Glycoreticular lesion segmentation method based on wavelet frequency decomposition and enhancement network

Abstract

The invention discloses a glycoreticular lesion segmentation method based on wavelet frequency decomposition and an enhanced network, and belongs to the technical field of medical image processing. The method comprises the following steps of: performing high and low frequency decomposition on multi-level encoder features by using discrete wavelet transform; enhancing the low-frequency features by using a low-frequency enhancer; enhancing high-frequency features by using a high-frequency intensifier; generating a glycoreticular lesion segmentation map by using a segmentation decoder; and training the model by using the loss function. According to the method, high and low frequency decomposition is performed on the multi-level encoder features through discrete wavelet transform, and the low-frequency enhancer and the high-frequency enhancer are respectively adopted to perform explicit enhancement on the low-frequency features (semantic information) and the high-frequency features (detailed information), so that the problem of gain imbalance of the semantic information and the detailed information in the prior art is solved; according to the method, the internal semantic characterization and detail characterization of the features are refined at the same time, and the accuracy of diabetic retinopathy segmentation is effectively improved.

Description

Technical Field

[0001] This invention relates to a method for segmenting diabetic retinopathy lesions based on wavelet frequency decomposition and enhancement networks, belonging to the field of medical image processing technology. Background Technology

[0002] Diabetic retinopathy (DR) is the most common eye complication of diabetes and a leading cause of vision impairment and blindness in the working-age adult population. Regular DR screening is widely recognized as a crucial component of primary diabetes care. During DR screening, ophthalmologists diagnose DR lesions by identifying areas of DR in color fundus images. However, in areas with limited medical resources, each ophthalmologist must screen a large number of fundus images, resulting in a heavy workload. Therefore, DRLS (Automatic Diabetic Retinopathy Segmentation) technology has crucial application value in assisting diagnosis and reducing the workload of ophthalmologists.

[0003] To achieve high-precision segmentation of diabetic retinopathy, the model needs to capture both accurate semantic and detailed information. This necessity stems from the high similarity between lesions (e.g., hemorrhages and microaneurysms both belong to the red lesion category) and the high similarity between lesions and the background (e.g., microaneurysms and blood vessels both belong to the red object category). Figure 1 As shown in the first row), this type of characteristic requires precise semantic information to distinguish different lesion types; at the same time, accurate detailed information is crucial for the segmentation and boundary delineation of minute lesions (such as...). Figure 1 As shown in the second row).

[0004] To obtain enhanced semantic information and detailed information, a popular approach is to design complex multi-level feature fusion schemes (such as...). Figure 2 (a) shows that this scheme is based on the characteristics of high-level encoder features being rich in semantic information and low-level encoder features being rich in detail information, and thus fuses them. Furthermore, some researchers have noted that the quality of multi-level features also affects the final semantic and detail representation, so they propose using task-specific encoders (such as M2MRF, HACDR-Net, etc.) Figure 2 (b) shows that multi-level features suitable for DRLS tasks are directly obtained, and feedback mechanisms (such as WSRFNet, etc.) are used. Figure 2 (c) shows the refinement of multi-level features extracted by a simple encoder in the spatial domain.

[0005] While the above methods enhance feature representations overall, their inherent semantic information (represented by low-frequency components) and detailed information (represented by high-frequency components) are not simultaneously enhanced in a definite manner. This feature enhancement approach may lead to an imbalance between semantic and detailed information gains—that is, one type of information is enhanced while the other is weakened, thus reducing the overall enhancement effect (e.g., Figure 1 (d) is shown. Summary of the Invention

[0006] To address the problems existing in the background technology, the present invention provides a method for segmenting diabetic retinopathy lesions based on wavelet frequency decomposition and enhancement network.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for segmenting diabetic retinopathy lesions based on wavelet frequency decomposition and enhancement networks, the method comprising the following steps:

[0008] S1: Use discrete wavelet transform to decompose the features of the multi-level encoder into high and low frequencies;

[0009] S101: Utilizing the backbone network through a CNN-based encoder After adjusting the number of channels through two convolutional layers, multi-level encoder features are extracted from the fundus image. Where: H is the first Hierarchical 2D encoder feature map The height, W is the first Hierarchical 2D encoder feature map width, Indicates the first The hierarchy is from bottom to top;

[0010] S102: Using discrete wavelet transform to transform the features of the multi-level encoder... The feature map of the hierarchical two-dimensional encoder is decomposed to obtain four wavelet subbands and four filters:

[0011] (1)

[0012] In formula (1):

[0013] , , , Indicates four wavelet sub-bands;

[0014] , , , This represents four filters;

[0015] It is a convolution operation with a stride of 2;

[0016] S103: Due to wavelet subband It is the first Hierarchical 2D encoder feature map The low-frequency components, therefore, the first Hierarchical low-frequency features Therefore, we can obtain the following: the low-frequency features obtained after decomposing the features of the multi-level encoder using discrete wavelet transform are: ;

[0017] S104: Considering wavelet subbands , as well as Includes the first Hierarchical 2D encoder feature map Therefore, a cross-channel splicing operation is used to merge all high-frequency components to obtain the first high-frequency information. Hierarchical high-frequency features Therefore, we can obtain the high-frequency features obtained by using discrete wavelet transform to decompose the features of the multi-level encoder as follows: .

[0018] S2: Use a low-frequency enhancer to enhance low-frequency characteristics;

[0019] S201: Pass Convolutional layers for the highest level of low-frequency features Enhancement is performed to obtain the highest level of enhanced low-frequency features. ;

[0020] S202: For the first Enhanced low-frequency features of the layer Upsampled features are obtained by upsampling by a factor of 2 using bilinear interpolation. ;

[0021] S203: Upsample the features With the Low-frequency characteristics of the hierarchy Merging is performed through element-wise addition.

[0022] S204: Application The convolutional layer enhances the fused features, resulting in the first... Low-frequency features after layer enhancement :

[0023] (2)

[0024] In formula (2):

[0025] express Convolution operation.

[0026] S3: Enhance high-frequency characteristics using a high-frequency enhancer;

[0027] S301: Apply FFT to obtain the... High-frequency characteristics Fourier spectrum :

[0028] (3)

[0029] In formula (3):

[0030] Indicates the first High-frequency characteristics Fourier spectrum The real part;

[0031] Indicates the first High-frequency characteristics Fourier spectrum The imaginary part;

[0032] Represents the imaginary unit;

[0033] S302: Enriching the first layer with complex convolutional layers High-frequency characteristics Fourier spectrum Post-output complex convolution features ;

[0034] The operation of the complex convolutional layer described in S302 includes the following steps:

[0035] S30201: Perform two-dimensional complex convolution operation:

[0036] (4)

[0037] In equation (4):

[0038] Denotes the complex convolution kernel, where: Denotes the real part of the complex convolution kernel. Represents the imaginary part of the complex convolution kernel;

[0039] This indicates the input of a complex variable, where: This represents the real part of the input complex number variable. Represents the imaginary part of the input complex number variable;

[0040] This indicates complex value output, where: This represents the real part of the complex value output. Represents the imaginary part of the complex value output;

[0041] S30202: Real part of the complex value output and the imaginary part of complex value output All are standardized in multiple batches;

[0042] S30203: The real part of the complex value output obtained from S30202. and the imaginary part of complex value output All activation operations are performed using the complex ReLU function.

[0043] S303: Convolution features of complex numbers Input into a complex channel attention block; output complex channel attention block features. ;

[0044] The operation of the complex channel attention block described in S303 includes the following steps:

[0045] S30301: Spatial information of complex convolution features is aggregated using channel-wide average pooling to obtain pooling features. Spatial information from complex convolution features is aggregated using global max pooling to obtain pooled features. ;

[0046] S30302: After merging the two pooling features mentioned above through addition, complex channel attention is obtained through one-dimensional complex convolution and complex sigmoid function. ,Right now:

[0047] (5)

[0048] In formula (5):

[0049] This indicates that the complex sigmoid function is applied to the real and imaginary parts of the features from the complex convolution, respectively.

[0050] Represents one-dimensional complex convolution;

[0051] S30303: The final output complex channel attention block features are obtained by weighting the input complex convolutional features through complex channel attention.

[0052] (6)

[0053] In formula (6):

[0054] ⊙ represents the Hadamard product operation.

[0055] S304: Attention block features for complex channels Input into a complex spatial attention block; output complex spatial attention block features. ;

[0056] The operation of the complex spatial attention block described in S304 includes the following steps:

[0057] S30401: Employs two activation functions: complex ReLU and complex sigmoid. Two-dimensional complex convolution aggregates complex channel attention block features to generate complex spatial attention. ,Right now:

[0058] (7)

[0059] In equation (7):

[0060] This indicates that the real and imaginary parts of the complex channel attention block features are calculated by applying the complex ReLU and complex sigmoid activation functions, respectively.

[0061] S30402: The final output complex spatial attention block features are obtained by weighting the input complex channel attention block features using complex spatial attention blocks.

[0062] (8)

[0063] S305: Attention Block Features in Complex Space Applying IFFT, we obtain the first... Enhanced high-frequency characteristics of the level ;

[0064] S306: Utilizes the same structure as the low-frequency enhancer to achieve the output of the high-frequency enhancer. .

[0065] S4: Use a segmentation decoder to generate a segmentation map of diabetic retinopathy lesions;

[0066] S401: Enhanced low-frequency features are obtained through inverse discrete wavelet transform. and enhanced high-frequency features Reconstructed as wavelet-enhanced features ;

[0067] S402: Aggregate adjacent encoder layer information using residual methods to obtain aggregated adjacent encoder features:

[0068] (9)

[0069] In equation (9):

[0070] express Convolutional layer;

[0071] S403: Enhance wavelet features and aggregated adjacent encoder features The final enhanced feature is obtained by adding them together. ;

[0072] S404: Decoder, which will ultimately enhance the features. With upsampling features Integration and Convolutional layer processing yields the first Each decoder block output :

[0073] (10)

[0074] S5: Train the model using the loss function.

[0075] The loss function described in S5 is as follows:

[0076] (11)

[0077] In equation (11):

[0078] This represents the segmentation map output by the segmentation decoder;

[0079] This represents the output split diagram of the low-frequency booster.

[0080] This represents the output segmentation diagram of the high-frequency booster.

[0081] Indicates the true label;

[0082] These are the weights of the auxiliary loss function;

[0083] Let represent the dice loss function, where:

[0084] K is the total number of lesion categories;

[0085] This represents the prediction result for the k-th type of lesion;

[0086] This represents the true label of the k-th type of lesion;

[0087] This represents the l1 norm.

[0088] Compared with the prior art, the beneficial effects of the present invention are:

[0089] This invention decomposes multi-level encoder features into high and low frequencies using discrete wavelet transform, and explicitly enhances low-frequency features (semantic information) and high-frequency features (detail information) using low-frequency enhancers and high-frequency enhancers respectively. This solves the problem of imbalance between semantic and detail information gain in existing technologies, and achieves simultaneous refinement of the intrinsic semantic and detail representations of features. It overcomes the segmentation challenges caused by high similarity between lesions and high similarity between lesions and background, effectively improving the accuracy of diabetic retinopathy segmentation, and providing key technical support for auxiliary diagnosis and reducing the workload of ophthalmologists. Attached Figure Description

[0090] Figure 1 This is a comparison of the segmentation results of this invention and WSRFNet for diabetic retinopathy, where: (a) is a fundus image; (b) is the ground truth label of the comparison region; (c) is the segmentation result obtained using the method of this invention; and (d) is the segmentation result obtained using the WSRFNet method. The ground truth label of the comparison region (b) is the area marked by the green box in the fundus image (a). The areas marked or filled with red, green, and blue edges are hard exudates, hemorrhages, and microaneurysms, respectively, and the white solid box marks challenging lesions. In the segmentation results (c) obtained using the method of this invention and (d) obtained using the WSRFNet method, the white dashed boxes mark the missed and false detection areas, and the yellow dashed boxes mark the areas with low segmentation boundary accuracy.

[0091] Figure 2 Figure 1 shows a comparison of different feature processing schemes for diabetic retinopathy segmentation. (a) is a multi-level feature fusion scheme based on features rich in semantic information from high-level encoders and features rich in detail information from low-level encoders. (b) is a scheme that directly obtains multi-level features suitable for DRLS tasks through task-specific encoders (such as M2MRF and HACDR-Net). (c) is a scheme that refines multi-level features extracted by simple encoders in the spatial domain using feedback mechanisms (such as WSRFNet). Figure 2 (d) is a scheme that combines discrete wavelet transform (DWT), low-frequency enhancer (LFB), high-frequency enhancer (HFB), and inverse discrete wavelet transform (IDWT) for feature processing.

[0092] Figure 3 This is a flowchart of the present invention;

[0093] Figure 4 This is the first of the present invention Architecture diagram of a CCFAM block;

[0094] Figure 5 It is the segment decoder. The structure diagram of each decoding block shows that the information flow is color-coded, with blue arrows indicating only when... Activated when >1, orange arrow only when <5 indicates activation, and a black arrow indicates general activation. Detailed Implementation

[0095] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0096] A method for segmenting diabetic retinopathy lesions based on wavelet frequency decomposition and enhancement networks, the method comprising the following steps:

[0097] S1: Use discrete wavelet transform to decompose the features of the multi-level encoder into high and low frequencies;

[0098] Typically, semantic information and detail information reside in the low-frequency and high-frequency components of multi-level encoder features, respectively. Therefore, inspired by the effectiveness of discrete wavelet transform in frequency decomposition, this invention uses discrete wavelet transform to obtain the inherent semantic and detail information of each level of encoder features, such as... Figure 3 As shown in (b).

[0099] S101: Utilizing the backbone network through a CNN-based encoder After adjusting the number of channels to C=64 through two convolutional layers, multi-level encoder features are extracted from the fundus image. Where: H is the first Hierarchical 2D encoder feature map The height, W is the first Hierarchical 2D encoder feature map width, Indicates the first The hierarchy is from bottom to top;

[0100] S102: Using discrete wavelet transform to transform the features of the multi-level encoder... The feature map of the hierarchical two-dimensional encoder is decomposed to obtain four wavelet subbands and four filters:

[0101] (1)

[0102] In formula (1):

[0103] , , , Indicates four wavelet sub-bands;

[0104] , , , This represents four filters, whose values ​​are set according to the wavelet basis of the discrete wavelet transform.

[0105] It is a convolution operation with a stride of 2;

[0106] S103: Due to wavelet subband It is the first Hierarchical 2D encoder feature map The low-frequency components, therefore, the first Hierarchical low-frequency features Therefore, we can obtain the following: the low-frequency features obtained after decomposing the features of the multi-level encoder using discrete wavelet transform are: ;

[0107] S104: Considering wavelet subbands , as well as Includes the first Hierarchical 2D encoder feature map Therefore, a cross-channel splicing operation is used to merge all high-frequency components to obtain the first high-frequency information. Hierarchical high-frequency features Therefore, we can obtain the high-frequency features obtained by using discrete wavelet transform to decompose the features of the multi-level encoder as follows: .

[0108] S2: Use a low-frequency enhancer to enhance low-frequency characteristics;

[0109] The low-frequency enhancement module aims to enhance the low-frequency components of the encoder features at each level by utilizing multi-scale information. To this end, considering that features at different levels have different scales, this invention employs a structure similar to a feature pyramid network to fuse multi-level low-frequency features, such as... Figure 3 As shown in (c).

[0110] S201: Pass Convolutional layers for the highest level of low-frequency features Enhancement is performed to obtain the highest level of enhanced low-frequency features. ;

[0111] S202: For the first Enhanced low-frequency features of the layer Upsampled features are obtained by upsampling by a factor of 2 using bilinear interpolation. ;

[0112] S203: Upsample the features With the Low-frequency characteristics of the hierarchy Merging is performed through element-wise addition.

[0113] S204: Application The convolutional layer enhances the fused features, resulting in the first... Low-frequency features after layer enhancement :

[0114] (2)

[0115] In formula (2):

[0116] express Convolution operation.

[0117] S3: Enhance high-frequency characteristics using a high-frequency enhancer;

[0118] like Figure 3 As shown in (d), the high-frequency enhancer aims to enhance the high-frequency components of the encoder features at each level through two steps. The first step enhances key features and reduces noise in Fourier space. To achieve the goal of the first step, previous Fourier-based feature enhancement methods filter Fourier space features by learning a parameterized attention map. However, these methods ignore the fact that diverse inputs contain various noises and key features, which may make it difficult for their static filters to meet the diverse enhancement requirements of different inputs. To alleviate this problem, this invention proposes a CCFAM that utilizes dynamic channels and spatial attention generated from the input spectrum to refine the high-frequency components. Considering that the features in Fourier space are complex values, the attention map is obtained through complex convolution.

[0119] S301: Applying FFT (Fast Fourier Transform) to obtain the i-th level high-frequency features Fourier spectrum :

[0120] (3)

[0121] In formula (3):

[0122] Indicates the first High-frequency characteristics Fourier spectrum The real part;

[0123] Indicates the first High-frequency characteristics Fourier spectrum The imaginary part;

[0124] Represents the imaginary unit;

[0125] S302: Enriching the first layer with complex convolutional layers High-frequency characteristics Fourier spectrum Post-output complex convolution features ;

[0126] Convolution is widely used in speech signal processing to enhance the Fourier spectrum of signals. Inspired by its success, this invention applies complex convolution to process the complex-valued spectrum of image features. The operation of the complex convolution layer described in S302 includes the following steps:

[0127] S30201: Perform two-dimensional complex convolution operation:

[0128] (4)

[0129] In equation (4):

[0130] Denotes the complex convolution kernel, where: Denotes the real part of the complex convolution kernel. Represents the imaginary part of the complex convolution kernel;

[0131] This indicates the input of a complex variable, where: This represents the real part of the input complex number variable. Represents the imaginary part of the input complex number variable;

[0132] This represents a complex-valued output that captures information about the amplitude and phase from the input, where: This represents the real part of the complex value output. Represents the imaginary part of the complex value output;

[0133] It can be seen that complex convolution can be implemented by two real-valued convolutions, with the kernel weights being the real parts of the complex convolution kernel. The imaginary part of the complex convolution kernel ,like Figure 4 As shown in (a).

[0134] S30202: Real part of the complex value output and the imaginary part of complex value output All are standardized in multiple batches;

[0135] S30203: The real part of the complex value output obtained from S30202. and the imaginary part of complex value output All activation operations are performed using the complex ReLU function.

[0136] S303: Convolution features of complex numbers Input into a complex channel attention block; output complex channel attention block features. Assign higher weights to channels containing key information;

[0137] The architecture of complex channel attention blocks is as follows Figure 4 As shown in (b). The operation of the complex channel attention block described in S303 includes the following steps:

[0138] S30301: Spatial information of complex convolution features is aggregated using channel-wide average pooling to obtain pooling features. Spatial information from complex convolution features is aggregated using global max pooling to obtain pooled features. ;

[0139] S30302: The two pooling features mentioned above are combined by addition for further aggregation. Complex channel attention is obtained through one-dimensional complex convolution with a kernel size of 3 and a complex sigmoid function. ,Right now:

[0140] (5)

[0141] In formula (5):

[0142] This indicates that the complex sigmoid function is applied to the real and imaginary parts of the features from the complex convolution, respectively.

[0143] Represents one-dimensional complex convolution;

[0144] Instead of using fully connected operations to capture dependencies between all channels, this invention learns channel attention of complex values ​​in a local manner, inspired by the low computational cost of local attention.

[0145] S30303: The final output complex channel attention block features are obtained by weighting the input complex convolutional features through complex channel attention.

[0146] (6)

[0147] In formula (6):

[0148] This represents the Hadamard product operation.

[0149] S304: Attention block features for complex channels Input into a complex spatial attention block; output complex spatial attention block features. This is done to highlight key spectrum components and suppress noise-related spectral components.

[0150] The structure of complex spatial attention blocks is as follows Figure 4 As shown in (c). The operation of the complex spatial attention block described in S304 includes the following steps:

[0151] S30401: Employs two activation functions: complex ReLU and complex sigmoid. Two-dimensional complex convolution aggregates complex channel attention block features, compressing the number of input channels to 1, and generating complex spatial attention. ,Right now:

[0152] (7)

[0153] In equation (7):

[0154] This indicates that the real and imaginary parts of the complex channel attention block features are calculated by applying the complex ReLU and complex sigmoid activation functions, respectively.

[0155] Considering that spatial dependencies in the Fourier domain are generally weaker than those in the spatial domain, this invention sets the kernel size of the two complex convolutions to... Instead of a larger size, this reduces redundant computations while preserving basic Fourier domain information.

[0156] S30402: The final output complex spatial attention block features are obtained by weighting the input complex channel attention block features using complex spatial attention blocks.

[0157] (8)

[0158] S305: Attention Block Features in Complex Space By applying IFFT (Inverse Fast Fourier Transform), the enhanced high-frequency features of the i-th level are obtained. .

[0159] S306: The second step is to refine the information using multi-scale data. Considering that the target is the same as that of the low-frequency enhancer, except for the enhanced object, the same structure as the low-frequency enhancer is used to obtain the output of the high-frequency enhancer. .

[0160] S4: Use a segmentation decoder to generate a segmentation map of diabetic retinopathy lesions;

[0161] The segment decoder consists of five decoder blocks. Composition. Each decoder block Structure such as Figure 5 As shown.

[0162] Inspired by PMCNet, the segmentation decoder of this invention aims to further enhance semantic and detail information by integrating features from adjacent encoders, and aggregates the enhanced multi-level features for prediction, such as... Figure 3 As shown in (e).

[0163] S401: Enhanced low-frequency features are obtained through inverse discrete wavelet transform. and enhanced high-frequency features Reconstructed as wavelet-enhanced features ;

[0164] S402: Aggregate adjacent encoder layer information using residual methods to obtain aggregated adjacent encoder features:

[0165] (9)

[0166] In equation (9):

[0167] express Convolutional layer;

[0168] Considering low-level encoder features While it contains rich details, it also contains high-frequency noise, so this invention directly uses the low-frequency component. To integrate low-level information.

[0169] S403: Enhance wavelet features and aggregated adjacent encoder features The final enhanced feature is obtained by adding them together. ;

[0170] S404: In order to obtain each decoder block The output will ultimately enhance the features. With upsampling features Integration and Convolutional layer processing yields the first Each decoder block output :

[0171] (10)

[0172] Specifically:

[0173]

[0174]

[0175]

[0176] S5: Train the model using the loss function.

[0177] To ensure better aggregation in the low-frequency and high-frequency enhancers, this invention applies auxiliary supervision to the lowest-level outputs of these two modules. Simultaneously, primary supervision is applied to the final segmentation map predicted in the segmentation decoder.

[0178] In practice, the lowest-level outputs of the split decoder, low-frequency enhancer, and high-frequency enhancer are respectively... , and Input to The loss function is obtained from the convolutional layer, upsampling layer, and sigmoid activation function. The loss function described in S5 is as follows:

[0179] (11)

[0180] In equation (11):

[0181] This represents the segmentation map output by the segmentation decoder;

[0182] This represents the output split diagram of the low-frequency booster.

[0183] This represents the output segmentation diagram of the high-frequency booster.

[0184] Indicates the true label;

[0185] These are the weights of the auxiliary loss function, which are empirically set to 0.5;

[0186] Let represent the dice loss function, where:

[0187] K is the total number of lesion categories;

[0188] This represents the prediction result for the k-th type of lesion;

[0189] This represents the true label of the k-th type of lesion;

[0190] This represents the l1 norm.

[0191] Example 1:

[0192] The performance of this invention was evaluated on two popular datasets, IDRid and DDR.

[0193] This invention uses three evaluation criteria: mean area under the precision-recall curve (mAUPR), mean dice coefficient (mDice) per class, and mean intersection-union ratio (mIoU) per class.

[0194] Table 1 shows the performance comparison of the present invention with the existing state-of-the-art methods on the IDRid dataset.

[0195] Table 2 shows the performance comparison of the present invention with the existing state-of-the-art methods on the DDR dataset.

[0196] Table 1. Quantitative comparisons on the IDRiD dataset

[0197]

[0198] Table 2 Quantitative comparisons on the DDR dataset

[0199]

[0200] As can be seen from Tables 1 and 2, the present invention achieves state-of-the-art results on all evaluation metrics on both datasets.

[0201] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0202] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for segmentation of diabetic retinopathy based on wavelet frequency decomposition and enhanced network, characterized in that: The method comprises the following steps: S1: using a discrete wavelet transform to perform high-low frequency decomposition on the multi-level encoder features; S2: using a low-frequency enhancer to enhance the low-frequency features; S3: using a high-frequency enhancer to enhance the high-frequency features; S4: using a segmentation decoder to generate a segmentation map of the diabetic retinopathy; S5: training the model using a loss function.

2. The method according to claim 1, wherein the method is characterized by: The S1 comprises the following steps: S101: Utilize the backbone network by a CNN-based encoder and extract multi-level encoder features from the fundus image after adjusting the number of channels by two convolutional layers , wherein: H is the height of the two-dimensional encoder feature map of the level, W is the width of the two-dimensional encoder feature map of the level, , and represents the level from bottom to top; S102: Adopting discrete wavelet transform to decompose the first layer of the multi-level encoder feature and four filters are obtained. （1） In formula (1): , , , denotes four wavelet subbands; , , , denotes four filters; is a convolution operation with a step size of 2; S103: Because the wavelet sub-band is the first layer low-frequency component of the two-dimensional encoder feature map , therefore, the first layer low-frequency feature , so it can be obtained that the low-frequency feature obtained after the multi-layer encoder feature is decomposed by using the discrete wavelet transform is ; S104: Consider the wavelet sub-band 、 and contains the first level two-dimensional encoder feature map all high-frequency information, therefore, using cross-channel splicing operation to merge all high-frequency components to obtain the first level high-frequency feature , so can be obtained: using discrete wavelet transform to decompose the multi-level encoder feature obtained after the high-frequency feature is .

3. The method according to claim 2, wherein the method is characterized by: The S2 comprises the following steps: S201: pass through convolutional layer to the highest level of low-frequency features enhanced to obtain enhanced highest level of low-frequency features ; S202: obtaining the first feature of the first image hierarchical enhanced low-frequency features Up-sampling the feature by 2 times using bilinear interpolation operation ; S203: upsample the features with the first hierarchical low-frequency features are fused by an element-wise addition operation; S204: Apply The convolution layer strengthens the fused features to obtain the low-frequency features of the first layer The hierarchical enhancement low-frequency features : （2） In formula (2): representing convolution operation.

4. The method according to claim 3, wherein the method is characterized by: The S3 comprises the following steps: S301: apply FFT to obtain the Fourier spectrum of the first level high frequency feature :​ （3） In formula (3): representing the fourier spectrum of the high frequency features of the real part of the fourier spectrum representing the high frequency features of the fourier spectrum imaginary part; denotes the imaginary unit; S302: enrich the first feature through a complex convolution layer high-frequency features of the second level Fourier spectrum of the first feature output complex convolution features ; S303: input the complex convolution feature to the complex channel attention block to output a complex channel attention block feature ; S304: input the complex channel attention block features to the complex spatial attention block to output complex spatial attention block features ; S305: complex space attention block features application of IFFT, obtain the enhanced high-frequency features of the first level​ S306: Obtain the output of the high frequency enhancer using the same structure as the low frequency enhancer .

5. The method according to claim 4, wherein the method is characterized by: The operation of the complex convolution layer in S302 comprises the following steps: S30201: performing a two-dimensional complex convolution operation: （4） In formula (4): denotes a complex convolution kernel, where: denotes the real part of the complex convolution kernel, denotes the imaginary part of the complex convolution kernel; denotes an input complex variable, where: denotes the real part of the input complex variable, denotes the imaginary part of the input complex variable; denotes a complex-valued output, wherein: denotes the real part of the complex-valued output, denotes the imaginary part of the complex-valued output; S30202: complex batch normalization on real part of complex-valued output and imaginary part of complex-valued output both S30203: The real part of the complex value output obtained from S30202. and the imaginary part of complex value output All activation operations are performed using the complex ReLU function.

6. The method according to claim 5, wherein the method is characterized by: The operation of the complex channel attention block in S303 comprises the following steps: S30301: the spatial information of the complex convolution feature is aggregated to obtain a pooled feature through channel global average pooling ; the spatial information of the complex convolution feature is aggregated to obtain a pooled feature through global maximum pooling ; S30302: obtain the complex value channel attention through one-dimensional complex convolution and complex sigmoid function after combining the two pooling features by addition That is: （5） In formula (5): denotes applying a complex sigmoid function to the real and imaginary parts of the complex convolutional features, respectively; denotes a one-dimensional complex convolution; S30303: weighting the input complex convolution features by the complex-valued channel attention to obtain the final output complex channel attention block features: （6） In formula (6): ⊙ represents a Hadamard product operation.

7. The method according to claim 6, wherein the method is characterized by: The operation of the complex spatial attention block in S304 comprises the following steps: S30401: uses two bands with complex ReLU and complex sigmoid activation functions Two-dimensional complex convolution aggregates complex channel attention block features, generating complex spatial attention That is: （7） In formula (7): represents the computation of applying complex ReLU and complex sigmoid activation functions to the real and imaginary parts of the complex channel attention block feature, respectively; S30402: weighting the input complex channel attention block features by the complex spatial attention block to obtain the final output complex spatial attention block features: （8）。 8. The method according to claim 7, wherein the method is characterized by: The S4 comprises the following steps: S401: reconstruct the enhanced low-frequency feature and the enhanced high-frequency feature into a wavelet enhanced feature by inverse discrete wavelet transform ;​​ S402: aggregating adjacent encoder layer information in a residual manner to obtain aggregated adjacent encoder features: （9） In formula (9): denotes convolutional layer S403: add the wavelet enhanced features and the aggregated adjacent encoder features to obtain the final enhanced features ; S404: The decoder fuses the final enhanced features with the up-sampled features and processes them through convolutional layers to obtain the output of the decoder block : （10）。 9. The method according to claim 8, wherein the method is characterized by: The loss function in S5 is as follows: （11） In formula (11): denotes a partition decoder outputting a partition map; represents the low frequency enhancer output segmentation map; represents the high frequency enhancer output segmentation map; represents the true label; is a weight of the auxiliary loss function; represents the dice loss function, where: K is the total number of lesion categories; represents the prediction result of the kth lesion; ykdenotes the true label of the kth lesion; denotes the l1 norm.

Images