Optic disk image segmentation method and system, computer equipment and storage medium

Abstract

The invention provides an optic disc segmentation method. The optic disc segmentation method comprises the following steps: acquiring an optic disc image; an encoder is used for extracting multi-scale features of the optic disk image, the encoder comprises a plurality of variable waveform convolution blocks FWC Block, and each variable waveform convolution block FWC Block dynamically adjusts a receptive field through a dynamic convolution layer AKConv and performs feature compression in combination with a wavelet down-sampling layer HWD; a decoder is used for gradually recovering the resolution of the feature map, the decoder comprises a plurality of selective double attention fusion gates SDA Gate, and the features of the encoder and the decoder are fused through a channel attention and space attention mechanism to obtain a fused feature map; and obtaining an optic disc image segmentation result according to the fused feature map.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence medical technology, and in particular to an eye image processing technology. Background Technology

[0002] Myopia, a prevalent eye disease, not only affects an individual's visual function but can also induce various eye health problems. In certain populations, myopia can evolve into a more severe pathological condition known as pathological myopia. Besides significant high myopia, pathological myopia is more likely to cause fundus lesions, including retinal detachment, macular degeneration, and optic disc structural abnormalities. Changes in fundus structures, especially abnormalities in the optic disc region, are crucial for assessing and monitoring the progression of pathological myopia. In cases of pathological myopia, due to the continued growth of the eyeball, the optic disc may undergo morphological changes such as tilting, distortion, or increased concavity (specifically, increased concavity in the central portion of the optic disc). These changes can not only induce or exacerbate visual impairment but may also become risk factors for other eye diseases such as glaucoma. Therefore, meticulous observation and in-depth analysis of patients with pathological myopia are particularly critical.

[0003] To meet this clinical need, optic disc image segmentation technology was developed. By precisely segmenting the optic disc region from fundus images, doctors can comprehensively and meticulously assess the morphological characteristics and lesion status of the optic disc. The application of optic disc image segmentation technology, especially when combined with advanced computer vision and machine learning algorithms, has significantly improved the objectivity and accuracy of the assessment process.

[0004] However, traditional image segmentation techniques often fail to deliver ideal results when applied to images of optic disc lesions due to the atypical nature of image edge morphology. Furthermore, medical image segmentation requires significantly higher accuracy than ordinary image segmentation. Common downsampling methods (such as average pooling and max pooling), while reducing computational complexity, often lead to the loss of feature edge information, thus affecting segmentation accuracy. This feature loss problem is particularly prominent in medical image analysis, and current technologies urgently need to address this issue to improve the precise localization capability of the optic disc region. Summary of the Invention

[0005] This application aims to propose an optic disc image segmentation scheme and presents an efficient fully convolutional optic disc segmentation network, named FDAU-Net, which includes flexible wavelet convolution and dual attention modules. With its flexibility and low computational cost, this network improves segmentation accuracy and can handle irregularly shaped lesion regions. The FDAU-Net network mainly consists of a flexible wavelet convolution block (FWC block) and a selective dual attention fusion gate (SDA gate). The FWC block is used to capture contextual information and achieve downsampling, while the SDA gate further optimizes feature selection through multi-level feature fusion and attention mechanisms, thereby improving the accurate localization capability of the optic disc region. The disc image segmentation method based on this includes: acquiring a disc image; extracting multi-scale features of the disc image using an encoder, wherein the encoder contains multiple variable waveform convolutional blocks (FWC blocks), each FWC block dynamically adjusts its receptive field through a dynamic convolutional layer (AKConv) and combines it with a wavelet downsampling layer (HWD) for feature compression; progressively restoring the feature map resolution using a decoder, wherein the decoder contains multiple selective dual attention fusion gates (SDA gates), and fuses the features of the encoder and decoder through channel attention and spatial attention mechanisms to obtain a fused feature map; and obtaining the disc image segmentation result based on the fused feature map.

[0006] Optionally, the encoder includes:

[0007] The feature extraction block Stem performs preliminary processing on the input image to extract low-level features;

[0008] The five variable waveform convolutional blocks FWCBlock are sequentially connected from top to bottom after the feature extraction block Stem, with each layer having the following output feature channels: C1 to C5: 16, 32, 128, 160, and 256. Optionally, the variable waveform convolutional block FWC Block includes a dynamic convolutional layer AKConv, a depthwise convolutional layer, an LN layer, a pointwise convolutional layer, a BN layer, a SiLU layer, a pointwise convolutional layer, a BN layer, and a GSiLU layer connected in sequence.

[0009] Among them, an inverted bottleneck structure is used between the two sets of pointwise convolutional layers, which first increases the number of channels by four times and then compresses and restores them;

[0010] Among them, the first to fourth layers of variable waveform convolutional blocks (FWC Blocks) also include wavelet downsampling layers (HWD) connected after the GSiLU layers;

[0011] The wavelet downsampling layer HWD performs feature downsampling and channel depth enhancement simultaneously. A set of features is output before the wavelet downsampling layer HWD for skip connections, and a 1×1 convolutional layer is used to adjust the number of channels to achieve dimensionality consistency with the features of the subsequent decoder. The 1×1 convolutional layer after each variable waveform convolutional block (FWC Block) is skip connected to a selective dual attention fusion gate (SDA Gate) of the decoder.

[0012] Optionally, the variable waveform convolutional block (FWC Block) is defined as follows:

[0013] f l '=DWConv(AKConv(f l-1 ))+f l-1 ;

[0014] f″ l =σ1(BN{PWConv dim1→4 (LN(f′ l ))});

[0015] f″′ l =σ2(BN{PWConv dim4→1 (f″ l )})+f′ l ;

[0016] f l =HWD(f″′) l );

[0017] Among them, f l This represents the feature map output by the Lth layer variable waveform convolutional block (FWC Block); σ1 represents the SiLU activation function, σ2 represents the GSiLU activation function, BN represents batch normalization, LN represents layer normalization, DWConv represents depthwise convolution, and P... W Conv represents pointwise convolution.

[0018] Optionally, the decoder corresponds to the encoder in terms of the number of layers, and the decoder shown includes five general convolutional blocks from bottom to top; a selective dual attention fusion gate SDAGate is embedded before the second to fifth general convolutional blocks;

[0019] The feature map output by the variable waveform convolutional block (FWC Block) of the fifth layer of the encoder is used as the input of the general convolutional block of the first layer of the decoder. Each layer of the decoder obtains a deep feature image through upsampling and inputs it into the selective dual attention fusion gate (SDA Gate) of the next layer of the decoder.

[0020] The selective dual attention fusion gate (SDA Gate) of each layer of the decoder fuses the shallow features of the encoder with the upsampled deep features from the previous layer of the decoder to obtain a fused feature image, which is then passed through the general convolutional block to reduce the number of channels in the fused feature image.

[0021] The fused feature image output by the general convolutional block of the fifth layer of the decoder is the result of the visual disk image segmentation.

[0022] Optionally, the fused feature map obtained by fusing the features of the encoder and decoder through channel attention and spatial attention mechanisms includes:

[0023] For the second to fifth layers of the decoder, the shallow features from the skip connections after the 1×1 convolutional layer and the deep features from the upsampled layer above are input into the selective dual attention fusion gate SDAGate. The shallow features and the deep features are processed first by convolution, batch normalization (BN), and GELU activation functions, respectively.

[0024] Then, the two features are concatenated to obtain a concatenated feature map, and the feature dimensions are adjusted through a reshaping operation;

[0025] The stitched feature map is input into the Selective Channel Attention (SC-A) module. The SC-A module uses max pooling and average pooling to extract global information from the stitched feature map and generate global features representing each channel. Then, the channel weights are calculated using a multilayer perceptron (MLP) and normalized using the Softmax function to obtain the channel attention map.

[0026] The channel attention map and the spliced ​​feature map are multiplied element-wise, and the result is summed element-wise with the processed deep features to obtain the channel-weighted fused features.

[0027] The channel-weighted fused features are input into the spatial attention (SA) module. The spatial attention (SA) module first performs max pooling and average pooling operations on the input feature map to generate two spatial feature maps in the channel dimension. Then, global spatial information is combined through a convolutional layer, BN, and GELU activation function. The generated feature map is normalized using the Sigmoid function to obtain the spatial attention map.

[0028] The spatial attention map and the fused features are multiplied element-wise to obtain the multiplication result;

[0029] The result of the multiplication is then summed element-wise with the fused features to obtain the final fused feature map;

[0030] The fused feature map is reduced in feature dimension by a general convolutional block in the fifth layer of the decoder, and then the output is the visual disc image segmentation result.

[0031] Optionally, the selective dual attention fusion gate SDAGate is defined as follows:

[0032] f′ s =σ1(BN(Conv(f s )))

[0033] f′ d =f′ s +σ1(BN(Conv(f d )))

[0034] f c&r =Reshape H×W×C×2 (Concat[f′ s ;f′ d ])

[0035] f′ c&r =MLP(MaxPool(Sum(f c&r )))+MLP(AvgPool(Sum(f c&r )))

[0036] SC-AMap = Reshape 1×1×C×2 (σ3(f′ c&r ))

[0037] f sc-α =f c&r ⊙SC-AMap+f′ d

[0038] S-AMap=σ2(σ1(BN(Conv(Concat[MaxPool(f sc-a ); AvgPool(f sc-a )]))))

[0039] f out =f sc-a ⊙S-AMap+f sc-a

[0040] Among them, f d f represents the deep features after upsampling. s shallow features representing skip connections, f outσ1, σ2, and σ3 represent the output features of the Selective Dual Attention Fusion Gate (SDA Gate); σ1, σ2, and σ3 represent the GELU activation function, the Softmax function, and the Sigmoid function, respectively; BN represents batch normalization; MaxPool represents max pooling; AvgPool represents average pooling; MLP represents a multilayer perceptron with one hidden layer, where the feature map size of the hidden layer is C / r×1×1, and r is the scaling ratio; Conv represents convolution; concat represents concatenation; Reshape represents reshaping; and ⊙ represents dot product.

[0041] According to another aspect of this application, this application also provides a video disc splitting system, comprising:

[0042] The acquisition module acquires the video disc image;

[0043] The encoder module is used to extract multi-scale features of the visual disc image. The encoder contains multiple variable waveform convolutional blocks (FWC Blocks). Each FWC Block dynamically adjusts its receptive field through a dynamic convolutional layer (AKConv) and combines it with a wavelet downsampling layer (HWD) for feature compression.

[0044] The decoder module uses a decoder to gradually restore the feature map resolution. The decoder contains multiple selective dual attention fusion gates (SDA Gates) and obtains a fused feature map by fusing the features of the encoder and decoder through channel attention and spatial attention mechanisms.

[0045] The output module obtains the visual disk image segmentation result based on the fused feature map.

[0046] According to another aspect of this application, this application also provides a computer device, comprising:

[0047] A processor coupled to a memory storing computer-executable instructions, wherein the processor executes the computer-executable instructions stored in the memory, causing the processor to perform a video disc splitting method.

[0048] According to another aspect of this application, a computer-readable storage medium is also provided.

[0049] Used to store computer programs, the computer programs including instructions for implementing the video disc splitting method.

[0050] This application proposes an efficient fully convolutional network, FDAU Net, for optic disc segmentation. FDAU-Net integrates a flexible wavelet convolutional block (FWC block) that combines AKConv's dynamic receptive field adjustment with wavelet downsampling to achieve efficient feature extraction while preserving crucial edge information. Furthermore, we design a selective dual attention fusion gate (SDA gate) to optimize feature selection and fusion, significantly improving segmentation accuracy and robustness by utilizing channel and spatial attention mechanisms. To further enhance performance, we introduce MedAugment, an automated data augmentation method tailored for medical images, effectively increasing data diversity even with limited datasets. Experimental results on the iChallenge and IDRiD datasets demonstrate that FDAU-Net consistently achieves superior segmentation performance across key metrics, highlighting its potential for advancing clinical applications. Attached Figure Description

[0051] Figure 1 A schematic flowchart illustrating the disc image segmentation method provided in this application embodiment;

[0052] Figure 2 This is a schematic diagram of the overall network structure of the FADU-Net network model provided in the embodiments of this application;

[0053] Figure 3 A schematic diagram of the network structure of the Variable Waveform Convolutional Block (FWC Block) provided in the embodiments of this application;

[0054] Figure 4 A schematic diagram of the network structure of the Selective Dual Attention Fusion Gate (SDA Gate) provided in an embodiment of this application;

[0055] Figure 5 Examples of dataset sample images provided in embodiments of this application;

[0056] Figure 6 Examples of dataset image annotation samples provided in embodiments of this application;

[0057] Figure 7 Examples of dataset image enhancement samples provided in embodiments of this application;

[0058] Figure 8 A comparison diagram of the image segmentation results provided in this embodiment with those of other U-shaped networks. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] As one application scenario, the technical solution of this application is applied to ophthalmoscopy equipment, such as... Figure 1 As shown, the detailed steps are as follows:

[0061] Step S101: Establish the transmission relationship between the ophthalmoscope and the terminal device. This step is fundamental to the ocular data processing and transmission method, ensuring stable and efficient communication between the ophthalmoscope and the terminal device. In practice, wireless communication technologies such as Bluetooth and Wi-Fi are typically used to establish the connection between the two, ensuring the real-time performance and accuracy of data transmission.

[0062] Step S102: Receive the shooting request signal sent by the terminal device. This step is the starting point of the entire eye data processing and transmission method. The terminal device sends a shooting request signal to the ophthalmoscope, triggering the subsequent image acquisition and processing flow. The shooting request signal may include parameters such as shooting mode, exposure time, and focal length to ensure that the quality of the acquired fundus images meets the requirements of medical diagnosis.

[0063] Step S103: Receive the imaging control signal. After the imaging request signal is sent, an imaging control signal is generated via the imaging control button on the ophthalmoscopy to guide the ophthalmoscopy in capturing fundus images. The imaging control signal may include operations such as triggering imaging, adjusting focus, and changing the brightness of the light source to ensure stability and accuracy during image acquisition.

[0064] Step S104: Based on the shooting request signal and shooting control signal, control the image sensor to acquire at least one fundus image. Upon receiving the shooting request and shooting control signal, the image sensor inside the ophthalmoscopy will capture a fundus image according to preset parameters. The high resolution and fast response characteristics of the image sensor make the acquired fundus images clear and detailed, which helps doctors make accurate diagnoses.

[0065] Step S105: The fundus image is input into the processor, which calls a preset segmentation model to perform optic disc image segmentation on the fundus image, obtaining the optic disc image segmentation result. Fundus images contain a large amount of physiological information, and the optic disc, as a key part, is of great significance for the diagnosis of ophthalmic diseases. Therefore, the processor uses a pre-trained segmentation model to segment the optic disc region of the fundus image and extract the image information of the optic disc. The accuracy and efficiency of this step are crucial for subsequent diagnostic analysis. In this application, we innovatively load a self-developed image segmentation algorithm model into the ophthalmoscopy processor for efficient optic disc image segmentation.

[0066] Step S106: In response to the shooting request signal, the optic disc image segmentation result and the fundus image are returned to the terminal device. After completing the optic disc image segmentation, the ophthalmoscope sends the segmentation result and the original fundus image back to the terminal device. The doctor can view these images on the terminal device and perform further diagnostic analysis. In addition, the terminal device can also store, print, or share the received image data to meet the needs of different scenarios.

[0067] Through the detailed description of the above steps, we can see that the application of ocular data processing and transmission methods in ophthalmoscopy equipment is highly practical and advanced. It not only enables the rapid acquisition and transmission of fundus images but also utilizes advanced image processing technology to accurately analyze and interpret these images, providing strong diagnostic support for ophthalmologists.

[0068] As an optional embodiment:

[0069] The image capture request signal contains a unique identifier for the patient;

[0070] The processor marks the optic disc image segmentation result and the captured fundus image based on the patient's unique identification and the ophthalmoscope's unique device identification; that is, after completing the optic disc image segmentation, the processor marks the segmentation result together with the captured fundus image to ensure the correlation and integrity between the two.

[0071] Specifically, after returning the optic disc image segmentation result and the fundus image to the terminal device, the terminal device establishes a correspondence between the patient's file and the optic disc image segmentation result and the fundus image based on the patient's unique identifier, and records the unique device identifier of the ophthalmoscope used to capture the fundus image.

[0072] Specifically, when the image capture request signal is sent to the processor, the processor labels the optic disc image segmentation results and the captured fundus image based on the patient's unique identifier and the ophthalmoscopy's unique device identifier in the signal. This step is crucial because it ensures that each patient's medical data is closely linked to their identity, and also helps doctors quickly and accurately identify patient information during subsequent diagnostic processes.

[0073] After labeling is completed, the optic disc segmentation results and fundus images are returned to the terminal device. On the terminal device, based on the patient's unique identifier, the system creates a detailed file for the patient. This file not only records the patient's personal information but also includes their eye examination data, such as optic disc segmentation results and fundus images. Simultaneously, the file also records the unique device identifier of the ophthalmoscopy used to capture the fundus images, facilitating subsequent traceability and management.

[0074] It is worth mentioning that by establishing patient files and recording device identifiers, medical institutions can not only better manage patient medical information but also improve the accuracy and efficiency of diagnosis. When doctors need to review a patient's historical records, they can easily find relevant information through the patient's unique identifier, thus gaining a more comprehensive understanding of the patient's condition. Furthermore, recording device identifiers also helps medical institutions manage and maintain equipment, ensuring its proper operation and the accuracy of examinations.

[0075] As an optional embodiment:

[0076] Before the processor calls the preset segmentation model to perform optic disc image segmentation on the fundus image, the method further includes: when multiple fundus images are captured, using the processor to filter the multiple fundus images for image quality and image content, and selecting one fundus image that meets preset requirements from the fundus images as the fundus image input to the preset segmentation model; when only one fundus image is captured, using the processor to judge the image quality and image content of the fundus image, if it does not meet the requirements, issuing a voice prompt to recapture the image, and if it meets the requirements, using that fundus image as the fundus image input to the preset segmentation model.

[0077] In the entire process of fundus image processing, image quality and content selection is a crucial step. This step determines the accuracy of subsequent optic disc image segmentation and therefore must be performed rigorously and meticulously. When capturing multiple fundus images, we need to utilize advanced processing technology to strictly select these images.

[0078] First, the processor performs an image quality assessment on each fundus image. This includes checking the image's sharpness, contrast, brightness, and the presence of interfering factors such as blurring and noise. Using advanced image processing algorithms, the processor can automatically determine whether the image quality meets preset requirements. If the quality of a fundus image does not meet the requirements, the processor will exclude it from subsequent processing.

[0079] Secondly, the processor filters the content of the fundus images. This mainly involves checking whether the images contain complete fundus structures, such as the optic disc and blood vessels. It also needs to determine if there are any abnormalities such as lesions or hemorrhages in the images. Through analysis of the image content, the processor can select the fundus image that best meets the preset requirements as input for subsequent optic disc image segmentation.

[0080] When only one fundus image is captured, the processor still assesses the image quality and content. If the fundus image does not meet the preset requirements, the processor will issue a voice prompt instructing a recapture. This helps to promptly identify and correct problems during the capture process, ensuring the accuracy and reliability of subsequent processing. If the fundus image meets the requirements, the processor directly uses it as the fundus image input to the preset segmentation model.

[0081] The above application scenarios are merely examples and are not the only limitations of this invention. In practical applications, the methods and systems of this invention can be applied to various scenarios requiring optic disc segmentation, including but not limited to ophthalmic disease diagnosis, fundus image analysis, and medical image processing.

[0082] For example, in the diagnosis of ophthalmic diseases, doctors need to carefully analyze the patient's fundus images to identify possible lesions or abnormalities. By applying the method and system of this invention, doctors can acquire images of the optic disc region more quickly and accurately, thereby improving the efficiency and accuracy of diagnosis.

[0083] Furthermore, optic disc segmentation is also an important task in the field of medical image processing. By segmenting the optic disc in fundus images, more information about the eye structure can be obtained, providing strong support for further analysis and processing of medical images.

[0084] The technical solution of this application will be described in detail below, such as... Figure 1 As shown, this application provides a video disc segmentation method, including the following steps:

[0085] Step S101: Acquire the visual disk image and input it into the FADU-Net network model.

[0086] like Figure 2As shown, the FADU-Net network model adopts a classic U-shaped design in its overall structure, which is divided into two parts: encoder and decoder. Multi-scale feature fusion is achieved by skipping connections.

[0087] Step S102: Use an encoder to extract multi-scale features of the visual disc image. The encoder contains multiple variable waveform convolutional blocks (FWC Blocks). Each FWC Block dynamically adjusts its receptive field through a dynamic convolutional layer (AKConv) and combines it with a wavelet downsampling layer (HWD) for feature compression.

[0088] Step S103: Use the decoder to gradually restore the feature map resolution. The decoder contains multiple selective dual attention fusion gates (SDA Gates). The features of the encoder and decoder are fused through channel attention and spatial attention mechanisms to obtain a fused feature map.

[0089] Step S104: Obtain the visual disk image segmentation result based on the fused feature map.

[0090] The encoder includes:

[0091] The feature extraction block Stem performs preliminary processing on the input image to extract low-level features;

[0092] The five variable waveform convolutional blocks FWCBlock are sequentially connected from top to bottom after the feature extraction block Stem. The number of output feature channels for each layer is C1 to C5: 16, 32, 128, 160, and 256.

[0093] The Stem block performs preliminary processing on the input image, while each FWC block has feature extraction and downsampling functions, laying the foundation for multi-scale feature extraction.

[0094] like Figure 3 As shown, the Variable Waveform Convolutional Block (FWC Block) comprises a dynamically convolutional layer AKConv, a depthwise convolutional layer, an LN layer, a pointwise convolutional layer, a BN layer, a SiLU layer, a pointwise convolutional layer, a BN layer, and a GSiLU layer connected in sequence. The FWC Block is composed of several layers of different types connected in sequence, each layer having its specific function.

[0095] AKConv (Adaptive Kernel Convolution): This is a convolutional layer that can dynamically adjust the kernel size based on the input data, enabling the network to better adapt to different input features.

[0096] Deep convolutional layer: This is a traditional convolutional layer used to extract deep features from input data.

[0097] LN layer (Layer Normalization): This layer is used to normalize the data, helping to stabilize the training process and improve the model's generalization ability.

[0098] Pointwise Convolution: This is a special type of convolution operation with a kernel size of 1x1, mainly used to change the number of channels in the feature map.

[0099] BN (Batch Normalization) layer: This layer accelerates the training process and reduces internal covariate bias by normalizing the mean and variance of the layer inputs.

[0100] siLU layer (Scaled Exponential Linear Units): This is an activation function used to increase the non-linearity of the network, helping the network learn complex features.

[0101] Further pointwise convolutional layers: used to further adjust the number of channels in the feature map.

[0102] The second BN layer: used again for normalization.

[0103] GSiLU layer (Gated siLU): This is an activation function with a gating mechanism, which can more effectively control the flow of information and improve model performance.

[0104] The entire FWC Block, through the combination of these layers, is capable of performing complex transformations and feature extraction on the input data.

[0105] In this design, an inverted bottleneck structure is used between the two sets of pointwise convolutional layers. The number of channels is first increased fourfold, and then compressed back to its original size. This inverted bottleneck structure is designed to increase the network's non-linear expressive power while controlling computational complexity. In this structure, a "pointwise convolutional layer" means that each output pixel in the convolution operation is calculated from only a single input pixel. Unlike traditional convolution operations, pointwise convolution does not consider the relationships between pixels.

[0106] The inverted bottleneck structure between the two sets of pointwise convolutional layers means that the number of input channels is increased to four times the original number before performing pointwise convolution. This process can be understood as adding more "paths" or "channels" to the network, enabling it to capture more feature information. The subsequent compression then reduces the number of these increased channels to a smaller number, removing redundant information while retaining key features. The inverted bottleneck structure enhances feature extraction capabilities by first increasing the number of channels, and then reduces computation and prevents overfitting through compression.

[0107] Compared to U-Net and its derivative models, the FWC block uses wavelet downsampling instead of traditional pooling, preserving key details while reducing feature resolution. Compared to conventional pooling, wavelet downsampling can effectively preserve edge and texture information, providing higher-quality feature information for subsequent networks, thereby improving overall segmentation performance.

[0108] Among them, the first to fourth layers of variable waveform convolutional blocks (FWC Blocks) also include wavelet downsampling layers (HWD) connected after the GSiLU layers;

[0109] The wavelet downsampling layer HWD performs feature downsampling and channel depth enhancement simultaneously. A set of features is output before the wavelet downsampling layer HWD for skip connections, and a 1×1 convolutional layer is used to adjust the number of channels to achieve dimensionality consistency with the features of the subsequent decoder. The 1×1 convolutional layer after each variable waveform convolutional block (FWC Block) is skip connected to a selective dual attention fusion gate (SDA Gate) of the decoder.

[0110] Specifically, the Variable Waveform Convolutional Block (FWC Block) is defined as follows:

[0111] f′ l =DWConv(AKConv(f l-1 ))+f l-1 ;

[0112] f″ l =σ1(BN{PWConv dim1→4 (LN(f′ l ))});

[0113] f″′ l =σ2(BN{PWConv dim4→1 (f″ l )})+f′ l ;

[0114] f l =HWD(f″′) l );

[0115] Among them, f l This represents the feature map output by the Lth layer variable waveform convolutional block (FWC Block); σ1 represents the SiLU activation function, σ2 represents the GSiLU activation function, BN represents batch normalization, LN represents layer normalization, DWConv represents depthwise convolution, and P... W Conv represents pointwise convolution.

[0116] As an optional embodiment, the decoder corresponds to the encoder in terms of the number of layers. The decoder, from bottom to top, includes five general convolutional blocks, progressively restoring the spatial resolution of the feature maps from bottom to top. A selective dual attention fusion gate (SDA Gate) is embedded before the second to fifth general convolutional blocks. The structure of the selective dual attention fusion gate (SDA Gate) is as follows: Figure 4 As shown;

[0117] The feature map output by the variable waveform convolutional block (FWC Block) of the fifth layer of the encoder is used as the input of the general convolutional block of the first layer of the decoder. Each layer of the decoder obtains a deep feature image through upsampling and inputs it into the selective dual attention fusion gate (SDA Gate) of the next layer of the decoder.

[0118] The selective dual attention fusion gate (SDA Gate) of each layer of the decoder fuses the shallow features of the encoder with the upsampled deep features from the previous layer of the decoder to obtain a fused feature image, which is then passed through the general convolutional block to reduce the number of channels in the fused feature image.

[0119] The fused feature image output by the general convolutional block of the fifth layer of the decoder is the result of the visual disk image segmentation.

[0120] As an optional embodiment, the fused feature map obtained by fusing features from the encoder and decoder through channel attention and spatial attention mechanisms includes:

[0121] For layers 2 through 5 of the decoder, shallow features from the skip connections following the 1×1 convolutional layer and deep features from the upsampled previous layer are input into the selective dual attention fusion gate (SDA Gate).

[0122] The shallow features and the deep features are processed by first performing convolution, batch normalization (BN), and GELU activation functions, respectively.

[0123] Then, the two features are concatenated to obtain a concatenated feature map, and the feature dimensions are adjusted through a reshaping operation;

[0124] The stitched feature map is input into the Selective Channel Attention (SC-A) module, which dynamically allocates weights between shallow and deep features through channel attention. Specifically, max pooling and average pooling are used to extract global information from the stitched feature map, generating global features representing each channel. Then, a Multilayer Perceptron (MLP) is used to calculate the channel weights, and the results are normalized using the Softmax function to obtain the channel attention map. The normalization property of Softmax ensures that the weights between branches sum to 1, allowing for a reasonable allocation of the contribution ratio of each branch, and its competition mechanism can highlight the weights of key features. This design makes the allocation of channel resources more reasonable, making full use of the complementary information of shallow and deep features. Subsequently, the channel attention map is multiplied pointwise with the feature map, and residual connections with deep upsampled features are applied to further emphasize the importance of deep information and enhance the ability to capture complex regions.

[0125] The channel attention map and the spliced ​​feature map are multiplied element-wise, and the result is summed element-wise with the processed deep features to obtain the channel-weighted fused features.

[0126] The channel-weighted fused features are input into the spatial attention (SA) module, which enhances the local representation of the features through a spatial attention mechanism. The spatial attention (SA) module first performs max pooling and average pooling operations on the input feature map to generate two spatial feature maps in the channel dimension. Then, it combines global spatial information through a concatenation operation. The concatenation result is processed by a convolutional layer, BN, and GELU activation function. The generated feature map is normalized using the Sigmoid function to obtain the spatial attention map.

[0127] The spatial attention map and the fused features are multiplied element-wise to obtain the multiplication result;

[0128] The result of the multiplication is then summed element-wise with the fused features to obtain the final fused feature map;

[0129] The fused feature map is reduced in feature dimension by a general convolutional block in the fifth layer of the decoder, and then the output is the visual disc image segmentation result.

[0130] In summary, the selective dual attention fusion gate proposed in this application combines spatial and channel attention mechanisms. Unlike traditional methods, this module effectively fuses features at different levels through channel weighting and optimizes residual connections, significantly enhancing the alignment and fusion of shallow and deep features.

[0131] The Selective Dual Attention Fusion Gate (SDA Gate) is defined as follows:

[0132] f′ s=σ1(BN(Conv(f s )))

[0133] f′ d =f′ s +σ1(BN(Conv(f d )))

[0134] f c&r =Reshape H×W×C×2 (Concat[f′ s ;f′ d ])

[0135] f′ c&r =MLP(MaxPool(Sum(f c&r )))+MLP(AvgPool(Sum(f c&r )))

[0136] SC-AMap = Reshape 1×1×C×2 (σ3(f′ c&r ))

[0137] f sc-a =f c&r ⊙SC-AMap+f′ d

[0138] S-AMap=σ2(σ1(BN(Conv(Concat[MaxPool(f sc-a ); AvgPool(f sc-a )]))))

[0139] f out =f sc-a ⊙S-AMap+f sc-a

[0140] Among them, f d f represents the deep features after upsampling. s shallow features representing skip connections, f outThe SDA gate represents the output features of the selective dual attention fusion gate; σ1, σ2, and σ3 represent the GELU activation function, the Softmax function, and the Sigmoid function, respectively; BN represents batch normalization; MaxPool represents max pooling; AvgPool represents average pooling; MLP represents a multilayer perceptron with one hidden layer, where the feature map size of the hidden layer is C / r×1×1, and r is the scaling ratio; Conv represents convolution; concat represents concatenation; Reshape represents reshaping; and ⊙ represents dot product. It is worth noting that the SDA gate does not change the resolution or number of channels of the feature map during processing. To further adjust the feature dimension, standard convolutional layers are used in subsequent processing to reduce the number of channels to half the original size, thereby reducing the feature dimension.

[0141] As an optional embodiment, this application also includes: a dataset of fundus images for establishing a lightweight image segmentation network FDAU-Net model;

[0142] We used the iChallenge series datasets and the IDRiD dataset;

[0143] The iChallenge-PM dataset focuses on pathological myopia (PM) and contains 400 training images and 400 validation images, including 381 training images and 379 validation images with expert annotations, respectively.

[0144] The iChallenge-AMD dataset is used to study age-related macular degeneration (AMD) and contains 400 training images, of which 270 images have expert annotations.

[0145] The iChallenge-GON dataset is designed for glaucoma diagnosis, providing 400 training images and 400 validation images, along with expert annotations for the optic disc and optic cup regions.

[0146] The IDRiD dataset comprises 81 color fundus images acquired using a Kowa VX-10alpha digital fundus camera with a 50-degree field of view. These images show signs of diabetic retinopathy and include masked annotations of the optic disc region.

[0147] To ensure data processing consistency, this application filtered and integrated the iChallenge dataset, ultimately yielding 1,830 images. These images were divided into training-validation and test sets in an 8:2 ratio. The training-validation set was further divided into training and validation sets in a 9:1 ratio, resulting in 1,318 training images, 146 validation images, and 366 test images. For the IDRiD dataset, the standard segmentation method from the Diabetic Retinopathy: Segmentation and Grading Challenge was used, generating 54 training images and 27 test images. Figure 5 A dataset example is shown.

[0148] Figure 5 In the image, (a) represents an example of the iChallenge-PM dataset; (b) represents an example of the iChallenge-AMD dataset; (c) represents an example of the iChallenge-GON dataset; and (d) represents an example of the IDRiD dataset.

[0149] For clarity, all datasets maintain a 1:1 aspect ratio. Key statistics are summarized in Table 1 below.

[0150] Table 1. Distribution and characteristics of the four datasets

[0151]

[0152] As an optional embodiment, this application further includes: data preprocessing and enhancement steps;

[0153] In the iChallenge-GON dataset, experts independently labeled the optic disc and optic cup, such as Figure 6 As shown in section (c). Among them, Figure 6 (a) is the original iChallenge-GON image data; (b) is the label obtained after threshold segmentation; (c) is the real label of the original iChallenge-GON image.

[0154] Since the visual cup is located in the central region of the visual disk and is part of it, it is necessary to merge the labeled regions before model training. We adopted a simple pixel value processing method to achieve this, setting all pixel values ​​in the labels that are not 255 to 0. The merged labels are as follows: Figure 6 As shown in section (b).

[0155] Due to the unique characteristics of medical images, the accuracy of color and structure is crucial for diagnosis; any color distortion or structural changes can severely impact their diagnostic value. However, the scarcity of fundus image data necessitates data augmentation strategies. We employed an efficient data augmentation strategy. By performing operations in both pixel-level and spatial-level augmentation spaces, we expanded the diversity of the data while avoiding significant color distortion and structural changes, thus preserving the diagnostic value of the images.

[0156] Enhanced sample images, such as Figure 7 As shown in the figure, (a) is the original training image. Based on this, we generated five augmented images from (b) to (f) using the MedAugment strategy. These images were randomly augmented with 0-1 pixel enhancements (brightness adjustment, sharpening, noise addition, etc.) and 0-3 spatial enhancements (rotation, scaling, translation, etc.). Furthermore, we further employed random cropping and mesh masking in (g) and (h) to enrich the diversity of the dataset.

[0157] in, Figure 7 In the data augmentation results examples, (a) is the original image, (b) to (f) are MedAugment data augmentation results, (g) is the random cropping result, and (h) is the grid masking result.

[0158] This application also compares the technical performance: In the quantitative evaluation phase, we compare the proposed method with various classic U-shaped networks, specifically U-Net, U-Net++, U-Net3+, Attention U-Net, and TransUNet. Furthermore, we use this dataset to compare with other state-of-the-art (SOTA) methods, such as Phase Fusion PSPNet (PFPSPNet) (Rauf et al., 2021), 2C128N0D (Fang et al., 2021), Probabilistic Bubble (U-Net PB) (Fu et al., 2021), SimpleCNN-UNet (Xiao et al., 2024b), and MIU-Net (Xiao et al., 2024a). Specific results are shown in Tables 2 and 3, where Table 2 presents the results across the iChallenge dataset, and Table 3 details the performance of each iChallenge dataset. These two tables show that our proposed FADU-Net achieves leading results on the vast majority of performance metrics. Compared to other methods, FADU-Net shows improvements in key metrics such as F1, IoU, and Acc. The results in these two tables demonstrate that our proposed FADU-Net achieves leading performance on most metrics, with significant improvements in F1, IoU, and Acc compared to other methods. However, the improvement on the iChallenge-GON dataset is relatively limited, possibly because the optic discs of glaucoma patients are more regularly shaped and easier to segment, which may not fully showcase the advantages of our method. In contrast, on the iChallenge-AMD dataset, other methods generally perform only moderately, making the advantages of our method more pronounced.

[0159] Furthermore, we compared FADU-Net with ten other state-of-the-art methods on the IDRID dataset, including the method proposed by the team that performed best in the spectral disc segmentation competition, as well as DRNet (Hasan et al., 2021), OD MaskR-CNN (Maysanjaya et al., 2022), W-Net (Tang et al., 2024), U-Net PB, and SimpleCNN-UNet. Since many of these methods did not release their code, the experimental data were collected from research papers and official competition results. Table 4 shows that FADU-Net, pre-trained on the iChallenge dataset, achieved the best performance. Notably, the second-ranked method, Mask R-CNN from the ZJU BH-SGEX team, used a 1024×1024 input image, while FADU-Net only required a 512×512 image. This demonstrates that FADU-Net achieves competitive or better performance at lower resolutions, highlighting its efficiency and practicality in scenarios with limited computational resources.

[0160] Table 2 shows the experimental results of the iChallenge dataset.

[0161]

[0162] Table 3 Experimental results for the iChallenge-PM, iChallenge-GON, and iChallenge-AMD datasets

[0163]

[0164] Table 4 Experimental results of the IDRiD dataset

[0165]

[0166] Qualitative results analysis: such as Figure 8 As shown, in the qualitative evaluation section, the proposed method is compared with several U-shaped networks mentioned in the quantitative evaluation. Figure 8 As shown, the model's performance corresponds to the quantitative experimental results in Table 2. It can be seen that most methods perform well for relatively smooth and regular-shaped eyelid regions. However, our method demonstrates higher accuracy in preserving edge details, primarily due to the advantages of wavelet downsampling. For irregularly shaped diseased eyelid regions, the differences in segmentation performance between various methods become increasingly apparent. The proposed method performs better in segmenting irregular regions and fine details, consistently approaching the real-world situation more closely.

[0167] Figure 8Qualitative analysis: (a) original image, (b) image label, (c) U-Net, (d) U-Net++, (e) U-Net3+, (f) attention U-Net, (g) simple CNN-UNet, (h) the method of this application.

[0168] According to another aspect of this application, this application also provides a video disc splitting system, comprising:

[0169] The acquisition module acquires the video disc image;

[0170] The encoder module is used to extract multi-scale features of the visual disc image. The encoder contains multiple variable waveform convolutional blocks (FWC Blocks). Each FWC Block dynamically adjusts its receptive field through a dynamic convolutional layer (AKConv) and combines it with a wavelet downsampling layer (HWD) for feature compression.

[0171] The decoder module uses a decoder to gradually restore the feature map resolution. The decoder contains multiple selective dual attention fusion gates (SDA Gates) and obtains a fused feature map by fusing the features of the encoder and decoder through channel attention and spatial attention mechanisms.

[0172] The output module obtains the visual disk image segmentation result based on the fused feature map.

[0173] It should be understood that specific limitations regarding the video disc segmentation system can be found in the limitations of the video disc image segmentation method described above, and will not be repeated here. Each module / unit in the aforementioned video disc image segmentation system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules / units can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the operations corresponding to each module.

[0174] This application also provides a computer device, including: a processor coupled to a memory, the memory storing computer execution instructions, and the processor executing the computer execution instructions stored in the memory, causing the processor to perform a video disc image segmentation method.

[0175] This application also provides a computer-readable storage medium for storing a computer program, the computer program including instructions for implementing a video disc image segmentation method.

[0176] The memory stores computer program instructions that, when executed, cause the processor to perform the steps of the video disc image segmentation method.

[0177] It should be understood that the processor provided in the above embodiments of this application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0178] It should also be understood that the number of processors in the computer device in the above embodiments of this application can be one or more, and can be adjusted according to the actual application scenario. This is merely an illustrative example and is not intended to limit the scope. The number of memory in the embodiments of this application can be one or more, and can be adjusted according to the actual application scenario. This is merely an illustrative example and is not intended to limit the scope.

[0179] It should also be noted that when a computer device includes a processor (or processing unit) and a memory, the processor in this application may be integrated with the memory, or the processor and the memory may be connected through an interface. This can be adjusted according to the actual application scenario and is not limited.

[0180] This application provides a chip system including a processor for supporting a computer device (client or server) in implementing the functions of the controller involved in the above-described methods, such as processing data and / or information involved in the above-described methods. In one possible design, the chip system also includes a memory for storing necessary program instructions and data. This chip system can be composed of chips or may include chips and other discrete devices.

[0181] In another possible design, when the chip system is a chip within user equipment or an access network, the chip includes a processing unit and a communication unit. The processing unit may be, for example, a processor, and the communication unit may be, for example, an input / output interface, pins, or circuits. The processing unit can execute computer-executable instructions stored in the storage unit to cause the chip within the client or management server to perform the steps of a common-sense question-and-answer method. Optionally, the storage unit can be an on-chip storage unit, such as a register or cache, or it can be an external storage unit within the client or management server, such as read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, such as random access memory (RAM).

[0182] It should be understood that the methods and / or embodiments in this application can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processing unit, the functions defined in the methods of this application are performed.

[0183] It should be understood that the controller or processor mentioned in the above embodiments of this application may be a central processing unit (CPU), or one or more combinations of other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0184] It should also be understood that the number of processors or controllers in the computer devices or chip systems in the above embodiments of this application may be one or more, and can be adjusted according to the actual application scenario. This is merely an illustrative example and is not intended to limit the scope. Similarly, the number of memories in the embodiments of this application may be one or more, and can be adjusted according to the actual application scenario. This is merely an illustrative example and is not intended to limit the scope.

[0185] It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0186] In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, capable of transmitting, propagating, or transmitting a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0187] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0188] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0189] In another aspect, embodiments of this application also provide a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more computer-readable instructions, which may be executed by a processor to implement the steps of the methods and / or technical solutions of the various embodiments of this application. The computer may be the aforementioned computer device (client or server or other computer network communication device).

[0190] In a typical configuration of this application, the terminal and the service network devices each include one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0191] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0192] Computer-readable media include both permanent and non-permanent, removable and non-removable media, which can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, read-only optical disc (CD-ROM), digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0193] Furthermore, this application also provides a computer program stored in a computer device, which causes the computer device to execute the method for executing the control code.

[0194] It should be noted that this application can be implemented in software and / or a combination of software and hardware, for example, using an application-specific integrated circuit (ASIC), a general-purpose computer, or any other similar hardware device. In some embodiments, the software program of this application can be executed by a processor to implement the steps or functions described above. Similarly, the software program of this application (including related data structures) can be stored in a computer-readable recording medium, such as RAM memory, magnetic or optical drives, floppy disks, and similar devices. Furthermore, some steps or functions of this application can be implemented in hardware, for example, as circuitry that cooperates with a processor to perform the various steps or functions.

[0195] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within this application. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, the terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0196] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The singular forms "a," "the," and "the" used in the embodiments of this application are also intended to include the plural forms unless the context clearly indicates otherwise.

[0197] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0198] In the description of this application, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship. For example, A / B can mean A or B. "And / or" in this application is merely a description of the relationship between the related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Depending on the context, the words "if" or "if" as used herein can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0199] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method of optic disc segmentation, characterized by, The method comprises the following steps: obtaining an optic disc image; extracting multi-scale features of the optic disc image using an encoder, the encoder comprising a plurality of variable waveform convolution blocks (FWC Block), each variable waveform convolution block dynamically adjusting a receptive field through a dynamic convolution layer (AKConv) and combining a wavelet down-sampling layer (HWD) for feature compression; gradually restoring feature map resolution using a decoder, the decoder comprising a plurality of selective dual attention fusion gates (SDAGate), and fusing features of the encoder and the decoder through channel attention and spatial attention mechanisms to obtain a fused feature map; obtaining an optic disc image segmentation result according to the fused feature map.

2. The method of claim 1, wherein, The encoder comprises: a feature extraction block (Stem) for performing preliminary processing on an input image and extracting low-level features; five variable waveform convolution blocks (FWC Block) sequentially connected from top to bottom after the feature extraction block (Stem), each layer outputting a feature channel number of C1-C5: 16, 32, 128, 160, and 256.

3. The method of claim 2, wherein, The variable waveform convolution block (FWC Block) comprises a dynamic convolution layer (AKConv), a depth convolution layer, an LN layer, a point-wise convolution layer, a BN layer, an siLU layer, a point-wise convolution layer, a BN layer, and a GSiLU layer connected in sequence; wherein a reverse bottle neck structure is adopted between the two groups of point-wise convolution layers, the channel number is first expanded four times, and then compressed and restored; wherein the first to fourth variable waveform convolution blocks (FWC Block) further comprise a wavelet down-sampling layer (HWD) connected after the GSiLU layer; The wavelet down-sampling layer (HWD) simultaneously performs feature down-sampling and channel depth enhancement, outputs a group of features before the wavelet down-sampling layer (HWD) for skip connection, and uses a 1x1 convolution layer to adjust the channel number to achieve dimensional consistency with the subsequent decoder features; wherein the 1x1 convolution layer after each variable waveform convolution block (FWC Block) is respectively skip connected with one selective dual attention fusion gate (SDAGate) of the decoder.

4. The method of claim 3, wherein, The variable waveform convolution block (FWC Block) is defined as follows: f l ' = DWConv(AKConv(f l-1 ))+f l-1 ; f l "= σ1(BN{PWConv dim1→4 (LN(f l '))}); f l " = σ2(BN{PWConv dim4→1 (f l ")}) + f l '; f l = HWD(f l ") ; wherein f l represents the feature map output by the L-th layer variable waveform convolution block FWC Block; σ1 represents a SiLU activation function, σ2 represents a GSiLU activation function, BN represents batch normalization, LN represents layer normalization, DWConv represents a depth convolution, and PWConv represents a point-wise convolution.

5. The method of claim 3, wherein, The decoder corresponds to the encoder in terms of the number of layers, and the decoder comprises five general convolution blocks from bottom to top; a selective dual attention fusion gate (SDAGate) is embedded before the second to fifth general convolution blocks; the feature map output by the fifth variable waveform convolution block (FWC Block) of the encoder is input as the first general convolution block of the decoder, each layer of the decoder obtains a deep feature image through up-sampling and inputs the selective dual attention fusion gate (SDAGate) of the next layer of the decoder; the selective dual attention fusion gate (SDAGate) of each layer of the decoder fuses the shallow features of the encoder through skip connection and the deep features after up-sampling from the previous layer of the decoder to obtain a fused feature image, and reduces the channel number of the fused feature image through the general convolution block; wherein the fused feature image output by the fifth general convolution block of the decoder is the optic disc image segmentation result.

6. The method of claim 5, wherein, The fused feature map obtained by fusing the features of the encoder and the decoder through the channel attention and the spatial attention mechanism comprises: For the second layer to the fifth layer of the decoder, the shallow layer features from the 1x1 convolution layer after the jump connection and the deep layer features after the upsampling of the previous layer are input into the selective double attention fusion gate SDA Gate, and the shallow layer features and the deep layer features are processed by convolution, batch normalization BN and GELU activation function respectively; Then, the two features are spliced to obtain a spliced feature map, and the feature dimension is adjusted through a reshaping operation; The spliced feature map is input into the selective channel attention SC-A module, the selective channel attention SC-A module uses global information extraction on the spliced feature map through maximum pooling and average pooling to generate global features representing each channel; then, the channel weight is calculated using a multilayer perceptron MLP, and the channel attention map is obtained by normalizing the channel weight through a Softmax function; The channel attention map and the spliced feature map are multiplied element by element, and the result is summed element by element with the processed deep layer features to obtain a channel-weighted fused feature; The channel-weighted fused feature is input into the spatial attention S-A module; the spatial attention S-A module first performs maximum pooling and average pooling operations on the input feature map to generate two spatial feature maps in the channel dimension; then, the global spatial information is combined through a splicing operation, and the splicing result is processed through a convolution layer, BN and GELU activation function to generate a feature map. The spatial attention map is obtained by normalizing the feature map using the Sigmoid function; The spatial attention map and the fused feature are multiplied element by element to obtain a multiplication result; The multiplication result and the fused feature are summed element by element to obtain the final fused feature map; The fused feature map is output after reducing the feature dimension through the general convolution block of the fifth layer of the decoder to obtain the optic disc image segmentation result.

7. The method of claim 6, wherein, The selective double attention fusion gate SDA Gate is defined as follows: f′ s = σ1(BN(Conv(f s ))) f′ d = f′ s + σ1(BN(Conv(f d ))) f c&r = Reshape H×W×C×2 (Concat[f′ s ; f′ d ]) f′ c&r = MLP(MaxPool(Sum(f c&r )) ) + MLP(AvgPool(Sum(f c&r ))) SC-AMap = Reshape 1×1×C×2 (σ3(f′ c&r )) f sc-a = f s&r ⊙SC-AMap + f d S-AMap = σ2(σ1(BN(Conv(Concat[MaxPool(f sc-a ); AvgPool(f sc-a )])))) f out = f sc-a ⊙S-AMap + f sc-a wherein f d denotes the up-sampled deep features, f s denotes the shallow features of the skip connection, f out denotes the output features of the selective dual attention fusion gate SDA Gate; σ1, σ2and σ3denote GELU activation function, Softmax function and Sigmoid function, respectively; BN denotes batch normalization; MaxPool denotes max-pooling; AvgPool denotes average-pooling; MLP denotes a multi-layer perceptron with one hidden layer, wherein the feature map size of the hidden layer is C / r x 1 x 1, r being a scaling ratio; Conv denotes convolution; concat denotes concatenation operation; Reshape denotes reshaping operation; and denotes dot product operation.

8. An optic disc segmentation system characterized by, It comprises: An acquisition module acquires an optic disc image; An encoder module is configured to extract multi-scale features of the optic disc image, and the encoder comprises a plurality of variable wave convolution blocks FWC Block, each variable wave convolution block FWC Block dynamically adjusts the receptive field through a dynamic convolution layer AKConv and combines a wavelet down-sampling layer HWD to compress the features; A decoder module gradually recovers the resolution of the feature map using a decoder, and the decoder comprises a plurality of selective double attention fusion gates SDA Gate, which fuse the features of the encoder and the decoder through the channel attention and the spatial attention mechanism to obtain a fused feature map; An output module obtains an optic disc image segmentation result according to the fused feature map.

9. A computer device, comprising: It comprises: A processor coupled to a memory, the memory storing computer-executable instructions, and the processor executing the computer-executable instructions stored in the memory to cause the processor to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized by, A computer program product for storing a computer program comprising instructions for implementing the method according to any one of claims 1 to 7. A computer program product for storing a computer program comprising instructions for implementing the method according to any one of claims 1 to 7.

Images