constructed to learn and recognize the characteristics of CNV, helping image data analysts to perform more accurate quantitative index measurement and analysis, providing data support for the screening of test products in non-clinical research, and assisting doctors in making more accurate diagnosis and treatment decisions.

[0035] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of an embodiment of a voxel identification method for choroidal neovascularization of the present invention;

[0037] Figure 2 It is a structural schematic diagram of a voxel segmentation model based on a three-dimensional convolutional neural network of the present invention;

[0038] Figure 3 yes Figure 1 Comparison chart of the model training curve and the traditional 2D method training curve;

[0039] Figure 4 It is a structural schematic diagram of an embodiment of a voxel recognition system for choroidal neovascularization of the present invention. DETAILED DESCRIPTION

[0040] In order to make the technical means, creative features, objectives and effects of the invention easier to understand, the invention is further described below with reference to specific diagrams. However, the invention is not limited to the following implementation cases.

[0041] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0042] In addition, the terms "first", "second" and "third" are used only for descriptive purposes to distinguish one entity or operation from another entity or operation, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated or any such actual relationship or order between these entities or operations. For any single technical feature described or implied in the embodiments mentioned in this article, or any single technical feature displayed or implied in the drawings, the present invention still allows any combination or deletion between these technical features (or their equivalents) without any technical obstacles, so it should be considered that these more embodiments according to the present invention are also within the scope of the description of this article.

[0043] Terms such as “comprise” and “include” indicate that in addition to the components directly and explicitly stated in the specification and claims, the technical solution of the present invention does not exclude the situation where it has other components that are not directly or explicitly stated.

[0044] Figure 1 It is a flow chart of an embodiment of a voxel identification method for choroidal neovascularization of the present invention.

[0045] like Figure 1 As shown, the voxel recognition method may include the following steps:

[0046] Step S1: Acquire a choroidal neovascularization dataset, specifically a choroidal neovascularization two-dimensional tomographic image dataset.

[0047] In this step, the operator can use the optical tomography equipment to perform optical coherence tomography on the laser photocoagulation modeling spot area of ​​the experimental animal's eye. For each eye's different spots, a set of serialized two-dimensional cross-sectional images, i.e., full-volume images, can be obtained. Finally, a two-dimensional cross-sectional image sequence of different eye spots of different experimental animals is obtained as a choroidal neovascularization data set.

[0048] In the present invention, optical coherence tomography can provide reliable data support and key information for the subsequent training of deep learning models based on 3D recognition analysis of choroidal neovascularization and effect comparison. Specifically, optical coherence tomography is a non-invasive imaging technology that avoids harm to experimental animals and can be safely repeated. Its principle is to reconstruct the cross-sectional image of the tissue by measuring the interference signal of the reflected light, and to use a low-coherence light source to perform high-resolution imaging of biological tissues. The provided two-dimensional cross-sectional image can at least include details of the choroid layer and the retina. Therefore, the present invention uses optical coherence tomography to obtain the choroidal neovascularization data set, which has the advantages of being fast, non-invasive, and having high resolution, and can provide microstructural information of the tissue.

[0049] A two-dimensional cross-sectional image sequence refers to a series of two-dimensional images continuously acquired along a specific direction (such as an axis perpendicular to the region of interest, in this embodiment, the z-axis), which together constitute a three-dimensional data set that can provide a detailed view of the internal structure of the eye tissue. These images can be arranged in sequence and reorganized into a three-dimensional model by computer software, and each cross-sectional image can be analyzed to identify abnormal areas of tissue structure layer by layer (choroidal neovascularization). In this embodiment, the image obtained by optical coherence tomography is presented in the form of a grayscale image.

[0050] Step S2: Construct and preprocess a three-dimensional structure dataset. In this step, the two-dimensional cross-sectional image sequence of each different light spot is stacked into a three-dimensional structure to obtain a complete three-dimensional structure dataset. In addition, in order to avoid data leakage and ensure the accuracy of the results, in this embodiment, a light spot-based segmentation method is used to divide the three-dimensional dataset into a training set and a test set, that is, the full amount of images corresponding to each light spot are included in the training set or the test set respectively to form a complete image dataset, and the subsequent model training process performs image enhancement on the training set.

[0051] Data Dimensions Training set Validation set Test Set Total Light spot 80 - 20 100 picture 1241 138 401 1780

[0052] Table 1

[0053] Optionally, as shown in Table 1, the training set and the test set are divided into a ratio of 8:2 in the spot dimension, and the spots in the training set contain a total of 1241+138 images (i.e., the full set of images of each spot), and then a subset of the training set is randomly used as a validation set.

[0054] The training set is used to train the model. It iteratively optimizes the model parameters to minimize the loss function and is the main data source for model learning. The validation set is used for model selection and hyperparameter adjustment. It evaluates the effect of the model during the training process and helps prevent the model from overfitting on the training set. The test set is used to finally evaluate the performance of the model. The test set is completely independent of the training process and is finally used to evaluate the generalization ability of the model.

[0055] In an optional embodiment, step S2 may specifically include: stacking each image in the two-dimensional cross-sectional image sequence in a time series to obtain a three-dimensional Numpy array to effectively capture the structural information of each light spot that varies with depth.

[0056] In this step, three-dimensional training data is constructed for the two-dimensional cross-sectional image sequence of each light spot so as to be better used for model training and analysis. Each image in the two-dimensional cross-sectional image sequence is a grayscale image of size [h, w] pixels, wherein, in this embodiment, the height dimension h = 330 and the width dimension w = 770 of each image acquired by the optical coherence tomography device, that is, the size of the two-dimensional cross-sectional image sequence is [330, 770] pixels. By stacking the d two-dimensional cross-sectional image sequences corresponding to each light spot in a time series order, a three-dimensional Numpy array of size [d, 330, 770] can be generated. The depth dimension d represents the size of the three-dimensional data along the depth direction, and d in this embodiment is the number of stacked images. This three-dimensional array can effectively capture the structural information of each light spot that changes with depth, thereby providing rich features for subsequent deep learning models.

[0057] In an optional embodiment, a sequence of 25 consecutive two-dimensional cross-sectional images (optical coherence tomography is presented in the form of grayscale images) is obtained for the light spot A. A three-dimensional Numpy array of size [25, 330, 770] is obtained by stacking the two-dimensional cross-sectional image sequence corresponding to the light spot A. Figure 2 In the illustrated embodiment, 16 two-dimensional cross-sectional images of the light spot are selected for stacking, and in this case d=16, and the size of the three-dimensional Numpy array is [16, 330, 770].

[0058] It should be noted here that Numpy array is a very important data structure in Python programming language, which is provided by NumPy library. NumPy (Numeric Python) is an open source Python scientific computing library for large-scale numerical and matrix operations. Numpy arrays are stored contiguously in memory, which makes them faster than Python native lists when performing numerical calculations. NumPy library provides a large number of mathematical functions that can perform various mathematical operations on arrays, such as statistics, algebra, Fourier transform, random simulation, etc. In addition, Numpy arrays are highly compatible and can be used by many other scientific computing and data analysis libraries, such as SciPy, Pandas, Matplotlib, etc.

[0059] Step S3: Model construction.

[0060] In this step, if Figure 2 As shown, a voxel segmentation model is constructed based on a three-dimensional convolutional neural network. The structure of the voxel segmentation model may specifically include an input layer, a three-layer encoder structure, a three-layer decoder structure, and an output layer that are linked in sequence. The three-layer encoder structure may include a first encoder, a second encoder, and a third encoder that are linked in sequence, and the three-layer decoder structure may include a first decoder, a second decoder, and a third decoder that are linked in sequence, and in this embodiment, the third encoder and the third decoder are connected through a bottleneck layer, and the first encoder and the first decoder, the second encoder and the second decoder, and the third encoder and the third decoder are all connected by jump connections.

[0061] In an optional embodiment, step S3 may specifically include the following steps:

[0062] Step S3.1: The three-dimensional structure dataset is input into the input layer of the voxel segmentation model in the form of three-dimensional voxel bricks. The size of each voxel brick is d×h×w×c, wherein the depth dimension d represents the size of the three-dimensional voxel brick along the depth direction, the height dimension h represents the size of the three-dimensional voxel brick along the height direction, the width dimension w represents the size of the three-dimensional voxel brick along the width direction, and the number of channels c represents the number of channels of the three-dimensional voxel brick.

[0063] The "voxel brick" represents a three-dimensional image data block formed by stacking the two-dimensional cross-sectional image sequence of the light spot in step S2. In a three-dimensional image, a voxel is the basic unit that constitutes the image. Specifically, Figure 2 In the embodiment shown, the voxel brick size of the input layer is 16×330×770×1, that is, d=16, h=330, w=770, c=1. As described in the previous embodiment, d=16 means that the input is stacked using 16 two-dimensional cross-sectional images, and at this time d=16, the size of the three-dimensional Numpy array is [16, 330, 770]. The height dimension of each image acquired by the optical coherence tomography device is h=330, and the width dimension is w=770. The grayscale image acquired by the optical coherence tomography device is a single channel, so c=1.

[0064] Step S3.2: The three-layer encoder structure maps the three-dimensional voxel bricks to the vector space, performs feature extraction and downsampling to capture the high-level semantic information of the image. The three-layer decoder structure identifies each voxel of the input through the extracted feature vector, and restores the spatial details and resolution of the image through feature reconstruction and upsampling. The bottleneck layer combines the high-level features extracted by the third encoder with the feature map in the third decoder, and the jump connection directly passes the features of the three-layer encoder to the corresponding stages of the three-layer decoder to fuse feature information of different scales.

[0065] Specifically, each encoder includes two CRG (Contextual Residual Gating) layers to enhance feature extraction capabilities and improve the model's understanding of input data and feature expression capabilities. The first CRG layer of each encoder is responsible for capturing local low-level features in the input voxel brick data, such as basic information such as edges and textures, while the second CRG layer of each encoder further processes these features, extracting or integrating more complex patterns and contextual information to provide richer feature representations for subsequent encoders in other layers. The three-layer encoder structure gradually reduces the spatial resolution (h×w) of the input data while increasing the number of feature channels c. This downsampling process helps the model capture more abstract feature representations. At the same time, the output size of each layer of encoder is different, which helps to compress information in the network, remove redundant information, and retain the most important features. In short, as the network level deepens, each layer of encoder will further reduce the spatial size of the feature map to extract higher-level features. These hierarchical feature extractions are also the key to the deep learning model in the present invention to process complex data.

[0066] The bottleneck layer is the deepest part of the network and is used to capture and integrate high-level features from the encoder and provide enough information for the decoder to reconstruct the output. Figure 2The bottleneck layer shown uses two CRG layers to further extract and integrate features, helping the model extract deeper features from the output data of the third encoder. Also, the additional CRG layer can improve the robustness of the model to outliers and noise because it provides more parameters to learn and adapt to changes in the data. In deep networks, gradient disappearance is a common problem. Adding CRG layers can help alleviate this problem because each CRG layer contains residual connections, which helps the gradient flow in the network.

[0067] Each layer of the three-layer decoder structure gradually restores the spatial resolution of the image until the last layer (the first decoder) outputs voxel bricks with the same resolution and size as the input.

[0068] Although the output spatial resolution (h×w) of the three-layer decoder structure is increasing, the number of feature channels c and the depth will be reduced to maintain the balance of calculation. This upsampling operation will cause some spatial information to be lost. The present invention uses jump connections between the three-layer encoder structure and the three-layer decoder structure, that is, each layer of the decoder will receive the feature map of the jump connection from the corresponding layer of the encoder and fuse it with it to restore the lost spatial information, retain image details and improve segmentation accuracy.

[0069] In an optional embodiment, each encoder and decoder in the three-layer encoder structure and the three-layer decoder structure adopts a sequentially ordered combination of GroupNorm3d+Conv3d+ReLU substructures. GroupNorm3d divides the three-dimensional data of the three-dimensional voxel bricks on the input side of each encoder or decoder into multiple groups, and normalizes the data in each group. The normalized data is convolved through the Conv3d convolution kernel to extract three-dimensional image features. After the convolution operation, the ReLU activation function is applied to convert linear features into nonlinear features. In each encoder layer, GroupNorm3d is used to normalize the feature map output by each convolution layer, Conv3d is used to extract features in three-dimensional space, and the ReLU activation function is applied to the convolution layer so that the network can learn complex feature representations; in each decoder layer, GroupNorm is used to normalize the feature map output by each convolution layer, Conv3d is used to gradually restore the spatial resolution of the feature map and reduce the number of channels of the feature map to match the dimension of the original input three-dimensional data, and the ReLU function is used to reconstruct the details and structure of the input data and restore the complexity of the input data. In summary, by following the order of GroupNorm3d, Conv3d and ReLU, the model can be trained more stably, learn more effective features, have stronger expression ability, and perform well in various three-dimensional data processing tasks. The nonlinear activation function in the CRG layer of the present invention uses ReLU to increase the nonlinear expression ability of the model, so that the model can capture more complex features and patterns.

[0070] Step S3.3: The output layer outputs a voxel class prediction map of the same size as the 3D voxel bricks of the input layer.

[0071] After the above steps, the voxel segmentation model built based on the three-dimensional convolutional neural network can convert the fused feature map into the final category prediction map for output. The output of this model is the voxel category prediction of the same size as the input, that is, predicting whether each voxel is background or foreground. The subsequent volume statistics and area calculations of choroidal neovascularization can be statistically analyzed through the voxel-level prediction results.

[0072] Step S4: Model training.

[0073] In this step, images of three-dimensional structures are randomly selected from the training set as the validation set, and training is performed based on the voxel segmentation model on the training set and the validation set, and the generalization ability of the voxel segmentation model is evaluated using the test set. The voxel segmentation model of the present invention is an end-to-end 3D recognition algorithm, and the input when training the model is three-dimensional data, which is completely different from traditional 2D model training.

[0074] During the model training process, the present invention performs image enhancement on the training set to increase the diversity of training data, wherein the image enhancement process may include shape enhancement and input enhancement.

[0075] Specifically, shape enhancement can include affine transformation, nearest neighbor interpolation, and flipping. By simulating various transformations that images may encounter in the real world, the diversity of model training data is increased, and the robustness of the model to different perspectives and posture changes is improved. For example, when training a deep learning model, the generalization ability of the model is enhanced by randomly applying affine transformations to simulate the movement of the camera or the rotation of the target. Input enhancement can include Gaussian blur, median blur, additive Gaussian noise, and brightness adjustment.

[0076] In an optional embodiment, shape enhancement is used to increase the image diversity of the training set. With a probability of 30%, a set of affine transformations are applied: Image scaling: Scale to between 80% and 120% of its original size along the x-axis and y-axis respectively. Translation: Translate between -1% and 1% of its size along the x-axis and y-axis respectively. Shearing: Shear transformation between -5 degrees and 5 degrees. Rotation: Rotate between -179 degrees and 179 degrees. Use nearest neighbor interpolation (order=0) to ensure speed and quality of image processing. Horizontal flip: Flip the image horizontally with a probability of 50%. Vertical flip: Flip the image vertically with a probability of 50%.

[0077] In an optional embodiment, input enhancement uses one of the following operations with a probability of 15%: Gaussian blur: custom function, maximum kernel size is 3. Median blur: custom function, maximum kernel size is 3. Additive Gaussian noise: mean is 0, standard deviation range is between 0 and 0.05, noise is added per channel with a probability of 50%. Brightness adjustment: custom function, range is between -15 and 15.

[0078] "Using probability" refers to introducing randomness when performing operations or decisions, and this randomness is controlled by probability distribution. In this embodiment, the training data is enhanced by randomly applying various transformations (such as rotation, scaling, shearing, etc.), which are applied with a certain probability. During the model training stage, the training set is enhanced in the manner described above, aiming to improve the robustness and generalization ability of the model through various operations.

[0079] In an optional embodiment, the section indicators for evaluating the effect of the voxel segmentation model in step S4 at least include precision, recall and Dice coefficient, and the best model training parameters are selected by reverse deduction through the values ​​of the section indicators.

[0080] As for the model effect, the two-dimensional model trained by the 2D method and the three-dimensional voxel segmentation model provided by the present invention respectively identify and predict choroidal neovascularization on the same independent test set. The dimensionality reduction effect of the traditional 2D model prediction effect and the 3D model prediction effect of the present invention are shown in the following Table 2:

[0081] index 3D method (present invention) 2D method (traditional) Precision 92％ 83％ Recall 94％ 93％ Dice 93％ 87％

[0082] Table 2

[0083] Precision is the ratio of true positives (TP) to all samples predicted as positive, which is used to measure the accuracy of the model in predicting positive samples. Recall is the ratio of true positives to all actual positive samples, which is used to measure the ability of the model to identify all positive samples. The Dice coefficient measures the similarity between the predicted segmentation result and the true segmentation result by comparing the intersection and union of the two.

[0084] It can be seen from the index table that the precision and Dice of the 3D method of the present invention on a fixed data set are significantly higher than those of the 2D method. This shows that introducing 3D features and using 3D models for feature recognition can greatly improve the recognition effect.

[0085] The present invention reversely infers the optimal model parameters through the optimal state of the above-mentioned section indicators when adjusting the model parameters, and directly obtains the final result through a comprehensive model training. In an optional embodiment, the steps of adaptively adjusting the model parameters are as follows:

[0086] (1) Adjust the learning rate

[0087] The learning rate is a hyperparameter used in the optimization algorithm to adjust the model weight step size, which directly affects the convergence speed and stability of model training.

[0088] Observation indicators: monitor the loss value during training and the cross-sectional indicators on the validation set (such as the Dice coefficient, etc.).

[0089] Strategy: If the performance on the validation set stagnates after an initial rapid improvement, you may need to reduce the learning rate to refine the weight adjustment; if the improvement is slow, you may need to increase the learning rate

[0090] In this embodiment, the initial learning rate of the present invention is set to 2e-4, that is, 2×10 -4 .

[0091] (2) Weight Decay

[0092] Weight decay is a regularization technique used to prevent model overfitting by adding a penalty term proportional to the square of the weight to the loss function.

[0093] Metrics to watch: Check for signs of overfitting during training, such as training loss continuing to decrease while validation loss starts to increase.

[0094] Strategy: Increasing weight decay can help reduce overfitting, but too much weight decay may lead to underfitting. Find the balance point by adjusting the value of weight decay.

[0095] In this embodiment, the weight attenuation coefficient of the present invention is set to 1×10 -5 .

[0096] (3) Batch Size

[0097] Observation indicators: Evaluate the impact of different batch sizes on training stability and convergence speed.

[0098] Strategy: A larger batch size may speed up training, but it may also cause the model to get stuck in a poor local minimum. Experiment with different batch sizes to find the best balance.

[0099] (4) Learning Rate Scheduler

[0100] The benefit of using a learning rate scheduler in the present invention is that it can automatically adjust the learning rate according to the performance of the model on the validation set, which helps to select the best learning rate during the training process, avoids converging to a non-optimal solution too quickly, or oscillations caused by too high a learning rate in the late stage of training. This adaptive adjustment mechanism helps the model converge to better performance more stably during the training process.

[0101] Observation indicators: Use the segmentation indicators on the validation set to dynamically adjust the learning rate.

[0102] Strategy: When the performance on the validation set no longer improves, reduce the learning rate. Adjust the patience (how many epochs are allowed without improvement before changing the learning rate) and factor parameters (the multiplier for the learning rate decay) to find the best learning rate adjustment strategy.

[0103] In this embodiment, the present invention uses ReduceLROnPlateau as a learning rate scheduler, which reduces the learning rate when the performance on the validation set no longer improves. The factor is set to 0.2, which means that the learning rate will be multiplied by 0.2, that is, reduced to 20% of the original value.

[0104] (5) Data Augmentation

[0105] Observation metrics: Evaluate the impact of different data augmentation strategies on model performance.

[0106] Strategy: Increase or decrease the intensity and type of data augmentation and observe the impact on the performance of the validation set to determine the most effective data augmentation strategy.

[0107] In the present invention, the randomness introduced by using probability in the embodiment of step S2.2 above when performing data enhancement can best improve the robustness and generalization ability of the voxel segmentation model in this solution.

[0108] (6) Loss Function

[0109] The loss function is used to measure the difference between the model prediction and the true label, and the training goal is to minimize this loss value.

[0110] Observation metrics: Evaluate the impact of different loss functions on model performance.

[0111] Strategy: Try different loss functions, such as cross entropy loss, Dice loss, IoU loss, etc., and select the best loss function based on the slice index on the validation set.

[0112] like Figure 3 As shown, Figure 3 (a) is the model training curve of the traditional two-dimensional method. Figure 3 (b) is the model training curve of the three-dimensional method of the present invention. It can be seen that the convergence speed of the 3D model training provided by the present invention is significantly better than the model trained by the traditional two-dimensional method, which proves that the 3D voxel segmentation model of the present invention has stronger data fitting ability.

[0113] (7) Network architecture adjustment

[0114] Observe metrics: Monitor the impact of different network architectures on performance.

[0115] Strategy: Increase or decrease the number of network layers, change the size of convolution kernels, etc., and optimize the network architecture based on the cross-sectional indicators on the validation set.

[0116] In the present invention, the use of the network structure in step S3 above can achieve the best effect of the voxel segmentation model.

[0117] Step S5: Model application: Using the trained voxel segmentation model to identify and predict the choroidal neovascularization data set, output the prediction results of the choroidal neovascularization for subsequent medical statistical analysis, wherein the medical statistical analysis at least includes calculating the area and volume of the choroidal neovascularization.

[0118] In an optional embodiment, the volume calculation can be performed by calculating the total volume of the region of interest (choroidal neovascularization region) by multiplying the number of voxels belonging to the category in the statistical prediction map for each category by the volume of the voxel.

[0119] In other optional embodiments, medical staff can also use the output data of the 3D voxel segmentation model of the present invention to perform morphological analysis, connectivity analysis, statistical measurement, etc. For example, morphological operations such as opening and closing operations are performed on the prediction map to eliminate small holes and smooth boundaries to improve the segmentation results; or, connectivity analysis is performed on the output data to analyze the connected domains in the prediction map, identify independent structures or regions, and calculate the properties of each connected domain, such as volume, surface area and shape characteristics; or, various statistical measurements of the prediction map are calculated, including mean, median, standard deviation, etc., to describe the distribution characteristics of the data.

[0120] In order to achieve the above object, the present invention also provides a voxel recognition system for choroidal neovascularization. Figure 4 As shown, the system may specifically include an optical coherence tomography device and a server connected to each other. The optical coherence tomography device is used to obtain a cross-sectional image of a light spot in the eye of an experimental animal as a choroidal neovascularization data set, and a digital recognition platform of the server receives the choroidal neovascularization data set from the optical coherence tomography device, identifies and predicts the choroidal neovascularization data set, and outputs the prediction result of the choroidal neovascularization for subsequent medical statistical analysis, wherein the medical statistical analysis at least includes calculating the area and volume of the choroidal neovascularization.

[0121] In order to achieve the above-mentioned objectives, the present invention also provides a computer device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor runs the program, the steps of the voxel identification method for choroidal neovascularization as described in any of the aforementioned embodiments can be implemented.

[0122] The processor and the memory can be set separately or integrated together, for example, integrated on a system-on-chip (SOC) of a terminal device. It should be understood that the processor of the embodiment of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to execute, or a combination of hardware and software modules in the decoding processor to execute. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0123] In order to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, which stores executable instructions or programs. When the executable instructions or programs are processed and executed, the voxel identification method for choroidal neovascularization as described above is implemented.

[0124] The readable storage medium is, for example, a memory. The memory may be a volatile memory or a nonvolatile memory, or the memory may include both a volatile memory and a nonvolatile memory. Among them, the nonvolatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0125] If the integrated units in the above embodiments are implemented in the form of software functional units and sold or used as independent products, they can be stored in the above-mentioned readable storage medium. Based on such understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling one or more devices (which may be personal terminals, servers or network devices, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention.

[0126] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A voxel identification method for choroidal neovascularization, characterized in that: The steps include: Step S1: Acquire a choroidal neovascularization dataset; perform optical coherence tomography on the eyes of experimental animals to acquire a two-dimensional cross-sectional image sequence of different eye spots of the experimental animals as the choroidal neovascularization dataset; Step S2: constructing a three-dimensional structure data set and preprocessing it; stacking the two-dimensional cross-sectional image sequence of each different light spot into a three-dimensional structure to obtain a complete three-dimensional structure data set; dividing the three-dimensional data set into a training set and a test set; Step S3: Model construction; constructing a voxel segmentation model based on a three-dimensional convolutional neural network; wherein the structure of the voxel segmentation model includes an input layer, a three-layer encoder structure, a three-layer decoder structure and an output layer that are linked in sequence; wherein the three-layer encoder structure includes a first encoder, a second encoder, and a third encoder that are linked in sequence, and the three-layer decoder structure includes a first decoder, a second decoder, and a third decoder that are linked in sequence, and the third encoder and the third decoder are connected through a bottleneck layer, and the first encoder and the first decoder, the second encoder and the second decoder, and the third encoder and the third decoder are all connected by skip connections respectively; Step S4: model training; randomly selecting the image of the three-dimensional structure from the training set as a validation set, training based on the voxel segmentation model on the training set and the validation set, and using the test set to evaluate the generalization ability of the voxel segmentation model; Step S5: model application; using the trained voxel segmentation model to identify and predict the choroidal neovascularization dataset, outputting the prediction results of the choroidal neovascularization for subsequent medical statistical analysis, wherein the medical statistical analysis at least includes calculating the area and volume of the choroidal neovascularization.

2. The voxel identification method for choroidal neovascularization according to claim 1, characterized in that: The step S2 specifically includes: stacking each image in the two-dimensional cross-sectional image sequence in a time series to obtain a three-dimensional Numpy array to effectively capture the structural information of each light spot that varies with depth.

3. The voxel identification method for choroidal neovascularization according to claim 2, characterized in that: Each image in the two-dimensional cross-sectional image sequence is a grayscale image of size [h, w] pixels, where the height dimension h=330 and the width dimension w=770.

4. The voxel identification method for choroidal neovascularization according to claim 1, characterized in that: The step S3 specifically includes the following steps: Step S3.1: The three-dimensional structure data set is input into the input layer of the voxel segmentation model in the form of three-dimensional voxel bricks, and the size of each voxel brick is d×h×w×c, wherein the depth dimension d represents the size of the three-dimensional voxel brick along the depth direction, the height dimension h represents the size of the three-dimensional voxel brick along the height direction, the width dimension w represents the size of the three-dimensional voxel brick along the width direction, and the channel number c represents the number of channels of the three-dimensional voxel brick; Step S3.2: The three-layer encoder structure maps the three-dimensional voxel bricks to the vector space, performs feature extraction and downsampling to capture the high-level semantic information of the image; the three-layer decoder structure identifies each voxel of the input through the extracted feature vector, and restores the spatial details and resolution of the image through feature reconstruction and upsampling; the bottleneck layer combines the high-level features extracted by the third encoder with the feature map in the third decoder; the jump connection directly transfers the features of the three-layer encoder structure to the corresponding stages of the three-layer decoder structure to fuse feature information of different scales; Step S3.3: The output layer outputs a voxel category prediction map of the same size as the three-dimensional voxel brick of the input layer.

5. The voxel identification method for choroidal neovascularization according to claim 4, characterized in that: Each encoder and decoder of the three-layer encoder structure and the three-layer decoder structure adopts a sequentially ordered combination of GroupNorm3d+Conv3d+ReLU substructures, wherein the GroupNorm3d divides the input three-dimensional data of the three-dimensional voxel bricks into multiple groups, and normalizes the data in each group. The normalized data is convolved with the Conv3d convolution kernel to extract three-dimensional image features. After the convolution operation, the ReLU activation function is applied to convert linear features into nonlinear features; in each encoder layer, the GroupNorm3d is used to normalize the feature map output by each convolution layer, the Conv3d is used to extract features in three-dimensional space, and the ReLU activation function is applied to the convolution layer to enable the network to learn complex feature representations; in each decoder layer, the GroupNorm is used to normalize the feature map output by each convolution layer, the Conv3d is used to gradually restore the spatial resolution of the feature map and reduce the number of channels of the feature map to match the dimension of the original input three-dimensional data, and the ReLU function is used to reconstruct the details and structure of the input data and restore the complexity of the input data.

6. The voxel identification method for choroidal neovascularization according to claim 1, characterized in that: The section indexes for evaluating the voxel segmentation model in step S4 at least include precision, recall and Dice coefficient, and the best model training parameters are selected by reverse deduction through the values ​​of the section indexes.

7. The voxel identification method for choroidal neovascularization according to claim 1, characterized in that: The voxel segmentation model is an end-to-end 3D recognition algorithm.

8. A voxel recognition system for choroidal neovascularization, characterized in that: include: An optical coherence tomography device is used to obtain cross-sectional images of ocular spots of experimental animals as a choroidal neovascularization dataset; A server, wherein the digital recognition platform of the server receives the choroidal neovascularization dataset from the optical coherence tomography device, identifies and predicts the choroidal neovascularization dataset, and outputs the prediction result of the choroidal neovascularization for subsequent medical statistical analysis, wherein the medical statistical analysis at least includes calculating the area and volume of the choroidal neovascularization.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor runs the program, the steps of the voxel identification method for choroidal neovascularization according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions or a computer program, and when the computer-executable instructions or the computer program are processed and executed, the voxel identification method for choroidal neovascularization according to any one of claims 1 to 7 is implemented.

Images