Method, system, device and medium for segmenting fundus lesions in OCT images based on uncertainty

Abstract

The present invention relates to the field of medical image processing technology. The present invention discloses a method, system, device, and medium for segmenting fundus lesions in uncertainty-based OCT images, including: constructing an encoder-decoder network architecture; embedding an uncertainty guidance module at the encoder end to generate an uncertainty map and weightedly enhance input features; embedding a cross-layer fusion module in the jump connection between the encoder and decoder to process features at different levels through parallel maximum fusion and multiplication fusion branches, generating fused features that are then passed to the decoder; performing independent uncertainty quantification at the decoder end to generate uncertainty maps for each lesion category; and optimizing network parameters using a total loss function combining a target-level uncertainty incentive loss function and a pixel-level uncertainty loss function. This method can significantly improve the automatic recognition and segmentation performance of complex lesion areas, facilitating the accurate quantitative analysis of ophthalmic diseases.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing technology, and in particular to a method, system, device and medium for segmenting fundus lesions in uncertainty-based OCT images. Background Art

[0002] Fundus lesions can cause multiple lesions in the retina simultaneously. For example, common lesions of macular holes include macular holes (MH) and intraretinal cysts (IRC), while common lesions of macular edema include subretinal fluid (SRF) and pigment epithelial detachment (PED). Lesions are located in a certain correlation, and joint segmentation of multiple lesions can fully utilize this correlation, allowing the network to focus on the characteristics of different lesion types and avoid missegmentation of different lesion types.

[0003] Advances in optical coherence tomography (OCT) technology allow doctors to rapidly obtain cross-sectional images of a patient's retina, clearly observing the retinal structure and lesion areas, enabling detailed analysis. However, segmenting lesions in retinal OCT images faces several challenges: 1) In the early stages of the disease, some lesions are small, with pixel grayscale levels similar to those of surrounding healthy tissue, making them difficult to distinguish. 2) In the late stages of the disease, the retinal structure may be severely damaged, resulting in a loss of boundaries between lesions. 3) The morphological similarity, close proximity, and grayscale consistency of lesions further exacerbate segmentation challenges.

[0004] Medical image segmentation is the process of extracting regions of interest from medical images, providing quantitative information about structures such as lesions and organs. With the development of convolutional neural networks (CNNs), these technologies have been widely used to segment lesion regions in OCT images. CNNs are known for their excellent feature extraction capabilities and sensitivity to local contextual information, making them suitable for processing details and structures in images. However, the process by which they produce predictions is completely opaque, lacks interpretability, and may contain biases and errors. Furthermore, the predictions they provide lack the reliability required, making them difficult to convince doctors and patients. The method designed in this patent introduces uncertainty into the segmentation network, identifying areas that require attention while making predictions, thereby providing more accurate medical image segmentation results. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is: in response to the need for accurate identification of different categories of lesions during multi-class segmentation, an uncertainty-guided cross-layer fusion network (UGCFNet) is proposed, which integrates the uncertainty quantification method and can realize fully automatic joint segmentation of various lesion areas in retinal OCT images.

[0007] To solve the above technical problems, the present invention provides the following technical solution, which is a method for segmenting fundus lesions in uncertainty OCT images, comprising the following steps: constructing an encoder-decoder network architecture; embedding an uncertainty guidance module at the end of the encoder, quantizing the uncertainty of the feature map through Gaussian distribution sampling, generating an uncertainty map and weighted enhancing the input features; embedding a cross-layer fusion module in the jump connection between the encoder and the decoder, processing different levels of features through parallel maximum fusion and multiplication fusion branches, generating fused features and then passing them to the decoder; independently quantifying the uncertainty of each channel of the segmentation probability map at the end of the decoder, generating an uncertainty map for each lesion category; combining the target-level uncertainty excitation loss function and the pixel-level uncertainty loss function, and optimizing the network parameters through the total loss function.

[0008] As a preferred solution of the method for segmenting fundus lesions in uncertainty-based OCT images described in the present invention, the encoder-decoder network includes: the decoder adopts an efficient upsampling convolution module, and each level of decoding block sequentially includes 3×3 convolution, 2×2 upsampling, 3×3 depth-separable convolution, batch normalization, ReLU activation layer and 1×1 convolution;

[0009] The 5th level output of the encoder is processed by the uncertainty guidance module and then connected to the 5th level input of the decoder;

[0010] The cross-layer fusion modules are embedded between the 2nd to 4th levels of the encoder and the corresponding levels of the decoder to achieve cross-layer feature fusion.

[0011] As a preferred solution of the method for segmenting fundus lesions in uncertainty-based OCT images according to the present invention, the uncertainty guidance module comprises performing convolution on the input feature map to generate a mean feature map and a standard deviation feature map;

[0012] Sampling a Gaussian distribution with a mean of 0 and a variance of 1 to obtain a random feature map, multiplying the random feature map by σ point by point and then superimposing μ, repeating K times to obtain K random feature maps;

[0013] Calculating the variance of the random feature map and normalizing it to generate an uncertainty map U;

[0014] The uncertainty map U is multiplied element-by-element with the input feature map, and the original input feature map is superimposed to output the enhanced feature map.

[0015] As a preferred solution of the method for segmenting fundus lesions in uncertainty OCT images described in the present invention, the cross-layer fusion module includes: and low-resolution features Perform convolution processing separately to obtain intermediate features and ;

[0016] The intermediate features are fused element-by-element by maximum value and element-by-element by multiplication to obtain ;

[0017] right and Gaussian kernel convolution, 1×1 convolution and Sigmoid activation are performed respectively to generate feature weights P and Q;

[0018] The feature weights and and The enhanced feature maps are obtained by element-by-element multiplication and superimposing the original input maps. After splicing, the final fusion features are generated through 3×3 convolution.

[0019] As a preferred embodiment of the method for segmenting fundus lesions in uncertainty OCT images according to the present invention, the independent uncertainty quantification includes performing convolution on each channel of the segmentation probability map to generate a class mean feature map μ n and category standard deviation feature map σ n ;

[0020] The random feature map is obtained by sampling the Gaussian distribution with mean 0 and variance 1, and the random feature map is compared with σ n Multiply point by point and add μ n , repeat K times to obtain K random feature maps of each category;

[0021] Calculate the variance of the random feature map of the category and normalize it to generate the category uncertainty map ;

[0022] Based on type uncertainty diagram , prediction results and true labels , calculate the loss of each category and take the average to get the total pixel-level uncertainty loss , the specific formula is:

[0023] ,

[0024] in, is the pixel-level uncertainty loss function of the nth category, D represents the number of pixels, express The value of the i-th pixel in , for The value of the i-th pixel in , for The value of the i-th pixel in , n is the category number.

[0025] As a preferred solution of the method for segmenting fundus lesions in uncertainty OCT images described in the present invention, the target-level uncertainty excitation loss function includes combining Dice coefficient loss and cross entropy loss, and adjusting the loss contribution through dynamic weight parameters. The formula is:

[0026] ,

[0027] in, is the target level uncertainty incentive loss, represents the Dice coefficient loss, represents the cross entropy loss, Represents the weight corresponding to the Dice coefficient loss, Represents the weight corresponding to the cross entropy loss.

[0028] As a preferred solution of the method for segmenting fundus lesions in uncertainty OCT images described in the present invention, the total loss function is the target level uncertainty excitation loss and pixel-level uncertainty loss The weighted sum of is:

[0029] ,

[0030] in, is the total loss function, is the balance coefficient.

[0031] Another object of the present invention is to provide a system for segmenting fundus lesions in OCT images based on uncertainty.

[0032] To solve the above technical problems, the present invention provides the following technical solutions: a fundus lesion segmentation system in uncertainty OCT images, comprising: an image preprocessing module, a feature extraction and cross-layer fusion module, an uncertainty guidance and enhancement module, a segmentation decoding and post-processing module, and a loss calculation and network optimization module;

[0033] The image preprocessing module is used to perform standardization processing on the input retinal OCT image, including resizing, normalization and data augmentation operations such as random flipping, cropping and rotation;

[0034] The feature extraction and cross-layer fusion module includes an encoder and a cross-layer fusion module;

[0035] The encoder extracts multi-scale features based on the pre-trained ResNet34 network, capturing global semantics and local details;

[0036] Cross-layer fusion module: Through parallel maximum fusion and multiplication fusion branches, it integrates the features of different layers of the encoder and passes them to the decoder after fusion;

[0037] The uncertainty guidance and enhancement module is embedded at the end of the encoder to generate an uncertainty map and perform weighted enhancement on the input features, so that the network focuses on the blurred and small lesion areas;

[0038] The segmentation decoding and post-processing module is composed of a decoder;

[0039] The decoder uses an efficient upsampling convolution module to restore the feature map resolution level by level and output a segmentation probability map;

[0040] The post-processing unit performs morphological optimization on the segmentation results to improve edge smoothness and segmentation accuracy;

[0041] The loss calculation and network optimization module integrates target-level uncertainty incentive loss and pixel-level uncertainty loss, adjusts the contribution of the two types of losses through dynamic weights, and optimizes network parameters in combination with the total loss function.

[0042] The present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a method for segmenting fundus lesions in uncertainty-based OCT images are implemented.

[0043] The present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method for segmenting fundus lesions in uncertainty-based OCT images are implemented.

[0044] The present invention incorporates uncertainty into the design of retinal OCT image segmentation networks, effectively quantifying perceptual uncertainty and enhancing the network's understanding of blurred and small lesions. Compared to existing techniques, the full uncertainty is obtained in a single forward pass, saving computational resources. This invention proposes an uncertainty-guided cross-layer fusion network for retinal OCT image segmentation. This network innovatively combines a target-level uncertainty-incentive loss with a pixel-level uncertainty loss to improve segmentation accuracy and robustness. During the feature extraction stage, the network uses an uncertainty-guided module (UGM) to model the feature map of the last encoder layer, representing each pixel as a Gaussian distribution. The uncertainty map is calculated by K-fold sampling and variance is generated. This map guides the network to focus on high-uncertainty regions, thereby optimizing feature representation. Furthermore, the cross-layer fusion module (CLF) integrates features from different encoder layers, effectively fusing low-level details with high-level semantic information, enhancing the network's understanding of complex structures. In terms of loss function design, the network adopts a target-level uncertainty-incentive loss, enabling the model to adaptively assign weights to different loss terms. Furthermore, pixel-level uncertainty loss further improves the generalization and reliability of segmentation by quantifying the uncertainty of model predictions. The resulting uncertainty map not only enhances segmentation accuracy but also provides clinicians with a confidence assessment of model predictions, assisting in medical decision-making. The implementation of UGCFNet significantly enhances the automatic recognition and segmentation of complex lesions, facilitating the precise quantitative analysis of ophthalmic diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 A flowchart of a method for segmenting fundus lesions in OCT images based on uncertainty is provided in one embodiment of the present invention.

[0047] Figure 2 A schematic diagram of an uncertainty guidance module of a method for segmenting fundus lesions in uncertainty OCT images provided by one embodiment of the present invention.

[0048] Figure 3 A schematic diagram of a cross-layer fusion module for a method for segmenting fundus lesions in uncertain OCT images provided in one embodiment of the present invention.

[0049] Figure 4The best eight segmentation results in the AI-challenger dataset of the fundus lesion segmentation method in uncertainty OCT images provided in one embodiment of the present invention are displayed.

[0050] In the figure: the eight segmentation results are:

[0051] (a) Original image, (b) True label, (c)UGCFNet, (d)UHA-Net, (e)Unet++, (f)CPFNet, (g)MsTGANet, (h)UA-AHF, (i)Att-Unet, (j)HyFormer. DETAILED DESCRIPTION

[0052] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0053] Example 1, reference Figure 1-Figure 3 , which is the first embodiment of the present invention, provides a method for segmenting fundus lesions in OCT images based on uncertainty, comprising:

[0054] S1: Build the encoder-decoder network architecture.

[0055] Furthermore, Figure 1 The paper demonstrates the Uncertainty-Guided Cross-Layer Fusion Network (UGCFNet) architecture, which represents a significant improvement over the traditional encoder-decoder framework. The core innovation of UGCFNet lies in the novel Uncertainty-Guided Module (UGM) designed at the encoder end. This module dynamically adjusts feature map weights by quantifying the uncertainty in the feature extraction process, allowing the network to focus on areas of high uncertainty. The Cross-Level Fusion (CLF) module integrates feature information at different scales and outputs the segmentation results after progressive decoding. The network calculates both pixel-level uncertainty loss and target-level uncertainty excitation loss through sophisticated uncertainty quantification. This dual optimization mechanism enables the model to focus on lesion areas with the most ambiguous characteristics, significantly improving segmentation accuracy.

[0056] The pre-trained ResNet34 is used as the network encoder. The encoder consists of five levels and extracts multi-scale features through multi-layer convolution. It can simultaneously capture global semantic information and morphological details that are crucial for segmentation tasks.

[0057] The decoder is constructed using the Efficient Upsampling Convolutional Block (EUCB) and consists of five levels of decoding blocks. Each level of decoding block sequentially includes a 3×3 convolution, a 2×2 upsampling layer, a 3×3 depthwise separable convolution, batch normalization, a ReLU activation layer, and a 1×1 convolution. The depthwise separable convolution significantly reduces the number of parameters and improves efficiency. The final 1×1 convolution is used to adjust the number of channels to match the output of the corresponding CLF module.

[0058] An uncertainty guidance module (UGM) is embedded between the fifth-level encoder and the fifth-level decoder. The uncertainty of the underlying feature map extracted by the encoder is quantified in the UGM to enhance the underlying feature map.

[0059] It should be noted that a cross-layer fusion (CLF) module is embedded between the jump connections of the codec. The third CLF module is used to fuse the output features of the UGM with the output features of the fourth-level encoder. Its output is added to the fourth-layer decoding features and input into the third-level decoding block. The second CLF module is used to fuse the output features of the third-level encoder with the output features of the fourth-level encoder. Its output is added to the third-layer decoding features and input into the second-level decoding block. The first CLF module is used to fuse the output features of the second-level encoder with the output features of the third-level encoder. Its output is added to the second-layer decoding features and input into the first-level decoding block. The output features of the first-level encoder are directly added to the first-level decoding features, then upsampled to the original input size and output as the segmentation probability map.

[0060] S2: An uncertainty guidance module is embedded at the end of the encoder to quantify the uncertainty of the feature map through Gaussian distribution sampling, generate an uncertainty map and perform weighted enhancement on the input features.

[0061] Furthermore, Figure 2 As shown in the figure, the input feature F is first passed through two 1×1 convolutional layers to generate the mean feature map μ and the standard deviation feature map σ respectively. In the figure, B, H, and W represent the batch size, height, and width respectively. In order to obtain pixel-level uncertainty, a Gaussian distribution with a mean of 0 and a variance of 1 is sampled at each pixel position. The resulting random feature map is multiplied by σ point by point and then added to μ to obtain a random feature map. The above sampling is repeated K times to obtain a set of random feature maps Set. Then, the variance of Set is calculated point by point and normalized to obtain the uncertainty map ( ). This completes the uncertainty quantification process. It can be expressed as:

[0062] ,

[0063] Where UQ( ) represents uncertainty quantification.

[0064] The uncertainty map is used as the weight-enhanced feature map, and the formula is:

[0065] ,

[0066] in is the input feature map of UGM, is the output feature map of the enhanced UGM, Represents element-wise multiplication.

[0067] S3: A cross-layer fusion module is embedded in the jump connection between the encoder and decoder. Features at different levels are processed through parallel maximum fusion and multiplication fusion branches, and the generated fusion features are passed to the decoder.

[0068] Furthermore, the cross-layer fusion module employs an uncertain spatial attention mechanism for feature fusion, embedded within skip connections. Maximization and multiplication fusion are commonly used methods for feature fusion. To leverage the advantages of both, this module processes these two operations separately through dual parallel branches, concatenating the results.

[0069] like Figure 3 As shown in the figure, C represents the number of channels of the current layer feature map. Assume that in the input of CLF, the high-resolution feature is , the low-resolution features are obtained after 2 times upsampling and Features of the same size , these two features are first processed by 3×3 convolutional layers to obtain and , and then use the element-by-element maximum value and element-by-element multiplication to fuse and get the maximum value respectively. and multiplied value .

[0070] ,

[0071] ,

[0072] ,

[0073] ,

[0074] Where Max() represents the element-level maximization operation, Represents a 3×3 convolution.

[0075] Then, and After being processed by 7×7 Gaussian kernel convolution, 1×1 convolution and Sigmoid activation layer in sequence, the feature weights P and Q are obtained.

[0076] ,

[0077] ,

[0078] represents the sliding window convolution operation, Gaussian convolution kernel representing learnable variance.

[0079] Enhance the generated feature weights respectively and ,get and :

[0080] ,

[0081] ,

[0082] Finally, the features of the two branches are concatenated and fused through the convolutional layer:

[0083] ,

[0084] S4: At the decoder end, each channel of the segmentation probability map is independently quantified to generate an uncertainty map for each lesion category.

[0085] Furthermore, each channel of the segmentation probability map output by the network is independently quantified using the uncertainty quantization method described in S2 to obtain the uncertainty map of each category. .

[0086] ,

[0087] in Indicates the category number, Represents the nth channel of the segmentation probability map.

[0088] For the nth category, based on the uncertainty diagram of the category , prediction results for this category The true label of the category , calculate the pixel-level uncertainty loss function , the expression is:

[0089] ,

[0090] Where D represents the number of pixels, express The value of the i-th pixel in . for The value of the i-th pixel in , for The value of the i-th pixel in .

[0091] Final pixel-level uncertainty loss For different categories Average of:

[0092] ,

[0093] Where N represents the number of segmentation categories.

[0094] S5: Combine the target-level uncertainty incentive loss function and the pixel-level uncertainty loss function to optimize the network parameters through the total loss function.

[0095] The loss function includes region-based loss and distribution-based loss. Dice coefficient loss (DC) and cross entropy loss (CE) are used to implement region-based and distribution-based loss calculations respectively. Therefore, the target level uncertainty excites the loss. It can be expressed as:

[0096] ,

[0097] in, represents the Dice coefficient loss, represents the cross entropy loss, Represents the weight corresponding to the Dice coefficient loss, Denotes the weights corresponding to the cross-entropy loss, and these two weights are dynamically learnable. This method effectively alleviates the impact of uneven category distribution while balancing global and local accuracy, making the model more robust to noisy or ambiguous data, ultimately improving overall segmentation performance.

[0098] Ultimately, the total loss Expressed as:

[0099] ,

[0100] in is the balance coefficient, set to 0.1.

[0101] Example 2, reference Figure 4 , which is an embodiment of the present invention, provides a method for segmenting fundus lesions in OCT images based on uncertainty. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0102] The experimental data used in this invention is a dataset provided by the AI-challenger open challenge. The image resolution is 1024×512. The experimental dataset includes 3D scans of 83 patients, including 10,624 2D OCT images of retinal edema areas (REA), subretinal fluid (SRF), and pigment epithelial detachment (PED) lesions. The manual segmentation labels for each image were manually annotated under the guidance of professional ophthalmologists. The dataset was divided into training, validation, and test sets in a ratio of 32:8:43, with data volumes of 4096, 1024, and 5504, respectively. Before input into the network, the images and manual segmentation labels were downsampled to 512×256 images. Data augmentation strategies such as random folding, cropping, and rotation were added during the training process.

[0103] The proposed UGCFNet was trained in an end-to-end manner with a batch size of 8 for training, validation, and testing. Considering GPU memory and training time costs, OCT images and manually segmented labeled images were downsampled to 512 × 256 pixels before input. The poly learning rate algorithm was applied to optimize the network, with the learning rate lr calculated as follows:

[0104] ,

[0105] The basic learning rate The model was trained on a GeForce RTX 3090 GPU with 24GB of memory and was developed using the public PyTorch platform. The model was trained for 100 epochs. After each epoch, the model was validated on the validation set data. The model parameters with the highest average DSC in the validation set were retained as the final network model parameters for testing. K was set to 50 for the uncertainty quantification.

[0106] a) Segmentation index

[0107] To evaluate the segmentation results of REA, SRF, and PED in OCT images, three evaluation metrics are used: Dice Similarity Coefficient (DSC), Intersection of Union (IoU), Sensitivity (Sen), and Accuracy (ACC). They are defined as follows:

[0108] ,

[0109] ,

[0110] ,

[0111] Where TP, TN, FP, and FN are the true positives, true negatives, false positives, and false negatives of the segmentation, respectively. The metrics for each category are calculated separately and the overall average is calculated.

[0112] b) Ablation experiments on different components of the network

[0113] Table 1 Ablation experiments on different components of the network (%)

[0114] ,

[0115] To verify the effectiveness of the proposed components, we designed an ablation experiment. Table 1 shows the ablation experiment results of all components. The baseline model only contains an encoder-decoder structure with a simple skip connection, and its loss function uses the sum of the Dice coefficient and the cross entropy loss. After the loss, further introduction As shown in Table 1, with the gradual addition of various components, the model's DSC and IoU indicators continue to improve. Ultimately, on both datasets, our method achieves a 5.21% DSC improvement over the baseline model.

[0116] c) Comparative experiment

[0117] Table 2 Segmentation performance of different segmentation networks (%)

[0118] ,

[0119] Table 2 shows a comparison of the segmentation performance of the proposed UGCFNet with different segmentation networks. The proposed network achieves the best performance in PED's single-class DSC, average DSC, and average sensitivity of the average intersection over union (IoU). The average DSC, average IoU, and average Sen metrics are 79.47%, 68.62%, and 79.33%, respectively, demonstrating that the network can accurately capture smaller objects. It also achieves the second-best result in REA's single-class DSC, reaching 80.90%, and a good result in SRF's single-class DSC, reaching 82.29%, demonstrating good and balanced segmentation performance. The experimental results shown in Table 2 demonstrate that the proposed method achieves the best segmentation performance on the AI-Challenger dataset, with an average DSC reaching the highest level of 79.47%. The DSC and IoU metrics for various lesions outperform all compared methods. Figure 4Examples of the segmentation effects of various advanced methods on the AI-Challenger dataset are shown: the first row of results shows that this method accurately segments the SRF area, while most other methods have obvious under-segmentation; the second row contains two types of lesions, REA and PED, and this method more accurately depicts the outline of the REA area; in the third row, compared with other methods, this scheme achieves good segmentation for all three types of lesions (especially PED).

[0120] Experiments demonstrate that the proposed network overcomes the challenges of segmenting multiple lesion types, such as large horizontal spans, widely varying lesion morphologies, large or small areas, and loss of tissue structure leading to the disappearance of lesion boundaries. The proposed hybrid network accurately segments multiple lesion types in retinal OCT images, surpassing several other advanced segmentation networks.

[0121] Example 3 is the third embodiment of the present invention, which differs from the first two embodiments in that:

[0122] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0123] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0124] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0125] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the aforementioned embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or combination of the following technologies known in the art may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0126] Example 4, an embodiment of the present invention, provides a system for segmenting fundus lesions in uncertainty-based OCT images, including an image preprocessing module, a feature extraction and cross-layer fusion module, an uncertainty guidance and enhancement module, a segmentation decoding and post-processing module, and a loss calculation and network optimization module;

[0127] The image preprocessing module is used to perform standardization processing on the input retinal OCT image, including resizing, normalization and data augmentation operations such as random flipping, cropping and rotation;

[0128] The feature extraction and cross-layer fusion module includes an encoder and a cross-layer fusion module;

[0129] The encoder extracts multi-scale features based on the pre-trained ResNet34 network, capturing global semantics and local details;

[0130] CLF module: It integrates features from different levels of the encoder through parallel maximum fusion and multiplication fusion branches, and passes the fusion to the decoder;

[0131] The uncertainty guidance and enhancement module is embedded at the end of the encoder to generate an uncertainty map and perform weighted enhancement on the input features, so that the network focuses on the blurred and small lesion areas;

[0132] The segmentation decoding and post-processing module is composed of a decoder;

[0133] The decoder uses the efficient upsampling convolution module EUCB to restore the feature map resolution step by step and output the segmentation probability map;

[0134] The post-processing unit performs morphological optimization on the segmentation results to improve edge smoothness and segmentation accuracy;

[0135] The loss calculation and network optimization module integrates target-level uncertainty incentive loss and pixel-level uncertainty loss, adjusts the contribution of the two types of losses through dynamic weights, and optimizes network parameters in combination with the total loss function.

[0136] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for segmenting fundus lesions in uncertainty-based OCT images, characterized by: include, Build an encoder-decoder network architecture; An uncertainty guidance module is embedded at the end of the encoder to quantify the uncertainty of the feature map through Gaussian distribution sampling, generate an uncertainty map and perform weighted enhancement on the input features; The uncertainty guidance module includes performing convolution on the input feature map to generate a mean feature map μ and a standard deviation feature map σ; Sampling a Gaussian distribution with a mean of 0 and a variance of 1 to obtain a random feature map, multiplying the random feature map by σ point by point and then superimposing μ, repeating K times to obtain K random feature maps; Calculating the variance of the random feature map and normalizing it to generate an uncertainty map U; Multiply the uncertainty map U by the input feature map element by element, and then superimpose the original input feature map to output the enhanced feature map; A cross-layer fusion module is embedded in the jump connection between the encoder and decoder. Features at different levels are processed through parallel maximum fusion and multiplication fusion branches, and the generated fusion features are passed to the decoder. At the decoder end, each channel of the segmentation probability map is independently quantified to generate an uncertainty map for each lesion category; The independent uncertainty quantification includes performing convolution on each channel of the segmentation probability map to generate a category mean feature map μ n and category standard deviation feature map σ n ; The random feature map is obtained by sampling the Gaussian distribution with mean 0 and variance 1, and the random feature map is compared with σ n Multiply point by point and add μ n , repeat K times to obtain K random feature maps of each category; Calculate the variance of the random feature map of the category and normalize it to generate the category uncertainty map U n ; Based on the type uncertainty diagram U n , prediction results and the true label y n , calculate the loss of each category and take the average to get the total pixel-level uncertainty loss. The specific formula is: in, is the pixel-level uncertainty loss function of the nth category, D represents the number of pixels, Indicates U n The value of the i-th pixel in , y n The value of the i-th pixel in , for The value of the i-th pixel in , n is the category number; The target-level uncertainty excitation loss function and the pixel-level uncertainty loss function are combined to optimize the network parameters through the total loss function.

2. The method for segmenting fundus lesions in OCT images based on uncertainty according to claim 1, wherein: The encoder-decoder network includes a decoder that uses an efficient upsampling convolution module. Each level of decoding block contains 3×3 convolution, 2×2 upsampling, 3×3 depth-separable convolution, batch normalization, ReLU activation layer and 1×1 convolution in sequence; The 5th level output of the encoder is processed by the uncertainty guidance module and then connected to the 5th level input of the decoder; The cross-layer fusion modules are embedded between the 2nd to 4th levels of the encoder and the corresponding levels of the decoder to achieve cross-layer feature fusion.

3. The method for segmenting fundus lesions in OCT images based on uncertainty according to claim 2, wherein: The cross-layer fusion module includes performing convolution processing on the high-resolution feature F1 and the low-resolution feature F2 respectively to obtain the intermediate feature F ' 1 and F ' 2; Perform element-by-element maximum fusion and element-by-element multiplication fusion on the intermediate features to obtain the maximum value F max and the multiplied value F mul ; F max and F mul Gaussian kernel convolution, 1×1 convolution and Sigmoid activation are performed respectively to generate feature weights P and Q; Combine the feature weights with F max and F mul The enhanced feature maps are obtained by element-by-element multiplication and superimposing the original input maps. After splicing, the final fusion features are generated through 3×3 convolution.

4. The method for segmenting fundus lesions in OCT images based on uncertainty according to claim 3, wherein: The target level uncertainty incentive loss function includes combining Dice coefficient loss and cross entropy loss, and adjusting the loss contribution through dynamic weight parameters. The formula is: in, is the target level uncertainty incentive loss, represents the Dice coefficient loss, represents the cross entropy loss, λ DC Represents the weight corresponding to the Dice coefficient loss, λ CE Represents the weight corresponding to the cross entropy loss.

5. The method for segmenting fundus lesions in OCT images based on uncertainty according to claim 4, wherein: The total loss function is the target level uncertainty incentive loss and pixel-level uncertainty loss The weighted sum of is: in, is the total loss function, and λ is the balance coefficient.

6. A system for segmenting fundus lesions in uncertain OCT images, comprising the method for segmenting fundus lesions in uncertain OCT images according to any one of claims 1 to 5, characterized in that: It includes image preprocessing module, feature extraction and cross-layer fusion module, uncertainty guidance and enhancement module, segmentation decoding and post-processing module, loss calculation and network optimization module; The image preprocessing module is used to perform standardization processing on the input retinal OCT image, including resizing, normalization and data augmentation operations such as random flipping, cropping and rotation; The feature extraction and cross-layer fusion module includes an encoder and a cross-layer fusion module; The encoder extracts multi-scale features based on the pre-trained ResNet34 network, capturing global semantics and local details; Cross-layer fusion module: Through parallel maximum fusion and multiplication fusion branches, it integrates the features of different layers of the encoder and passes them to the decoder after fusion; The uncertainty guidance and enhancement module is embedded at the end of the encoder to generate an uncertainty map and perform weighted enhancement on the input features, so that the network focuses on the blurred and small lesion areas; The segmentation decoding and post-processing module is composed of a decoder; The decoder uses an efficient upsampling convolution module to restore the feature map resolution level by level and output a segmentation probability map; The post-processing unit performs morphological optimization on the segmentation results to improve edge smoothness and segmentation accuracy; The loss calculation and network optimization module integrates target-level uncertainty incentive loss and pixel-level uncertainty loss, adjusts the contribution of the two types of losses through dynamic weights, and optimizes network parameters in combination with the total loss function.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images