Three-dimensional segmentation method for human fundus amd high-reflection area based on deep learning

Abstract

The application provides a three-dimensional segmentation method for age-related macular degeneration (AMD) high-reflection area of human fundus, an electronic device and a storage medium, and the method comprises the following steps: obtaining an initial fundus optical coherence tomography image output by an optical coherence tomography system; a deep learning network for segmenting the AMD high-reflection area of the fundus is pre-trained, the network can automatically lock the segmentation area and extract feature information of different sizes; the initial fundus optical coherence tomography image is input into the deep learning network model for segmenting the AMD high-reflection area of the fundus, which is obtained by pre-training, and the initial fundus optical coherence tomography image is three-dimensionally segmented by the deep learning network model to obtain a three-dimensional stereoscopic lesion area image. The deep learning model can correctly, intuitively and clearly segment the AMD high-reflection area of the fundus of a patient, thus providing great convenience for clinical diagnosis and follow-up of AMD, and improving the acceptance and recognition of the patient to the diagnosis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of OCT medical image segmentation, in particular to a three-dimensional segmentation method of an optical coherence tomography image of an AMD high reflection area of a human fundus, an electronic device and a storage medium. BACKGROUND

[0002] Age-related macular degeneration (AMD) is a fundus disease that often occurs in people over 60 years old and is one of the main causes of blindness in the elderly population worldwide. Wet AMD is a late form of AMD, accounting for more than half of AMD blindness cases. Subretinal high reflection area (SHM) can indicate the risk of progression of wet AMD, so its size and shape are closely related to the clinical diagnosis and monitoring of AMD.

[0003] Optical coherence tomography (OCT) has been widely used in ophthalmic examinations and related biomedical research fields due to its non-invasive, high-resolution and three-dimensional imaging of biological tissues. At present, many AMD fundus segmentation models based on deep learning can only present lesions in two-dimensional form, but cannot show the three-dimensional morphology of actual AMD lesions. Therefore, in solving the problem of clinical AMD segmentation, it is expected to propose a high-efficiency three-dimensional segmentation model of AMD fundus lesions to assist doctors in diagnosis, thereby reducing the probability of misdiagnosis and missed diagnosis and improving the efficiency of diagnosis. SUMMARY

[0004] In order to solve the problems existing in the prior art, the present application provides a three-dimensional segmentation method of an AMD high reflection area of a human fundus based on deep learning, a device and a medium, which can automatically identify and three-dimensionally segment the AMD high reflection area of the human fundus.

[0005] To achieve the above-mentioned purpose, the technical solutions adopted by the present application are as follows:

[0006] In a first aspect, the present application provides a three-dimensional segmentation method of an AMD high reflection area of a human fundus based on deep learning, comprising:

[0007] Obtaining an initial OCT image of an AMD patient fundus, wherein the initial OCT image of the AMD patient fundus has an AMD high reflection area;

[0008] Inputting the initial OCT image of the AMD patient fundus into a three-dimensional segmentation deep learning model trained in advance, and performing three-dimensional segmentation on the initial OCT image of the fundus by the three-dimensional segmentation deep learning model to obtain a three-dimensional result of the AMD high reflection area of the fundus.

[0009] The three-dimensional segmentation deep learning model is a U-Net structure composed of an encoder and a decoder connected in sequence.

[0010] The initial diseased fundus OCT image is input into the three-dimensional segmentation deep learning network model for AMD high-reflection region of the fundus pre-trained, and the initial diseased fundus OCT image is three-dimensionally segmented by the three-dimensional segmentation deep learning network model for AMD high-reflection region of the fundus to obtain a three-dimensional target AMD high-reflection region image.

[0011] Before the initial AMD patient fundus OCT image is input into the three-dimensional segmentation deep learning model pre-trained, the following steps are further included:

[0012] An initial AMD patient fundus OCT image data set is constructed, and the initial AMD patient fundus OCT image data set contains a plurality of data pairs, each data pair containing an original fundus OCT image and a true value image in which a fundus AMD high-reflection region is artificially labeled;

[0013] An initial model is constructed, and the initial model includes: a structure based on an improved U-Net, sequentially connected with an encoder, a hollow pyramid pooling module and a decoder; wherein the encoder and the decoder are connected through a skip connection containing a Squeeze-and-excitation module;

[0014] The initial model is trained based on a sample data set to obtain a three-dimensional segmentation deep learning model.

[0015] The initial model is constructed, including:

[0016] A U-Net is used as the backbone of the initial model, and the encoder and the decoder are connected through a skip connection, so as to mine more feature information; the encoder is a structure composed of a self-defined mixed down-sampling module and a Squeeze-and-Excitation module, which can avoid the gradient problem caused by the increase of the number of layers while realizing the redistribution of the weights between the channels of the image; the decoder is composed of a deconvolution layer and a series of conventional convolution activation layers, and is connected to the corresponding part of the encoder through a skip connection; the decoder up-samples the input feature map, so that the feature map is restored to the original size, and then a channel attention module is used to adaptively calculate the attention weight of each channel feature and adjust the utilization rate of different features; unlike the traditional skip connection, each skip connection in the model also passes through a Squeeze-and-Excitation module to reduce the loss of feature information;

[0017] The self-defined hybrid down-sampling module is a dense residual structure integrating a regular convolution layer and a Relu activation function; after the input image is subjected to feature extraction by the five self-defined hybrid down-sampling modules, a dense low-resolution feature map is generated;

[0018] After the low-resolution feature map generated by the self-defined hybrid down-sampling module is received, the Squeeze-and-Excitation module adaptively corrects the corresponding intensity of the features between the channels of the feature map, thereby improving the segmentation effect;

[0019] After the low-resolution feature map is generated by the decoder, a dilated pyramid pooling module is used to further mine feature information, which is then input into the decoder for up-sampling; the channel attention module after the encoder causes the weight of each channel in the up-sampled feature map to be redistributed, while keeping the original regional features and filtering out unnecessary features.

[0020] The construction process of the sample data set includes:

[0021] An initial OCT image of the high-reflection area of the human fundus AMD is obtained, and the initial OCT image of the human fundus has an AMD high-reflection area to be segmented;

[0022] The data set is preprocessed, and the preprocessing operation is selected from one or more of the following operations: data set enhancement, adjusting picture parameters, cropping picture size, and target area manual annotation operation;

[0023] Each data pair corresponding to each sample is combined, and a training set, a test set and a validation set are divided;

[0024] Optionally, the initial model is trained based on the sample data set to obtain the three-dimensional segmentation deep learning model, including the following steps:

[0025] Step one, the training set obtained is used to train the network model, and the proportion of the training, test set and validation set obtained is 8:1:1; the initial AMD patient fundus OCT image after preprocessing is input into the network to extract features by the self-defined hybrid down-sampling module, to obtain high-level semantic features and low-level semantic features; the high-level semantic features are input into the dilated pyramid pooling module to further extract high-level semantic information, which is independently subjected to five dilated convolution layers, the five feature maps obtained are merged by splicing, and finally a 1 1 convolution compression feature;

[0026] Step two, the feature map obtained in step one is up-sampled, and then the low-level semantic feature map with the same resolution after up-sampling is found, which is subjected to 3 3, the jump connection composed of the convolution and the squeeze-and-excitation module is spliced and combined, then the operation is repeated four times until the final output size returns to the original input size, and the extraction of spatial information is completed, finally, the feature map is passed through the channel attention module, so that the weights between channels are redistributed according to the priority, and finally the segmentation prediction result is obtained;

[0027] Step three, calculating the segmentation loss according to the obtained prediction result;

[0028] Step four, repeating steps one to three, updating the network parameters according to the segmentation loss, and saving the network weight parameters when the segmentation loss meets the preset condition, and the model is the three-dimensional segmentation deep learning model.

[0029] Optionally, the loss information determined according to the prediction result and the corresponding true value image includes:

[0030] According to the segmentation result of the segmentation network and the true value image, the model loss function is calculated

[0031] (1)

[0032] (2)

[0033] (3)

[0034] By adjusting the hyperparameters of the Tversky loss function , the balance between false positives, true negatives, true positives and false negatives can be controlled. According to the proportion of false positives and false negatives in the prediction result, adjusting the hyperparameters α and β can improve the segmentation accuracy and reduce outliers;

[0035] And the binary cross entropy loss function can eliminate the data imbalance problem caused by the large difference between the lesion area and the background volume, so as to further improve the segmentation effect.

[0036] In a second aspect, the embodiments of the present application also provide an electronic device, comprising: a processor, a storage medium and a bus, the storage medium storing program instructions executable by the processor, when the electronic device is running, the processor and the storage are communicated through the bus, and the machine readable instructions are executed by the processor to perform the steps of segmenting the three-dimensional AMD high-reflective lesion area OCT image of the human fundus as described above.

[0037] In a third aspect, the embodiments of the present application further provide an electronic device, comprising a processor, a storage medium and a bus, the storage medium stores program instructions executable by the processor, when an application program is running, the processor and the storage medium communicate through the bus, the processor executes the program instructions to perform the steps of the three-dimensional segmentation of the AMD high-reflection lesion area of the human fundus in the first aspect.

[0038] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is read and executes the steps of the three-dimensional segmentation of the AMD high-reflection lesion area of the human fundus in the first aspect.

[0039] The beneficial effects of the present application are:

[0040] The present application provides a three-dimensional segmentation model of the AMD high-reflection area of the human fundus based on the improved U-Net, which can provide convenience and reliability when applied to clinical diagnosis and monitoring of AMD, reduce the probability of misdiagnosis and missed diagnosis, improve the diagnosis efficiency, and has good development prospects in the medical field. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0042] Figure 1 It is an application scenario schematic diagram of the three-dimensional segmentation method of the human fundus AMD optical coherence tomography image of the present application.

[0043] Figure 2 It is a flowchart of the three-dimensional segmentation method of the human fundus AMD optical coherence tomography image of the present application.

[0044] Figure 3 It is a whole network structure schematic diagram of the three-dimensional segmentation method of the human fundus AMD optical coherence tomography image of the present application.

[0045] Figure 4 It is a flowchart of the method for constructing the human fundus OCT image data set provided by the embodiments of the present application.

[0046] Figure 5 It is a schematic diagram of the optical coherence tomography system structure provided by the embodiments of the present application.

[0047] Figure 6 A three-dimensional segmentation algorithm based on the optical coherence tomography image of the human fundus AMD of the application is used in the comparison chart of the test set segmentation sample.

[0048] Figure 7 The recognition result schematic diagram of three preferred embodiments in the clinical practice of the three-dimensional segmentation algorithm based on the optical coherence tomography image of the human fundus AMD of the application.

[0049] Figure 8 The system structure schematic diagram of the initial model training provided by the embodiment of the application.

[0050] Figure 9 The structure block diagram of an electronic device provided by the embodiment of the application. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. It should be understood that the drawings in the application only serve the purpose of description and illustration, and are not used to limit the protection scope of the application. In addition, it should be understood that the schematic drawings are not drawn according to the actual proportions. The flowcharts show the operations implemented according to some embodiments of the application. It should be understood that the operations of the flowcharts can not be implemented in sequence, and the steps without logical context relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowcharts or removed from the flowcharts by those skilled in the art under the guidance of the content of the application.

[0052] In addition, the described embodiments are only some of the embodiments of the application, not all the embodiments. The components of the embodiments of the application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the application. Based on the embodiments of the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0053] It should be noted that the term "comprising" will be used in the embodiments of the application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0054] Figure 1 The application scenario schematic diagram structure schematic diagram of the three-dimensional segmentation method based on the optical coherence tomography image of the human fundus AMD of the application. As Figure 1As shown, the present application is applicable to the scene of obtaining OCT images of any sample. The sample may be a biological sample, etc. The scene involves an optical coherence tomography (OCT) system and an electronic device. First, the OCT system scans any sample to output an initial fundus OCT image, and then the initial diseased fundus OCT image is input into the electronic device. The electronic device uses a trained segmentation deep learning network model to perform three-dimensional segmentation on the fundus OCT image, and finally obtains a three-dimensional lesion area result.

[0055] Figure 2 A flowchart of a three-dimensional segmentation method based on human fundus AMD optical coherence tomography images according to the present application is shown in Figure 1. As shown, the method is applied in the electronic device described above. The electronic device may be a desktop computer, a notebook computer, etc. having computing processing capability and display function, or it may also be a server. The method comprises the following steps: Figure 2

[0056] S101, obtaining an initial diseased fundus OCT image output by an optical coherence tomography system.

[0057] Optionally, the OCT system can include a light source, a coupler, a reference mirror, a sample arm, a reference arm, and a spectrometer. The specific principle is as follows: the system can use a broadband light source as the light source, for example, a broadband light source with a center wavelength of 850 nm and a full width at half maximum of 165 nm. The light emitted by the broadband light source passes through the coupler and is irradiated to the sample and the reference mirror through the sample arm and the reference arm, respectively. The reflected light in the two light paths converges in the coupler under certain conditions to produce an interference signal. The interference signal is output to the spectrometer, and the spectrometer transmits the detected interference signal to the computer. The computer performs gray scale imaging and outputs an initial OCT image. The initial fundus OCT image has an AMD lesion area.

[0058] S102, inputting the initial fundus OCT image into a pre-trained deep learning network model for three-dimensional segmentation of fundus AMD high reflection area, and performing three-dimensional segmentation on the initial fundus OCT image by the deep learning network model to obtain a three-dimensional AMD high reflection lesion area image.

[0059] ​Optionally, the three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus is trained according to the method provided in the present application. The three-dimensional segmentation deep learning network model trained can be combined with the OCT system in step S01, and the initial fundus OCT image obtained in step S01 can be input into the three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus, so as to perform three-dimensional segmentation on the initial fundus OCT image by the three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus, and obtain the result of three-dimensional AMD high-reflection area.

[0060] The embodiment obtains an initial fundus optical coherence tomography image output by an optical coherence tomography system, inputs the initial fundus optical coherence tomography image into a three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus pre-trained, performs three-dimensional segmentation on the initial fundus optical coherence tomography image by the three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus, and obtains the result of three-dimensional AMD high-reflection area.

[0061] Optionally, the speckle denoising deep learning network model is a structure based on an improved U-Net, and sequentially connected are an encoder, a hollow pyramid pooling module, and a decoder. The encoder and the decoder are connected by a skip connection containing a Squeeze-and-excitation module.

[0062] Optionally, step S102 inputs the initial fundus OCT image into the three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus pre-trained, performs three-dimensional segmentation on the initial fundus OCT image by the three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus, and obtains the result of three-dimensional AMD high-reflection area, which can include:

[0063] Optionally, the initial fundus OCT image containing the lesion area is input into the encoder of the three-dimensional segmentation deep learning network model for AMD high-reflection area of fundus, the initial fundus OCT image is down-sampled by the convolution layer and the pooling layer in the encoder, and the feature image extracted by down-sampling is input into the hollow pyramid pooling module to further extract multi-size information, and then the output feature image is sent to the decoder to be up-sampled by the inverse convolution layer and the pooling layer in the decoder, so as to restore to the original size of the image, and then the output image is output after channel attention module for channel weight redistribution, and finally the output image is the target lesion image.

[0064] Figure 3A system structure schematic diagram of a three-dimensional segmentation method based on an optical coherence tomography image of an AMD of a human fundus is provided.

[0065] Optionally, the initial fundus OCT image is input into the three-dimensional segmentation deep learning network, and after five mixed down-samplings, a low-resolution feature map is obtained; then the feature map is further extracted by using a hollow pyramid pooling module, so that feature information of different sizes is learned; the feature map output by the hollow pyramid pooling module is input into the decoder part, and after five mixed up-samplings, the feature map is restored to the same size as the initial input image. Finally, a channel attention module is used to enhance the channel of the feature map useful for the current segmentation task and suppress the feature channel not useful for the current task, and finally the segmentation result is obtained.

[0066] The self-defined mixed down-sampling module is specifically composed of two layers of conventional convolution normalization layers, a convolution layer with a step of 2, and a layer of conventional convolution normalization layers connected in sequence. After the first two layers of convolution normalization operations, the input features are added to the self-defined mixed down-sampling module through a skip connection, so that the self-defined mixed down-sampling module forms a residual structure.

[0067] The mixed up-sampling module is specifically composed of a layer of conventional convolution normalization layers, a layer of transposed convolution layers, and a layer of conventional convolution normalization layers connected in sequence. After the transposed convolution realizes up-sampling, the result is spliced and fused with the value of the corresponding skip connection, and then convolution normalization is performed.

[0068] The skip connection connecting the encoder and the decoder is specifically that a Squeeze-and-excitation module is added at the skip connection, which can avoid the gradient disappearance phenomenon to a certain extent.

[0069] Figure 4 A flowchart of a method for constructing an OCT image dataset of a human fundus is provided for the embodiments of the present application.

[0070] Figure 5 A schematic diagram of the structure of an optical coherence tomography system is provided for the embodiments of the present application.

[0071] As shown in Figure 5 The OCT system includes a spectrometer, a fiber coupler, a collimating mirror, a first lens, a second lens, a third lens, a two-dimensional scanning galvanometer, a first polarization controller, a second polarization controller, and a mirror.

[0072] Optionally, the OCT system can be used to scan the fundus of each AMD patient to obtain a plurality of initial fruit fly heart OCT images corresponding to each sample.

[0073] For example, the plurality of initial Drosophila heart OCT images can be 100 initial fundus OCT images.

[0074] S302, respectively, each initial fundus OCT image as an initial fundus OCT image in the image pair (referred to as an initial image), according to the selected fundus AMD high reflection area target segmentation region, the AMD high reflection area target segmentation region on each initial image is manually labeled as the ground truth image in the image pair, and each initial image and the corresponding ground truth image are combined into an image pair.

[0075] Optionally, for each sample, the corresponding plurality of initial images and the ground truth image corresponding to the initial object are combined to obtain an image pair, and by combining other initial images with their corresponding ground truth images, a plurality of image pairs are obtained.

[0076] For example, an AMD patient sample corresponds to 100 initial images, and the sample can have 100 image pairs.

[0077] S303, integrate the image pairs corresponding to each fundus sample to obtain a fundus optical coherence tomography image dataset.

[0078] The dataset obtained by combining all image pairs corresponding to all sample objects is the fundus AMD optical coherence tomography image dataset, and a predetermined proportion of the fundus AMD optical coherence tomography image dataset can be used as a training dataset and a test dataset, for example, 90% of the fundus AMD optical coherence tomography image dataset can be used as a training dataset, and 10% of the fundus AMD optical coherence tomography image dataset can be used as a test dataset.

[0079] For example, if there are 100 sample objects, each sample object can obtain 100 image pairs corresponding to each sample object through the above S301-S303, and the Drosophila heart sample optical coherence tomography image dataset includes 100x100 image pairs.

[0080] Figure 6 A three-dimensional segmentation algorithm based on human fundus AMD optical coherence tomography images of the present application is a test set segmentation sample comparison chart. Figure 4 From left to right are the original fundus OCT image (left), the predicted 3D lesion image (middle), and the 3D lesion image manually labeled by an ophthalmologist (right). It can be seen that the three-dimensional prediction results not only approach the actual morphology of the lesion, but also retain the surface details, which is conducive to AMD diagnosis.

[0081] Figure 7The recognition result schematic diagram of nine preferred embodiments of a three-dimensional segmentation algorithm based on the human fundus AMD optical coherence tomography image in the clinical practice of the application. Figure 5 The nine three-dimensional fundus AMD high reflection area segmentation graphs are all from different AMD patients. It can be seen that the three-dimensional shapes and sizes of the fundus AMD high reflection areas of different patients are different. This proves the robustness of the three-dimensional segmentation method based on the human fundus AMD optical coherence tomography image provided in the embodiments of the application.

[0082] Figure 8 The system structure schematic diagram of the initial model training provided in the embodiments of the application.

[0083] Figure 9 The structure block diagram of an electronic device 300 provided in the embodiments of the application. As shown in the figure, the electronic device can include a processor 301, a memory 302. Figure 9

[0084] Optionally, a bus 303 can also be included, wherein the memory 302 is used to store machine readable instructions executable by the processor 301, when the electronic device 300 is running, the processor 301 communicates with the memory 302 through the bus 303, and the machine readable instructions are executed by the processor 601 to execute the method steps in the above method embodiments.

[0085] The embodiments of the application also provide a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to execute the method steps in the above three-dimensional segmentation fundus AMD high reflection area method embodiments.

[0086] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and device can refer to the corresponding process in the method embodiments, and will not be repeated in the application. In the several embodiments provided in the application, it should be understood that the disclosed system, device and method can be implemented by other ways. The device embodiments described above are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed mutual elements can be indirect coupling or communication connection through some communication interface, device or module, which can be electrical, mechanical or other forms.

[0087] ​In addition, each function unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. When the functions are realized in the form of software function units and sold or used as an independent product, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0088] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A three-dimensional segmentation method for the high-reflectivity area (AMD) of the human fundus based on deep learning, characterized in that, include: Acquire initial fundus OCT images of AMD patients, wherein AMD high reflectivity regions are present in the initial fundus OCT images of AMD patients; The initial fundus OCT image of the AMD patient is input into a pre-trained three-dimensional segmentation deep learning model, which performs three-dimensional segmentation on the initial fundus OCT image to obtain the corresponding three-dimensional result of the fundus AMD high reflectivity region. The 3D segmentation deep learning model is a U-Net structure composed of an encoder and a decoder connected in sequence; Before inputting the initial fundus OCT images of the AMD patient into the pre-trained three-dimensional segmentation deep learning model, the following steps are also included: A dataset of fundus OCT images of initial AMD patients was constructed, which contained multiple data pairs, each containing a raw fundus OCT image and a ground truth image with manually annotated high reflectivity areas of fundus AMD. An initial model is constructed, comprising: an improved U-Net-based structure, in which an encoder, a hollow pyramid pooling module, and a decoder are connected in sequence; wherein the encoder and decoder are connected by a skip connection containing a Squeeze-and-excitation module; The initial model is trained based on the sample dataset to obtain the three-dimensional segmentation deep learning model; The construction of the initial model includes: U-Net is used as the backbone of the initial model, with skip connections between the encoder and decoder. The encoder consists of a custom hybrid downsampling module and a Squeeze-and-Excitation module. The decoder consists of a deconvolutional layer and a series of regular convolutional activation layers, connected to the corresponding parts of the encoder via skip connections. The decoder upsamples the input feature map to restore it to its original size, and then uses a channel attention module to adaptively calculate the attention weights of each channel feature to adjust the utilization of different features. Each skip connection in the model also passes through a Squeeze-and-Excitation module. The custom hybrid downsampling module is a dense residual structure that integrates conventional convolutional layers and ReLU activation functions; after feature extraction by the five layers of the custom hybrid downsampling module, a dense low-resolution feature map is generated from the input image. After receiving the low-resolution feature map generated by the custom hybrid downsampling module, the Squeeze-and-Excitation module adaptively corrects the feature intensity of each channel in the feature map. After the decoder generates a low-resolution feature map, a hollow pyramid pooling module is used to further mine feature information, and then it is input into the decoder for upsampling; the channel attention module after the encoder makes the weight of each channel in the upsampled feature map redistribute, filtering out unnecessary features while maintaining the original region features. The process of constructing the sample dataset includes: Obtain an initial OCT image dataset of fundus images from AMD patients, wherein the initial OCT images of fundus images from AMD patients contain AMD high reflectivity regions to be segmented; Preprocess the dataset, including but not limited to dataset augmentation, adjusting image parameters, cropping image size, and manually annotating target regions; The data pairs corresponding to each sample are combined and divided into training set, test set and validation set; The process of training the initial model based on the sample dataset to obtain the 3D segmentation deep learning model includes the following steps: Step 1: Train the network model using the obtained training set, with the ratio of training, test, and validation sets being 8:1:

1. Input the preprocessed initial fundus OCT images of AMD patients into the network and extract features through a custom hybrid downsampling module to obtain high-level and low-level semantic features. The high-level semantic features are then input into a dilated pyramid pooling module for further extraction of high-level semantic information. These features are independently processed through five dilated convolutional layers, and the resulting five feature maps are merged and combined. Finally, a 1-bit multiplication table is used to perform the final merging.

1. Convolutional compression features; Step two: Upsample the feature map obtained in step one, then find the low-level semantic feature map with the same resolution after upsampling, and process it through 3... The feature map is composed of three convolutional and Squeeze-and-Excitation modules, which are then concatenated and merged. This operation is repeated four times until the final output size is restored to the original input size, thus completing the extraction of spatial information. Finally, this feature map is passed through a channel attention module, which redistributes the weights of each channel according to their priority, and finally the segmentation prediction result is obtained. Step 3: Calculate the segmentation loss based on the obtained prediction results; Step four: Repeat steps one through three, backpropagate the segmentation loss and update the network parameters. When the segmentation loss meets the preset conditions, save the network weight parameters. This model is the three-dimensional segmentation deep learning model.

2. The three-dimensional segmentation method for the high reflectivity region (AMD) of the human fundus based on deep learning according to claim 1, characterized in that, The step of calculating the segmentation loss based on the obtained prediction results includes: Based on the predicted results and the ground truth image, the model loss function is calculated using the following formula. 。

Images