Computer system for processing OCT (Optical Coherence Tomography) image, image processing method and medium

Abstract

The invention provides a computer system for processing an OCT image, an image processing method and a medium. The computer system comprises a processor; the image classification model comprises a feature extraction module used for performing feature extraction on each of a plurality of slice images of a target object to obtain slice features of each slice image, a first classifier used for classifying the slice features to obtain a first prediction result, and a second classifier used for classifying the first prediction result to obtain a second prediction result. The first classifier is used for determining the first prediction result and the slice feature as slice information of each slice image, and the second classifier is used for determining the weight of each slice image based on the slice information of at least part of the multiple slice images, performing weighted summation on the slice feature based on the weight to obtain a fusion feature, and obtaining a second prediction result based on the fusion feature; and a memory for storing instructions that, when executed by the processor, obtain a first prediction result and a second prediction result through the image classification model. Therefore, the accuracy of the user-level prediction result can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence, and in particular to a computer system, image processing method and medium for processing OCT images. Background Technology

[0002] Optical coherence tomography (OCT) technology can provide high-resolution three-dimensional structural images of the retina, whose data is essentially composed of multiple consecutive two-dimensional slice images (hereinafter sometimes simply referred to as slices). The pathological morphology of fundus diseases such as macular edema or neovascularization changes continuously in three-dimensional space. Therefore, it is necessary to comprehensively analyze the information carried by multiple or even all slice images to improve the accuracy of prediction.

[0003] Artificial intelligence (AI) models, such as convolutional neural networks, have demonstrated powerful capabilities in analyzing individual slice images and identifying pathological features within them. However, the aggregation methods employed in existing technologies are relatively rudimentary when transitioning from analyzing individual slice images to forming a "patient-level" prediction for the entire 3D OCT image. For example, simple voting (i.e., majority rule) or max pooling (i.e., selecting the prediction with the highest probability among all slices as the final result) are two commonly used strategies.

[0004] However, these methods have some drawbacks: on the one hand, they are easily interfered with by a large number of normal or artifact-containing sections, thus diluting or ignoring the "signals" emitted by a few but crucial early lesion sections; on the other hand, they fail to distinguish the importance of information contained in different sections. For example, the weight of key lesion sections located in the fovea of ​​the macula should obviously differ from that of sections located in the peripheral region in predicting the corresponding disease. In this case, an "information gap" is created from two-dimensional information to three-dimensional conclusions, resulting in insufficient accuracy and robustness of the final patient-level prediction results. Summary of the Invention

[0005] This disclosure is made in view of the above-mentioned situation, and its purpose is to provide a computer system, image processing method and medium for processing OCT images that can improve the accuracy of user-level prediction results.

[0006] Therefore, a first aspect of this disclosure provides a computer system for processing OCT images, comprising: at least one processor; a deep learning-based image classification model, comprising: a feature extraction module for extracting features from each of multiple slice images from the fundus of a target object to obtain slice features of each slice image, wherein the slice images are two-dimensional OCT images; a first classifier for classifying the slice features of each slice image to obtain a first prediction result for each slice image, the first prediction result including the probability of belonging to a corresponding category, the first prediction result and the slice features of each slice image being used as slice information of each slice image; and a second classifier for classifying the multiple slice images based on the slice features. The slice information of at least a portion of the slice images determines the weight of each slice image of the at least a portion of the images. Based on the weight of each slice image of the at least a portion of the images, the slice features of the at least a portion of the images are weighted and summed to obtain a fusion feature. Based on the fusion feature, a second prediction result is obtained, the second prediction result reflecting whether there is an anomaly in the target object. At least one memory is used to store instructions, the operations implemented by the instructions when executed by the at least one processor include: acquiring the multiple slice images of the fundus of the target object, inputting the multiple slice images into the image classification model to obtain a first prediction result and a second prediction result for each slice image.

[0007] In this scenario, information from multiple slice images can be integrated to obtain user-level prediction results. Furthermore, integrating information from multiple slice images can reduce the impact of inaccurate slice-level predictions, thereby improving the accuracy of user-level predictions. Additionally, the second classifier can calculate an adaptive weight for each slice image to generate highly condensed and information-rich user-level features, thus improving the accuracy and robustness of user-level predictions. Moreover, determining slice weights using slice features and the first prediction result allows for the incorporation of more information and perspectives from the first classifier's judgment, thereby improving the accuracy of user-level predictions through complementarity and error correction.

[0008] Additionally, in the computer system disclosed in the first aspect, optionally, the feature extraction module includes an encoder and a cue decoder connected to the encoder; the encoder receives each slice image, and the cue decoder receives learnable cue parameters and the output of the encoder and outputs slice features for each slice image. In this case, these category-aware learnable cue parameters can guide the cue decoder to extract discriminative features (i.e., enabling the cue decoder to learn disease-related information for a single slice, such as discriminative retinal information) to generate slice features, thereby improving the accuracy of subsequent slice-level classification and user-level classification.

[0009] Furthermore, in the computer system disclosed in the first aspect, the encoder may optionally be a fine-tuning-based VisionFM; the fine-tuning method is the LoRA method. In this case, since VisionFM uses more OCT data during pre-training, the extracted features are more useful for classifying OCT images, thus improving the robustness of the model. Additionally, the LoRA method can reduce the additional overhead of fine-tuning for OCT classification tasks while maintaining the model's general visual representation.

[0010] Additionally, in the computer system disclosed in the first aspect, optionally, the second classifier includes a cross-attention-based fusion module. This fusion module receives a feature evaluation vector of the at least partial image and a first prediction result of the at least partial image to calculate an attention weight for each slice of the at least partial image and use it as the weight. The feature evaluation vector is the result of feature evaluation of all slice features of the at least partial image. In this case, cross-attention can generate dynamic weights that are closely related to the classification task context, thereby helping the model learn more flexible feature representations. Furthermore, since the fusion module receives a feature evaluation vector, it can simplify the calculation of complex relationships between slice features into scalar processing, and facilitate the integration of other more suitable methods for analyzing slice features (e.g., value analysis and uncertainty analysis) to guide the model to focus on more important content.

[0011] Furthermore, in the computer system disclosed in the first aspect, optionally, the second classifier includes a first attention module and a second attention module; the first attention module receives all slice features of the at least partial image to generate first attention weights processed by the tanh activation function and used as a value-aware signal; the second attention module receives all slice features of the at least partial image to generate second attention weights processed by the sigmoid activation function and used as an uncertainty-aware signal; the value-aware signal and the uncertainty-aware signal are fused to obtain the feature evaluation vector. Thus, valuable slices and slices with high recognition reliability can be dynamically highlighted, facilitating subsequent robust user-level prediction.

[0012] Furthermore, in the computer system according to the first aspect of this disclosure, optionally, the second classifier is further configured to identify at least one abnormal slice image from the plurality of slice images based on a first prediction result for each slice image and use it as the at least partial image, wherein the normal probability and abnormal probability of each slice image are obtained based on the probability of belonging to the corresponding category, and at least one abnormal slice image is identified from the plurality of slice images based on the normal probability and abnormal probability of each slice image. In this case, the number of slice images that need to be weighted can be greatly reduced, thereby improving prediction efficiency. In addition, the model can be made to measure weights only based on the probability of abnormal slices. Furthermore, diseased areas are usually small and few in number in slices; removing normal slices allows the model to focus more on abnormal slices.

[0013] Furthermore, in the computer system disclosed in the first aspect, optionally, the loss function used to train the image classification model is the sum of a first loss function, a second loss function, and a third loss function; the first loss function is the difference between slice features of slice images with the same slice-level annotation; the second loss function is the difference between the first prediction result and the slice-level annotation; and the third loss function is the difference between the second prediction result and the user-level annotation. In this case, the first loss function can enhance both slice-level and user-level feature representations, thereby simultaneously promoting embedding representations with both classification discriminative power and compactness at both classification levels.

[0014] Furthermore, in the computer system involved in the first aspect of this disclosure, optionally, the first loss function satisfies the formula:

[0015] in, Let the first loss function be... The number of slice images in the training dataset used to train the image classification model. For the first training dataset A set of slice images with the same slice-level annotation. The number of elements in the set. For the training dataset of the first Slice features of a slice image For the set of Slice features of a slice image For the training dataset of the first Slice features of a slice image, "This is a dot product operation." For temperature coefficient, It is an exponential function. This is a filtering condition, indicating that the selected slice image is not the first slice in the training dataset. A slice image.

[0016] A second aspect of this disclosure provides an OCT image processing method based on an image classification model, comprising: acquiring multiple slice images of the fundus of a target object, wherein the slice images are two-dimensional OCT images; inputting the multiple slice images into a deep learning-based image classification model to obtain a first prediction result and a second prediction result for each slice image, wherein the image classification model includes: a feature extraction module for performing feature extraction on each of the multiple slice images of the fundus of the target object to obtain slice features of each slice image, wherein the slice images are two-dimensional OCT images; and a first classifier for classifying each slice image... The slice features are classified to obtain a first prediction result for each slice image. The first prediction result includes the probability of belonging to the corresponding category. The first prediction result and the slice features of each slice image are used as the slice information of each slice image. A second classifier is used to determine the weight of each slice image of the at least some images based on the slice information of the at least some images. The slice features of the at least some images are weighted and summed based on the weight of each slice image of the at least some images to obtain a fusion feature. A second prediction result is obtained based on the fusion feature. The second prediction result reflects whether there is an anomaly in the target object.

[0017] A third aspect of this disclosure provides a computer-readable storage medium storing at least one instruction that, when executed by a processor, implements the OCT image processing method as described in the second aspect of this disclosure.

[0018] According to this disclosure, a computer system, image processing method, and medium for processing OCT images are provided, which can improve the accuracy of user-level prediction results. Attached Figure Description

[0019] This disclosure will now be explained in further detail by way of example only with reference to the accompanying drawings.

[0020] Figure 1 This is a schematic diagram illustrating an example of the processing environment involved in this disclosure.

[0021] Figure 2 This is an exemplary block diagram illustrating a computer system as described in this disclosure.

[0022] Figure 3 This is a schematic diagram illustrating the structure of the image classification model involved in the example of this disclosure.

[0023] Figure 4This is a schematic diagram illustrating the structure of the prompt decoder involved in the example of this disclosure.

[0024] Figure 5 This is a schematic diagram illustrating the structure of the second classifier involved in the example of this disclosure.

[0025] Figure 6 This is an exemplary flowchart illustrating the image processing method involved in the examples of this disclosure.

[0026] Figure 7 This is an exemplary block diagram illustrating an electronic device to which the present disclosure is based. Detailed Implementation

[0027] Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same components, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the components or the shapes of the components may differ from actual figures. It should be noted that the terms "comprising" and "having," and any variations thereof, in this disclosure, do not necessarily limit the process, method, system, product, or apparatus to the explicitly listed steps or units, but may include or have other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0028] First, let me introduce the relevant terminology used in this disclosure.

[0029] An "encoder" can refer to a network structure or module in deep learning used to transform input data into a latent representation (such as a feature vector). Conversely, a "decoder" can refer to a structure used to transform the latent representation generated by the encoder back into the original input data.

[0030] "Annotation information" can refer to the labels assigned to an image after it has been annotated. It can also be called the gold standard. In the field of artificial intelligence, annotation information can be used to indicate what a neural network needs to predict or learn. Additionally, "annotation" can refer to the process of labeling or annotating a dataset (such as a training dataset). Annotation can include operations such as classifying images, text, or other types of data, marking bounding boxes, or semantic segmentation. For example, for medical images, the labels provided by annotation can be markings for target regions (such as lesion areas) on the medical image, or they can be classifications for the entire medical image.

[0031] As mentioned above, existing solutions have some drawbacks. The inventors have designed an adaptive aggregation technique for multiple slice images, which treats multiple slice images as a whole to obtain user-level prediction results, thereby achieving intelligent conversion from slice-level information to user-level prediction. Therefore, the inventors provide some solutions, and corresponding embodiments can at least solve some of the aforementioned problems, thereby improving the accuracy of user-level prediction results. Examples of this disclosure will be described in detail below.

[0032] This disclosure relates to a computer system for processing medical images, which can improve the accuracy of user-level prediction results. Furthermore, the computer system for processing medical images described in this disclosure may also be referred to as a computer system, computing device, or processing system, etc. Additionally, the medical image can be a two-dimensional image of a corresponding part of an organism. In some examples, the medical image can be a slice image in a three-dimensional OCT image (i.e., a two-dimensional OCT image), and accordingly, the computer system can also be referred to as a computer system for processing OCT images. In some examples, the corresponding part of the organism can be the fundus (e.g., the retina).

[0033] For ease of description, some examples below use the fundus as an example, and medical images are further described using OCT images of the fundus as an example. It should be noted that this does not imply any limitation on this disclosure. Unless there is a contradiction, the relevant descriptions also apply to two-dimensional images of other parts, two-dimensional images in other forms besides OCT images, or two-dimensional images of any three-dimensional images.

[0034] Examples of this disclosure will now be described in detail with reference to the accompanying drawings. Figure 1 This is a schematic diagram illustrating an example of the processing environment involved in this disclosure.

[0035] refer to Figure 1 The processing environment may include computer system 100. Computer system 100 can be any device with computing capabilities. For example, computer system 100 can be a cloud server, personal computer, mainframe, or distributed computing system. Computer system 100 can acquire input 200 and use deep learning model 300 to generate output 400 corresponding to input 200. In some examples, input 200 can be an image (e.g., a two-dimensional OCT image), and output 400 can be a prediction result, training parameters (e.g., weights), or performance metrics (e.g., accuracy, error rate, sensitivity, and specificity).

[0036] In some examples, the processing environment may also include a model training apparatus and a model application apparatus. The model training apparatus may be used to implement training methods for training the deep learning model 300 to obtain a trained model. The model application apparatus may be used to implement relevant methods for obtaining prediction results using the trained model. Additionally, the prediction results may be correlated with the output 400 of the trained model. In the examples of this disclosure, the trained deep learning model 300 may be a deep learning-based image classification model 130, which will be described later. In some examples, the image classification model 130 may include various deep learning models.

[0037] Figure 2 This is an exemplary block diagram illustrating a computer system 100 as described in this disclosure. Figure 3 This is a schematic diagram illustrating the structure of the image classification model 130 involved in the example of this disclosure.

[0038] In some examples, reference Figure 2 The computer system 100 may include at least one processor 110 (i.e., one or more processors 110) and at least one memory 120 (i.e., one or more memory 120). The one or more processors 110 may be any suitable processing device (e.g., processor core, microprocessor, ASIC, FPGA, controller, microcontroller, etc.) and may be a single processor or multiple processors operatively connected.

[0039] Additionally, memory 120 may store instructions that, when executed by processor 110, can implement the schemes involved in the examples of this disclosure (e.g., the image classification method described later). In some examples, memory 120 may include one or more non-transitory computer-readable storage media, such as RAM, ROM, EEPROM, EPROM, flash memory devices, disks, and combinations thereof.

[0040] In some examples, computer system 100 may include one or more server computing devices. If computer system 100 includes multiple server computing devices, these devices may operate in a sequential, parallel, or hybrid architecture.

[0041] In some examples, computer system 100 may include a deep learning-based image classification model 130. For example, computer system 100 may include one or more image classification models 130 in storage or otherwise.

[0042] In some examples, reference Figure 3The image classification model 130 may include a feature extraction module 131, which can be used to extract features from each slice image to obtain slice features D11 for each slice image. Specifically, the feature extraction module 131 can be used to extract features from each of multiple slice images from the fundus of the target object to obtain slice features D11 for each slice image (that is, the feature representation of each slice image, which can also be called slice-level information representation, depth feature representation, or high-dimensional features). For example, the OCT scan results of a patient can be divided into 20 slice images. After the feature extraction module 131, 20 slice features D11 can be obtained. In this case, each slice feature D11 contains rich image semantic information.

[0043] As mentioned above, slice images can be two-dimensional images. In some examples, slice images can be two-dimensional OCT images. In some examples, multiple slice images can be continuous in three-dimensional space. This helps the model identify continuous changes in pathological morphology in three-dimensional space, thereby improving the accuracy of prediction results.

[0044] In some examples, sliced ​​images can be derived from 3D images. (Reference) Figure 3 Two-dimensional OCT images can be derived from three-dimensional OCT images. For example, an OCT examination can yield at least one three-dimensional OCT image, and each three-dimensional OCT image can include multiple two-dimensional OCT images (i.e., a three-dimensional OCT image can be segmented into multiple slice images, see reference). Figure 3 ).

[0045] In some examples, multiple slice images of the target object can originate from a single 3D OCT image of the target object. Additionally, the target object can be any living organism (e.g., a human or animal).

[0046] In some examples, reference Figure 3 The feature extraction module 131 may include an encoder 1311 and a decoder 1312 connected to the encoder 1311. The encoder 1311 may receive each slice image, and the decoder 1312 may receive the output D12 of the encoder 1311 and output the slice features D11 of each slice image.

[0047] In some examples, encoder 1311 can be a fine-tuned VisionFM (Vision Foundation Model). In this case, because VisionFM uses more OCT data during pre-training, the extracted features are more useful for classifying OCT images, improving the model's robustness. VisionFM can be a pre-trained model. Specifically, VisionFM is an existing model pre-trained on a large OCT dataset with raw weights. VisionFM was pre-trained using 3.4 million ophthalmic images from over 500,000 subjects. In some examples, encoder 1311 can also be other models, such as RETFound or EyeClip.

[0048] Alternatively, fine-tuning can be based on the original weights of VisionFM. In some examples, the fine-tuning method can be LoRA (Low-Rank Adaptation). In this case, the number of parameters can be reduced and the model's performance can be improved, thereby increasing the efficiency of adapting VisionFM to the OCT classification task. This allows for maintaining the model's general visual representation while reducing the additional overhead of fine-tuning for the OCT classification task. In some examples, fine-tuning can also take other forms, such as QLoRA or an adapter.

[0049] Figure 4 This is a schematic diagram illustrating the structure of the prompt decoder 500 involved in the example of this disclosure.

[0050] In some examples, reference Figure 3 and Figure 4 Decoder 1312 can be a prompt decoder 500 (also known as a prompt-guided decoder). Prompt decoder 500 can receive learnable prompt parameters D13 and the output D12 of encoder 1311 and output slice features D11. In this case, these category-aware learnable prompt parameters D13 can guide prompt decoder 500 to extract discriminative features (i.e., enable prompt decoder 500 to learn disease-related information of a single slice, such as discriminative retinal information) to generate slice features D11, thereby improving the accuracy of subsequent slice-level classification and user-level classification.

[0051] Furthermore, compared to common attention mechanisms that primarily aim to construct correlations between sequence features, this approach allows for updating the learnable cue parameter D13 based on the output D12 of the encoder 1311, thereby obtaining cue parameters that measure the relevance of the slice image to the disease. In other words, the learnable cue parameter D13 serves as both the input and output of the cue decoder 500. Through iterative updates with a certain number of layers, the cue decoder 500 learns richer and more accurate slice-related information (i.e., the learnable cue parameter D13 ultimately output by the cue decoder 500 effectively expresses disease-related information).

[0052] Additionally, the learnable cue parameter D13 can be associated with the predicted category. For example, the cue decoder 500 can be guided by multiple learnable cue parameters D13, each corresponding to a category (e.g., a disease category).

[0053] In some examples, it is suggested that decoder 500 may be based on Transformer. (See reference) Figure 4 When the learnable cue parameter D13 and the output D12 of the encoder 1311 are input to the cue decoder 500, the learnable cue parameter D13 can be used as a query vector, and the output D12 of the encoder 1311 can be used as both a key vector and a value vector.

[0054] As an example, Figure 4 A schematic diagram of the structure of a Transformer-based cue decoder 500 is shown. From... Figure 4 As can be seen, the input (i.e., the learning cue parameter D13 and the output D12 of encoder 1311) can sequentially pass through cross attention, residual connection & layer normalization (Add & Norm), self-attention, residual connection & layer normalization (Add & Norm), feedforward network (FFN), residual connection & layer normalization (Add & Norm) before outputting (i.e., slice feature D11).

[0055] In some examples, a reference is returned. Figure 3 The image classification model 130 may include a first classifier 132. The first classifier 132 (also called a slice-level classification head or a two-dimensional image classification head) can be used to classify the slice features D11 of each slice image to obtain a first prediction result D14 for each slice image. Thus, a preliminary prediction result for each slice image can be obtained.

[0056] Additionally, the first prediction result D14 can reflect whether each slice image contains anomalies (i.e., whether each slice image contains pathological information). In some examples, the first prediction result D14 may include the probability of belonging to a corresponding category. For ease of description, the category classified by the first classifier 132 will be referred to as the first category, and the probability of the category will be referred to as the first category probability. For example, for the first category probability in vector form, each dimension of the vector can represent the predicted probability that the slice image belongs to a specific category.

[0057] Additionally, the first category can reflect the presence of anomalies in the sliced ​​image. In some examples, the first category can include normal and abnormal. In this case, the first category probability can include a normal probability and an abnormal probability. In some examples, the first category can include belonging to the corresponding disease and not belonging to the corresponding disease. In this case, the first category probability can include a belonging probability and a not belonging probability. In some examples, the first category can include normal and at least two diseases (i.e., multi-class prediction). In this case, the first category probability can include a normal probability and a probability for each disease. In some examples, the first category can be associated with retinal diseases. For example, the first category can include normal, age-related macular degeneration (AMD), choroidal neovascularization (CNV), central line serous chorioretinopathy (CSC), diabetic retinopathy (DR), macular edema (ME), macular hole (MH), epiretinal membrane (ERM), and retinitis pigmentosa (RP).

[0058] In some examples, the first classifier 132 can be based on a linear classifier. Specifically, the first classifier 132 can be used to perform a linear transformation (e.g., through a linear transformation layer) on the slice features D11 and then input them into a linear classifier (e.g., a softmax activation function) to generate a first class probability (also referred to as a probability distribution). Thus, features adapted to the classification task can be generated through linear transformation.

[0059] In this configuration, each slice image, after passing through the feature extraction module 131 and the first classifier 132, generates a probability distribution for at least one first category and obtains slice features D11 for each slice image. This facilitates subsequent selection of slice images containing pathological information (e.g., retinal pathological features) based on the first prediction result D14 and / or calculation of the weight of each slice image.

[0060] For ease of description, the first prediction result D14 and the slice feature D11 of each slice image are used as the slice information of each slice image.

[0061] In some examples, reference Figure 3The image classification model 130 may include a second classifier 133. The second classifier 133 can be used to classify all slice features D11 of at least a portion of multiple slice images (hereinafter referred to as the target slice set) to obtain a second prediction result D15 (i.e., multiple slice features D11 of the target slice set can be treated as a whole for classification to obtain a user-level (e.g., patient-level) prediction result). In this case, a user-level feature representation can be constructed from multiple slice images to obtain a user-level prediction result; for example, multiple two-dimensional slice images of a three-dimensional (3D) OCT image can be aggregated to obtain a patient-level prediction result.

[0062] In some examples, reference Figure 3 The first classifier 132 and the second classifier 133 can share the feature extraction module 131. This facilitates their joint influence on the network parameters of the feature extraction module 131 through training. In some examples, the second classifier 133 can receive the output from the feature extraction module 131 (i.e., slice feature D11) and the output from the first classifier 132 (i.e., the first prediction result D14) to obtain the second prediction result D15.

[0063] Figure 5 This is a schematic diagram illustrating the structure of the second classifier 133 involved in the example of this disclosure.

[0064] In some examples, the second classifier 133 can be used to determine the weight of each slice image in the target slice set (hereinafter referred to as slice weight) based on the slice information (i.e., the slice information sequence) of all slice images in the target slice set. The slice features D11 of the target slice set are then fused based on the weights of each slice image to obtain fused features D17 (i.e., user-level features). This allows for the calculation of an adaptive weight for each slice image (e.g., an abnormal slice containing pathological information) to generate highly condensed and information-rich user-level (e.g., patient-level) features. Furthermore, by using the slice features D11 of the slice images and the first prediction result D14 to determine the slice weights, the judgment of the first classifier 132 can be referenced to gather more information, thereby improving the accuracy of the user-level prediction results through complementarity and error correction.

[0065] Additionally, slice weights can represent the contribution (i.e., the degree of attention) of each slice image in obtaining the second prediction result D15. For example, the first slice might have a weight of 0.01, the second slice might have a weight of 0.2, and so on.

[0066] In some examples, the weights of slice images other than the target slice set in multiple slice images can be set to 0. For instance, if the target slice set consists of anomalous slices, the weights of normal slices in multiple slice images can be set to 0.

[0067] In some examples, fusion can be a weighted sum (e.g., multiplying and summing slice weights with slice features D11 (e.g., the anomaly feature D111 described later)). Specifically, the fusion feature D17 based on the weighted sum can satisfy the formula: , in, For the fusion feature D17, For the target slice set, The index of the slice image in the target slice set. For the first Slice weights of a sliced ​​image. For the first Slice features D11 of a slice image.

[0068] Additionally, the second prediction result D15 can reflect whether the target object has any anomalies (i.e., whether the target object has any anomalies from multiple slice images as a whole). In some examples, the second prediction result D15 may include the probability of belonging to the corresponding category. For ease of description, the category classified by the second classifier 133 will be referred to as the second category, and the probability of the second category will be referred to as the second category probability. The relevant descriptions are similar to those of the first category and the first category probability, and will not be repeated here. In some examples, for the image classification model 130, the second category may be consistent with the first category. For example, the first category may include normal and at least two diseases, and the second category may include normal and at least two diseases.

[0069] In some examples, the target slice set can be all images of multiple slice images of the target object. In some examples, the weights of slice images in the target slice set that do not contain anomalies can be set to 0 or values ​​close to 0. In some examples, the weights of slice images in the target slice set that do not contain anomalies can also be calculated using the fusion module described later.

[0070] The inventors discovered that while slice-level prediction can determine whether a sliced ​​image is abnormal, this predictive value is lost when performed as an independent task. Therefore, the inventors extracted anomalous slices based on first-class probability, increasing the model's focus on obviously anomalous slices, thus fully utilizing the information and improving prediction accuracy.

[0071] In some examples, the target slice set can be slice images with anomalies among multiple slice images of the target object. Specifically, the second classifier 133 is also used to identify at least one slice image with anomalies (hereinafter referred to as an anomalous slice) from multiple slice images of the fundus of the target object based on the first prediction result D14 of each slice image and use it as the target slice set. In this case, the number of slice images that need to be weighted can be greatly reduced, thereby improving prediction efficiency (for example, the lesions in ophthalmic slice images are very small, and there may be only 10 abnormal slices and 90 normal slices out of 100 slice images). In addition, the model can measure weights only based on the probability of anomalous slices. Furthermore, the diseased area is usually small and few in number in the slice; removing normal slices allows the model to focus more on anomalous slices.

[0072] In some examples, reference Figure 5 Based on the first category probability of the first prediction result D14, the normal probability and abnormal probability of each slice image can be obtained (that is, the normal / abnormal probability D16, which can also be called the binary classification probability). Based on the normal and abnormal probabilities of each slice image, at least one abnormal slice can be identified from multiple slice images. That is, the slice-level multi-class prediction is converted into a binary decision to identify abnormal slices. Thus, the slice information of the abnormal slice (i.e., the abnormal feature D111 and the abnormal prediction result D141) can be obtained.

[0073] For example, after obtaining the first category probability (e.g., normal: 0.3, first disease: 0.2, second disease: 0.2, ...), the first category probability can be adjusted to normal and abnormal probabilities (e.g., normal: 0.3, abnormal: 0.7), and abnormal slices can be selected based on the normal and abnormal probabilities. (Reference) Figure 5 For example, anomaly feature D111 can be obtained by filtering the slice features D11 of multiple slice images of the target object, and anomaly prediction result D141 can be obtained by filtering the first prediction result D14 of multiple slice images of the target object.

[0074] In some examples, slices with a higher probability of being abnormal than normal can be considered as anomalous slices.

[0075] As mentioned above, the weight of each slice image can be determined based on the slice information of all slice images in the target slice set. In some examples, slice information can be processed based on cross-attention to obtain slice weights. In this case, cross-attention can generate dynamic weights that are closely related to the classification task context, thereby helping the model learn more flexible feature representations. In addition, cross-attention receives feature evaluation vectors, which on the one hand simplifies the calculation of the complex relationship between recognition slice features D11 to scalar processing, and on the other hand facilitates the integration of other more suitable methods for analyzing slice feature D11 (such as value analysis and uncertainty analysis) to guide the model to focus on more important content.

[0076] Specifically, refer to Figure 5 The second classifier 133 may include a fusion module 1331 based on cross-attention. The fusion module 1331 may receive feature evaluation vectors of the target slice set (e.g., all anomalous slices) and a first prediction result D14 of the target slice set (e.g., anomaly prediction result D141) to calculate the attention weight of each slice image in the target slice set and use it as the slice weight. Thus, it is possible to calculate the normalized slice weight for each slice image.

[0077] In some examples, when the feature evaluation vector of the target slice set and the first prediction result D14 of the target slice set are input into the fusion module 1331, the first prediction result D14 can be used as a query vector, and the feature evaluation vector can be used as both a key vector and a value vector.

[0078] In some examples, the feature evaluation vector can be adjusted via a linear layer and then input into the fusion module 1331 along with the first prediction result D14 of the target slice set.

[0079] Furthermore, the feature evaluation vector of the target slice set can be determined by all slice features D11 of the target slice set. Specifically, the feature evaluation vector can be the result of feature evaluation of all slice features D11 of the target slice set. The feature evaluation vector can be used to guide the generation of slice weights. In some examples, the feature evaluation vector can be determined by performing value analysis and uncertainty analysis on each of all slice features D11 of the target slice set through an attention mechanism. This allows for the dynamic highlighting of valuable slices and the identification of highly reliable slices, facilitating subsequent robust user-level predictions.

[0080] In some examples, reference Figure 5The second classifier 133 may include a first attention module 1332 (also called a value-aware module) and a second attention module 1333 (also called an uncertainty-aware module). The first attention module 1332 receives all slice features D11 of the target slice set to generate first attention weights, which can serve as value-aware signals. The second attention module 1333 receives all slice features D11 of the target slice set to generate second attention weights, which can serve as uncertainty-aware signals. For example, the slice features D11 of all abnormal slices of a target object can be input together into the first attention module 1332 and the second attention module 1333 respectively. The first attention module 1332 determines the value of each abnormal slice, and the second attention module 1333 determines the uncertainty of each abnormal slice.

[0081] Additionally, the value-perceived signal can be used to measure the information value of slice feature D11 in obtaining the second prediction result D15 (that is, the value-perceived signal can be used to assess the importance or contribution of the information contained in slice feature D11 to the final decision). The uncertainty-perceived signal can be used to measure the reliability of the first prediction result D14.

[0082] In some examples, the value range of the value-aware signal can be [-1, 1], and the value range of the uncertainty-aware signal can be [0, 1].

[0083] In some examples, the first attention module 1332 and the second attention module 1333 can receive all slice features D11 of the target slice set in parallel.

[0084] In some examples, reference Figure 5 The first attention module 1332 may include a tanh activation function, and the first attention weights may have already been processed by the tanh activation function. Therefore, the range of the tanh activation function is symmetric, which helps to reflect both positive and negative information input. Specifically, the tanh activation function can receive the attention score generated by the first attention module 1332 and output the first attention weights.

[0085] In some examples, reference Figure 5 The second attention module 1333 may include a sigmoid activation function, and the second attention weights may have already been processed by the sigmoid activation function. Therefore, the range of values ​​for the sigmoid activation function helps to reflect the reliability of the prediction results. Specifically, the sigmoid activation function may receive the attention scores generated by the second attention module 1333 and output the second attention weights.

[0086] In this case, value-perceived signals and uncertainty-perceived signals can be obtained by using different activation functions (i.e., tanh activation function and sigmoid activation function).

[0087] In some examples, the value-perceived signal and the uncertainty-perceived signal can be fused to obtain the feature evaluation vector. This allows the feature evaluation vector to incorporate the "value" and "reliability" of slice feature D11. In some examples, the value-perceived signal and the uncertainty-perceived signal can be multiplied element-wise (i.e., element-wise product) to obtain the feature evaluation vector.

[0088] In some examples, reference Figure 5 The second classifier 133 can also be used to obtain a second prediction result D15 based on the fusion feature D17. Specifically, the second classifier 133 can be used to classify the fusion feature D17 and output the second prediction result D15. Thus, information from multiple slice images can be integrated to obtain a user-level prediction result. Furthermore, integrating information from multiple slice images can reduce the impact of inaccurate slice-level prediction results, thereby improving the accuracy of the user-level prediction result. For example, although the first prediction result D14 of a certain slice image may be biased, the bias can be corrected by performing an overall prediction using multiple slices.

[0089] In some examples, reference Figure 5 The second classifier 133 may include a classification head 1334, which can be used to classify the fused feature D17 and output a second prediction result D15. In some examples, the classification head 1334 may be based on a linear classifier. Specifically, refer to... Figure 5 The classification head 1334 can be used to linearly transform the fused feature D17 (e.g., through a linear transformation layer) and then input it into a linear classifier (e.g., a softmax activation function) to generate a second class probability (also known as a probability distribution). In this case, linear transformation can generate features adapted to the classification task. Furthermore, deep networks themselves possess powerful non-linear mapping capabilities, and the final features are often linearly separable. Compared to using complex non-linear classifiers, this reduces computational complexity, minimizes disruption to end-to-end trainability, and increases the stability of gradient propagation. In some examples, the classification head 1334 can also be based on a non-linear classifier (e.g., MLP (Multilayer Perceptron), XGBoost, etc.).

[0090] In some examples, an initial image classification model 130 can be trained using a training dataset to obtain a trained image classification model 130. In some examples, the training dataset may include a set of slice images of multiple sample target objects. Each image in each set of slice images may have slice-level annotation information (i.e., one slice image corresponds to one annotation, also known as slice-level annotation), and each set of slice images for each sample target object may have user-level annotation information (i.e., one set of slice images for one sample target object corresponds to one annotation, also known as user-level annotation). For example, multiple patients can be selected, with each patient associated with a set of slice images, corresponding slice-level annotation information, and user-level annotation information. It should be noted that the annotation information matches the category to be classified (e.g., the first category and the second category), which will not be elaborated further here.

[0091] To enhance slice-level and user-level feature representations, the inventors introduced supervised contrastive loss on top of standard cross-entropy loss, thereby simultaneously promoting embedded representations with classification discriminative power and compactness at both classification levels.

[0092] In some examples, the loss function used to train the image classification model 130 may include a first loss function, a second loss function, and a third loss function. In some examples, the loss function may be the sum of the first loss function, the second loss function, and the third loss function.

[0093] Additionally, the first loss function can be a supervised contrastive loss based on slice features D11 and slice-level annotations. In some examples, the first loss function can be the difference between slice features D11 of slice images with the same slice-level annotations in the training dataset. For example, suppose there are three slice features D11: a, b, and c, where features a and b have the same annotation information, and c is different. The first loss function can shorten the distance between features a and b, and increase the distance between features c, making slice features D11 with different annotation information more discriminative. In some examples, the first loss function can satisfy the formula:

[0094] in, For the first loss function, The number of slice images in the training dataset (i.e., the number of slice images used for training). To compare with the training dataset A set of slice images with the same slice-level annotation. The number of elements in the set. For the training dataset, the first Slice features D11 of a slice image. For the set of Slice features D11 of a slice image. For the training dataset, the first Slice features D11 of a slice image, "This is a dot product operation." For temperature coefficient, It is an exponential function. This is a filtering condition, indicating that the selected slice image is not the first slice in the training dataset. A slice image. Additionally, the temperature coefficient can be a hyperparameter that can be adjusted based on model performance (e.g., it can be set to 0.07).

[0095] Alternatively, the second loss function can be the difference between the first prediction D14 of the sliced ​​image used for training and the slice-level annotations. In some examples, the second loss function can be the cross-entropy loss function.

[0096] Alternatively, the third loss function can be the difference between the second prediction result D15 of multiple slice images of the target object used for training and the user-level annotation. In some examples, the third loss function can be the cross-entropy loss function.

[0097] To this end, this disclosure also provides some training examples. Specifically, a set of slice images of each sample target object can be input into VisionFM to obtain the encoded features of each slice image (i.e., the output of VisionFM). The encoded features and the learnable cue parameter D13 are then input into the cue decoder 500 to output slice features D11. Thus, the learnable cue parameter D13 can effectively express the disease-related information of the slice. The output slice features D11 can be input into the first classifier 132 to obtain the first prediction result D14 (i.e., the slice-level prediction result) with the first category probability. The multiple slice features D11 output by the cue decoder 500 and the multiple first category probabilities obtained by the first classifier 132 can be input into the second classifier 133. The second classifier 133 extracts abnormal slices according to the first category probabilities and assigns weights to each abnormal slice according to the slice features D11 of all abnormal slices and the first category probabilities. Finally, the slice features D11 of all abnormal slices are fused and predicted to obtain the second prediction result D15 (i.e., the user-level prediction result). Finally, the first prediction result D14 and the second prediction result D15 can be output together. Based on the differences between the first prediction result D14 and the slice-level annotation, the differences between the second prediction result D15 and the user-level annotation, and the differences in slice features D11 of slice images with the same slice-level annotation, a loss function is constructed to continuously optimize the image classification model 130 until the loss meets certain conditions.

[0098] In some examples, the image classification model 130 can be trained end-to-end. Specifically, during model training, the data input from the input end is used to obtain a predicted result at the output end. This predicted result is compared with the true result (i.e., the labeled information) to obtain an error. This error is propagated through each layer of the model (i.e., backpropagation), and the representation of each layer can be adjusted based on this error until the model converges or achieves the expected result. This is end-to-end training. In other words, end-to-end training also delegates the task of feature extraction to the model, directly inputting raw data or data with slight preprocessing, allowing the model to perform feature extraction itself.

[0099] In this context, multiple modules within the image classification model 130 (e.g., feature extraction module 131, first classifier 132, and second classifier 133) can mutually promote and improve each other, thereby enhancing not only the accuracy of user-level prediction results but also the accuracy of slice-level prediction results. For example, in OCT images, not every slice contains obvious pathological information (e.g., a slice may contain images surrounding pathological information, which is difficult for both humans and machines to accurately identify). However, by training them together, they can collectively influence the feature extraction module 131, thereby improving the accuracy of feature extraction and, consequently, the accuracy of subsequent predictions.

[0100] This disclosure also relates to an OCT image processing method based on an image classification model 130, which may be simply referred to as an image processing method.

[0101] Figure 6 This is an exemplary flowchart illustrating an image processing method according to an example of this disclosure. Furthermore, at least one step in the image processing method can be implemented by at least a portion of the components of the computer system 100. It should be noted that, unless contradictory, the relevant description of the computer system 100 can also be applied to the image processing method.

[0102] In some examples, reference Figure 6 The image processing method may include acquiring multiple slice images of the fundus of the target object (step S101), and inputting the multiple slice images into the image classification model 130 to obtain a first prediction result D14 and a second prediction result D15 for each slice image (step S102).

[0103] Figure 7 This is an exemplary block diagram illustrating an electronic device 8 as described in this disclosure.

[0104] The examples of this disclosure also relate to an electronic device 8, see reference 8. Figure 7The electronic device 8 may include a processor 801 (e.g., a central processing unit or a graphics processing unit) and a memory 802. The memory 802 may store a computer program that, when executed, implements one or more steps of the image processing method or image classification method described above.

[0105] In some examples, memory 802 may include a computer-readable storage medium, which may include, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory, etc. For example, computer-executable instructions may be loaded from memory unit 807 (described later) into random access memory (RAM) to execute computer-executable instructions. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, USB storage, flash memory, etc.

[0106] In some examples, reference Figure 7 The processor 801 and memory 802 can be connected to each other via bus 803. In some examples, the electronic device 8 may also include an input / output interface 804, which can be connected to bus 803. In some examples, the electronic device 8 may also include an input unit 805 (e.g., touch screen, keyboard, mouse, camera, microphone, etc.), an output unit 806 (e.g., display, speaker, etc.), a storage unit 807 (e.g., magnetic tape, hard disk, flash memory, etc.), and / or a communication unit 808, which can be connected to the input / output interface 804. Additionally, the communication unit 808 can be used to enable the electronic device 8 to communicate wirelessly or wiredly with other electronic devices.

[0107] Examples of this disclosure also disclose a computer-readable storage medium that can store at least one instruction, which, when executed by a processor, implements one or more steps of the image processing method or image classification method described above. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0108] The scheme disclosed in this example can simulate the comprehensive judgment ability of ophthalmologists, assigning different weights to slices of different importance, thereby achieving a qualitative leap from massive two-dimensional data to accurate three-dimensional prediction.

[0109] To verify the effectiveness of the scheme disclosed herein (hereinafter referred to as "this scheme"), the inventors compared its performance with various baseline aggregation methods. The experiments were conducted using the PyTorch deep learning framework on a hardware environment consisting of four NVIDIA A6000 GPUs. During the training phase of the image classification model 130, the Adam optimizer was used (initial learning rate 3×10⁻⁶). -4 Weight decay 1×10 -5 The cosine annealing learning rate scheduling strategy, which includes 5 warm-up cycles, is adopted, and encoder 1311 is configured as VisionFM with LoRA parameters (rank). =8, scaling factor =16, random inactivation rate=0.1), decoder 1312 is the cue decoder 500, the second classifier 133 is equipped with a module for value analysis and uncertainty analysis of slice feature D11, and is trained for 50 cycles with batch size 4.

[0110] Training data: The training was conducted on an internal test set containing 40,672 slides from 1,964 patients and an external test set containing 67,531 slides from 3,300 patients.

[0111] The comparison methods include this scheme, ABMIL as described in the literature (Ilse, M., Tomczak, JM & Welling, M. Attention-based Deep Multiple Instance Learning. Preprint at https: / / doi.org / 10.48550 / arXiv.1802.04712 (2018)), S4MIL as described in the literature (Fillioux, L., Boyd, J., Vakalopoulou, M., Cournède, P.-H. & Christodoulidis, S. Structured StateSpace Models for Multiple Instance Learning in Digital Pathology. Preprint at https: / / doi.org / 10.48550 / arXiv.2306.15789 (2023)), and M4 as described in the literature (Li, J. et al. M4: Multi-proxy multi-gate mixture of experts network for multiple instance learning in histopathology image analysis. Medical Image Analysis 103, 103561(2025)). ABMIL is a classic attention-based multi-instance learning aggregation method, while S4MIL and M4 are other advanced multi-instance learning aggregation methods.

[0112] Experimental results: The user-level predicted F1-score of this disclosure is 94.39% and 94.05% in internal and external tests, respectively. The performance comparison with other comparative methods is shown in Table 1.

[0113]

[0114] Table 1 Performance comparison of the comparison methods Experimental results clearly demonstrate that the adaptive aggregation scheme proposed in this disclosure, which simultaneously considers information value and uncertainty, significantly outperforms (p<0.05) several existing classical and advanced aggregation methods. This proves that this disclosure has achieved a substantial technical breakthrough in addressing the core technical problem of generating accurate and robust 3D user-level predictions from massive 2D slice information.

[0115] The above embodiments and their detailed experimental verification results demonstrate that the end-to-end automated prediction computer system 100 and related methods and media proposed in this disclosure, through their innovative two-stage architecture, powerful visual foundation model, and novel adaptive aggregation mechanism, successfully address many pain points of existing technologies. It not only achieves a high degree of automation and efficiency improvement throughout the entire OCT image classification process, but also reaches industry-leading levels in accuracy and generalization ability, showcasing enormous clinical application potential and value.

[0116] Furthermore, the research disclosed herein has been approved by the Ethics Review Committee of Zhongshan Ophthalmic Center, Sun Yat-sen University (IRB-ZOC-SYSU, No.: 2023KYPJ344). All procedures were conducted in accordance with the principles of the Declaration of Helsinki. Informed consent is not required because the data was collected retrospectively and personal information has been de-identified.

[0117] While the present disclosure has been specifically described above in conjunction with the accompanying drawings and examples, it is to be understood that the foregoing description does not limit the present disclosure in any way. Those skilled in the art can make modifications and variations to the present disclosure as needed without departing from its essential spirit and scope, and all such modifications and variations shall fall within the scope of the present disclosure.

Claims

1. A computer system for processing OCT images, characterized in that, include: At least one processor; Deep learning-based image classification models include: The feature extraction module is used to extract features from each of multiple slice images of the fundus of the target object to obtain the slice features of each slice image. The slice images are two-dimensional OCT images. The first classifier classifies the slice features of each slice image to obtain a first prediction result for each slice image. The first prediction result includes the probability of belonging to the corresponding category. The first prediction result and the slice features of each slice image are used as the slice information of each slice image. The second classifier is used to determine the weight of each slice image of the at least some images based on the slice information of the multiple slice images, to perform a weighted summation of the slice features of the at least some images based on the weight of each slice image of the at least some images to obtain a fusion feature, and to obtain a second prediction result based on the fusion feature, wherein the second prediction result reflects whether there is an anomaly in the target object. At least one memory is used to store instructions, the operations implemented when executed by the at least one processor including: Obtain the multiple slice images of the fundus of the target object. The multiple slice images are input into the image classification model to obtain a first prediction result and a second prediction result for each slice image.

2. The computer system according to claim 1, characterized in that, The feature extraction module includes an encoder and a cue decoder connected to the encoder; The encoder receives each slice image, and the cue decoder receives learnable cue parameters and the output of the encoder and outputs the slice features of each slice image.

3. The computer system according to claim 2, characterized in that, The encoder is based on fine-tuning VisionFM; The fine-tuning method is the LoRA method.

4. The computer system according to claim 1, characterized in that, The second classifier includes a cross-attention-based fusion module, which receives a feature evaluation vector of the at least partial image and a first prediction result of the at least partial image to calculate the attention weight of each slice image in the at least partial image and use it as the weight. The feature evaluation vector is the result of feature evaluation of all slice features of the at least partial image.

5. The computer system according to claim 4, characterized in that, The second classifier includes a first attention module and a second attention module; The first attention module receives all slice features of the at least part of the image to generate first attention weights that have been processed by the tanh activation function and serve as a value-perceived signal; The second attention module receives all slice features of the at least partial image to generate second attention weights that have been processed by the sigmoid activation function and serve as an uncertainty-aware signal; The value perception signal and the uncertainty perception signal are fused to obtain the feature evaluation vector.

6. The computer system according to claim 1, characterized in that, The second classifier is further configured to identify at least one abnormal slice image from the plurality of slice images based on the first prediction result of each slice image and to use it as the at least partial image, wherein the normal probability and abnormal probability of each slice image are obtained based on the probability of belonging to the corresponding category, and at least one abnormal slice image is identified from the plurality of slice images based on the normal probability and abnormal probability of each slice image.

7. The computer system according to claim 1, characterized in that, The loss function used to train the image classification model is the sum of the first loss function, the second loss function, and the third loss function; The first loss function is the difference between slice features of slice images with the same slice-level annotation; The second loss function is the difference between the first prediction result and the slice-level annotation; The third loss function is the difference between the second prediction result and the user-level annotation.

8. The computer system according to claim 7, characterized in that, The first loss function satisfies the formula: in, Let the first loss function be... The number of slice images in the training dataset used to train the image classification model. For the first training dataset A set of slice images with the same slice-level annotation. The number of elements in the set. For the training dataset of the first Slice features of a slice image For the set of Slice features of a slice image For the training dataset of the first Slice features of a slice image, "This is a dot product operation." For temperature coefficient, It is an exponential function. This is a filtering condition, indicating that the selected slice image is not the first slice in the training dataset. A slice image.

9. An OCT image processing method based on an image classification model, characterized in that, include: Multiple slice images of the fundus of the target object are obtained, and the slice images are two-dimensional OCT images. The multiple slice images are input into a deep learning-based image classification model to obtain a first prediction result and a second prediction result for each slice image, wherein the image classification model includes: The feature extraction module is used to extract features from each of multiple slice images of the fundus of the target object to obtain the slice features of each slice image. The slice images are two-dimensional OCT images. The first classifier classifies the slice features of each slice image to obtain a first prediction result for each slice image. The first prediction result includes the probability of belonging to the corresponding category. The first prediction result and the slice features of each slice image are used as the slice information of each slice image. The second classifier is used to determine the weight of each slice image of the at least some of the slice images based on the slice information of the at least some of the slice images, to perform a weighted summation of the slice features of the at least some of the images based on the weight of each slice image of the at least some of the images to obtain a fusion feature, and to obtain a second prediction result based on the fusion feature, the second prediction result reflecting whether there is an anomaly in the target object.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the OCT image processing method as described in claim 9.

Images