Glaucoma picture detection method based on deep learning
Through the improved DRU-NET network model, combined with variable accumulation and residual networks, the problem of lack of data set resources in the existing glaucoma diagnostic method and the lack of segmentation accuracy in the existing glaucoma diagnosis method, and achieved higher diagnosis accuracy and stability , Meet the needs of large -scale glaucoma census.
Patent Information
- Application Number
- CN202411388965.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-08
- Publication Date
- 2025-01-21
AI Technical Summary
The existing glaucoma diagnostic methods have problems such as lack of data set resources, low segmentation accuracy of segmentation networks, and poor classification accuracy of classification networks, which are difficult to meet the needs of large-scale glaucoma census.
The improved DRU-Net network model is adopted, combined with variable accumulation and residual networks, and the pre-processing and feature extraction of the bottom image of the eye. The accuracy and stability of the blue light eye picture detection is calculated through ellipse fitting and clinical indicators.
It improves the accuracy and stability of the bottom image segmentation, enhances the ability to adapt to the bottom of the eyes of different lesions, improves the accuracy and robustness of glaucoma diagnosis, and meets the needs of large -scale census.
Smart Images
Figure CN119338766A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a glaucoma image detection method based on deep learning. Background Art
[0002] Glaucoma is an eye disease that causes irreversible damage to the optic nerve with a slow onset, no obvious early symptoms, and a high blindness rate. The early morphological features of the fundus structure mainly include optic atrophy, pathological depression of the optic disc, and loss of the disc edge. The clinical symptoms in the middle and late stages mainly include abnormally elevated intraocular pressure, dizziness, eye pain, continuous loss of vision, and a gradual narrowing of the visual field. Therefore, it is particularly important to promote large-scale glaucoma screening (especially for middle-aged and elderly high-risk groups) and early prevention and treatment of glaucoma in the community.
[0003] In the past, doctors mainly relied on their own clinical experience to analyze the fundus images of patients to diagnose early glaucoma. This method is time-consuming and laborious, and requires a high level of medical literacy from the doctors themselves. It also carries the risk of clinical misdiagnosis. In the face of large-scale glaucoma screening, this method that relies on manual analysis cannot meet medical needs at all.
[0004] With the rapid development of computer vision, many scholars have proposed many new deep learning network models and image processing algorithms to analyze medical images in order to diagnose glaucoma. For example, the image processing technology based on machine vision is used to segment the optic cup, optic disc, retinal blood vessels and other structures of the fundus image, and then various glaucoma indicators are calculated based on the segmented structures to judge the patient's condition. Although this can greatly reduce the workload of doctors, there are many problems in this field at present, such as the lack of fundus data set resources, the inability to meet the requirements of large-scale glaucoma census, the inability of existing segmentation networks to stably and accurately segment fundus images of different lesions, and the classification accuracy of existing classification networks is not ideal.
[0005] Specifically, the existing related research still has the following problems:
[0006] First, there are currently few publicly available fundus image datasets, and some datasets lack manual annotation of optic cups and discs. At the same time, the image quality of different datasets varies, and images of different contrast and brightness will have varying degrees of impact on the segmentation and classification results.
[0007] Second: The existing fundus image optic disc and optic cup segmentation methods cannot achieve good segmentation results for the irregular deformation of the fundus with severe lesions. The existing deep segmentation network cannot fully learn and utilize the characteristics of the fundus structure. The shallow features are not closely connected with the deep features, resulting in the extracted image features cannot accurately describe the contours of the optic cup and optic disc.
[0008] Third: Existing glaucoma diagnosis methods generally have problems such as unsatisfactory diagnostic accuracy and unstable diagnostic effects of cross-validation of multiple data sets. This is mainly because data sets are relatively scarce and the image sizes of different fundus data sets vary, which increases the probability of learning redundant information during model training. At the same time, existing deep classification networks are insufficiently optimized, the network classification accuracy is not high enough, and the robustness is poor. Summary of the invention
[0009] In order to overcome or alleviate one or more of the above technical problems, the purpose of the present invention is to provide a glaucoma image detection method based on deep learning.
[0010] The present invention provides the following technical solutions:
[0011] A glaucoma image detection method based on deep learning, comprising the following steps:
[0012] S1: Preprocessing of fundus images. For optic cup segmentation, use the smallest possible ROI image containing the optic disc as input to reduce external interference. Locate the optic disc on the grayscale fundus image based on the two features of circular shape and high brightness, extract the optic disc ROI area, and finally use the restricted contrast adaptive histogram equalization method to enhance the contrast and brightness of the optic disc ROI image, and resize it to 256×256 to obtain the preprocessed fundus image.
[0013] S2: Establish an improved DRU-net network model. On the basis of simplifying the U-net network, the variable convolution and residual network are integrated to establish the DRU-net network model, which is used to solve the problem of low segmentation accuracy of the segmentation network and unstable segmentation effect of different fundus lesions;
[0014] S3: Establishing a fundus image data set, making the fundus image preprocessed in step S1 into a data set, and generating a fundus image data set that can be recognized and utilized by the DRU-net network model;
[0015] S4: feature extraction, training the DRU-net network model to extract features based on the fundus image data set obtained in step S3, and selecting the Adam optimizer to continuously optimize and adjust the network parameters to obtain an accurate segmentation map of the image optic cup and optic disc;
[0016] S5: ellipse fitting, using the optic disc segmentation map obtained in step S4 and extracted features through DRU-net network training and output as the input of the ellipse fitting technology to generate the final optic disc segmentation result map;
[0017] S6: Calculate, by fitting the disc edge of the optic disc segmentation result image obtained in step S5 to calculate the cup-to-disc ratio and the retinal edge ratio as clinical detection indicators for glaucoma.
[0018] According to some embodiments, in step S2, the DRU-net network model adopts a residual convolution block, which uses residual mapping to cascade and superimpose the output results of several convolutions. The first two 3×3 convolution layers extract multi-dimensional features, and the last 1×1 convolution layer extracts features from the previous output again to reduce the amount of network calculation; a LeakyReLU activation function and a batch normalization layer are added before each convolution layer.
[0019] According to some embodiments, in step S5, after the optic cup and optic disc of the fundus image are segmented, the Dice coefficient and intersection-over-union ratio indicators are used to measure the performance of the algorithm in segmentation. The Dice coefficient calculates the similarity of two samples by the principle of set similarity and is widely used in the field of image semantic segmentation. The specific formula is as follows:
[0020]
[0021] Among them, X represents the experimental segmentation result of the optic cup and optic disc, and Y represents the corresponding manual segmentation annotation. |X| and |Y| represent the sum of the corresponding image pixels, respectively, and |X∩Y| is the overlapped part of the two images. From the formula, it can be seen that the value range of the Dice coefficient is [0,1];
[0022] Since the Dice coefficient can intuitively reflect the performance of the network, LOSSDice is used as the loss function of this network. The formula is as follows:
[0023] LOSSDice=-log(x,y) (3)
[0024] The IOU indicator is a classic indicator for measuring the performance of image semantic segmentation and target detection algorithms. The formula is as follows:
[0025]
[0026] Where A represents the experimental segmentation result of the optic cup and optic disc, B represents the corresponding manual segmentation annotation, A∩B represents the overlapping part of the two, and A∪B represents the union image of the two.
[0027] According to some embodiments, in step S6, CDR and NRR are calculated by segmenting the optic disc and cup regions, and CDR is obtained by calculating the area ratio of the optic cup region to the optic disc region, and the formula is:
[0028]
[0029] Among them, S cup Represents the area within the entire optic cup outline, N disc It represents the area within the entire optic disc outline including the optic cup area;
[0030] NRR is used to measure whether the fundus under test follows the ISNT rule. The formula is as follows:
[0031]
[0032] Among them, S Superior , S Inferior , S Nasal , S Temporal They represent the areas of the superior, inferior, nasal, and temporal regions of the disc rim, respectively;
[0033] When NRR>1, the eye is detected as normal; when NRR≤1, the fundus is detected to have a high risk of suspected glaucoma. If CDR>0.67 at the same time, the patient is detected as having strong glaucoma.
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] First, to address the problem of uneven data sets, a simplified U-net network is used to extract the optic disc ROI area. This method can quickly obtain the optic disc ROI image; data expansion is used to process all sample data, and the CLAHE image enhancement algorithm is used to improve the contrast and brightness of the optic disc area of the fundus.
[0036] Second, to address the problem of inaccurate segmentation of the optic cup and optic disc in the fundus, variable convolution and residual modules were introduced in the improvement of U-net to improve the utilization of shallow features and the segmentation accuracy of the fundus with different lesions; all convolution layers were optimized using the LeakyReLU activation function and batch normalization layer to improve the generalization ability of the network.
[0037] Third, to address the problem of unstable glaucoma image detection, the CDR and NRR are calculated by designing a mask and fitting the disk edge to detect glaucoma images.
[0038] Fourth, in order to address the complex problems of complex neural networks, by reducing the number of convolution kernels in deep convolutional layers and using 1×1 convolution to reduce network operation parameters, the amount of network calculations and the network training time can be reduced; by optimizing the convolution structure, the generalization ability of the network can be improved and network overfitting can be reduced.
[0039] Fifth, the image preprocessing method used reduces the redundant information on the fundus image, enhances the brightness and contrast of the optic disc area on the image, and expands the number of data set samples, so that the network training can focus more on learning the feature information related to glaucoma classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 This is a structural diagram of the DRU-net model provided in an embodiment of the present invention.
[0041] Figure 2 A residual convolution structure block provided by an embodiment of the present invention.
[0042] Figure 3 The optimized convolution provided by the embodiment of the present invention.
[0043] Figure 4 Flow chart of DRU-net segmentation of visual cup and visual disc provided by the embodiment of the present invention
[0044] Figure 5 The comparison results of the DRU-net optic cup and optic disc segmentation experiment provided in the embodiment of the present invention.
[0045] Figure 6 Four-side masks of the retinal disc edge provided by the embodiment of the present invention.
[0046] Figure 7 The present invention provides a glaucoma ocular disc edge extraction and index calculation process.
[0047] Figure 8 The comparison results of the glaucoma diagnosis experiment based on the segmented images provided in the embodiment of the present invention.
[0048] Fig. 9 A flowchart of a glaucoma image detection method based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0049] The present invention is described in detail below in conjunction with the embodiments and drawings, but it should be understood that the embodiments and drawings are only used to exemplify the present invention and do not constitute any limitation on the protection scope of the present invention. All reasonable changes and combinations within the scope of the inventive concept of the present invention fall within the protection scope of the present invention.
[0050] The present invention will be further described below in conjunction with the accompanying drawings.
[0051] Example 1
[0052] like Fig. 9 , this embodiment provides a glaucoma image detection method based on deep learning, which includes the following steps:
[0053] S1: Fundus image preprocessing,
[0054] For optic cup segmentation, the smallest possible ROI image containing the optic disc is used as input to reduce external interference; the position of the optic disc is located on the grayscale fundus image based on the two features of circular shape and high brightness, and the optic disc ROI area is extracted. Finally, the contrast and brightness of the optic disc ROI image are enhanced using the restricted contrast adaptive histogram equalization (CLAHE) method, and the size is scaled to 256×256 (at this accuracy, the model training speed and accuracy can be better balanced), and the preprocessed fundus image is obtained as the input image for optic cup segmentation;
[0055] S2: Establish an improved DRU-net network model,
[0056] The model is based on the simplified U-net network and integrates variable convolution and residual network to establish the DRU-net network model, which is used to solve the problem of low segmentation accuracy of the segmentation network and unstable segmentation effect of fundus with different lesions;
[0057] S3: Establish fundus image dataset,
[0058] The fundus image preprocessed in step S1 is made into a data set, and a fundus image data set that can be recognized and utilized by the DRU-net network model is generated;
[0059] S4: Feature extraction,
[0060] Based on the fundus image dataset obtained in step S3, the DRU-net model is trained to extract features, and the Adam optimizer is selected to continuously optimize and adjust the network parameters so as to obtain an accurate segmentation map of the image optic cup and optic disc;
[0061] S5: ellipse fitting,
[0062] The optic disc segmentation map obtained in step S4 and outputted through feature extraction by DRU-net network training is used as the input of the ellipse fitting technology to generate the final optic disc segmentation result map;
[0063] S6: Calculation,
[0064] The optic disc segmentation result image obtained in step S5 is fitted with the disc edge to calculate the cup-to-disc ratio and the retinal edge ratio, which are used as clinical detection indicators for glaucoma.
[0065] like Figure 1 , which is the DRU-net model structure diagram obtained by step S2, and its network structure is mirror-symmetrical, and five convolutional layers constitute the image encoding / decoding area (where all convolutions in the first two layers of the network are fixed convolutions, and the middle convolutions in the last three layers are variable convolutions);
[0066] like Figure 2The figure shows the structure of the residual convolution block used in the DRU-net model obtained in step S2. The convolution block uses residual mapping to cascade and stack the output results of several convolutions. The first two 3×3 convolution layers extract multi-dimensional features, and the last 1×1 convolution layer extracts features from the previous output again and significantly reduces the amount of network calculation.
[0067] Add the LeakyReLU activation function and batch normalization layer before each convolution layer. The LeakyReLU activation function can solve the problem that the output of the traditional ReLU function is 0 in the negative part. Its formula is shown in Equation 1. The batch normalization layer can normalize the input of each batch, which helps to speed up the network training process and improve the generalization ability of the model.
[0068]
[0069] The batch normalization function (BN) is introduced and added before the activation function to optimize the convolution module, such as Figure 3 As shown;
[0070] In the upsampling module, the number of feature channels (i.e., the number of convolution kernels) of the feature map of the next layer after each downsampling is set to twice that of the previous layer;
[0071] Furthermore, after the optic cup and optic disc of the fundus image are segmented in step S5, the Dice coefficient and intersection-over-union (IOU) indicators are used to measure the performance of the algorithm in segmentation. The Dice coefficient calculates the similarity between two samples based on the principle of set similarity and is widely used in the field of image semantic segmentation. The specific formula is as follows:
[0072]
[0073] Among them, X represents the experimental segmentation result of the optic cup and optic disc, and Y represents the corresponding manual segmentation annotation. |X| and |Y| represent the sum of the corresponding image pixels, respectively, and |X∩Y| is the overlapped part of the two images. From the formula, it can be seen that the value range of the Dice coefficient is [0,1].
[0074] Since the Dice coefficient can intuitively reflect the performance of the network, LOSSDice is used as the loss function of this network, and its formula is:
[0075] LOSSDice=-log(x,y) (3)
[0076] The IOU indicator is a classic indicator for measuring the performance of image semantic segmentation and target detection algorithms. The formula is as follows:
[0077]
[0078] A represents the experimental segmentation result of the optic cup and optic disc, B represents the corresponding manual segmentation annotation, A∩B represents the overlapped part of the two, and A∪B represents the union image of the two. From the formula, it can be seen that the larger the value of the IOU indicator, the more overlapped parts of the AB images, the greater the correlation, and the better the performance of the algorithm. The best result is that A and B completely overlap, and the correlation is 1; the worst result is that A and B are completely unrelated, and the correlation is 0. Therefore, the value range of the IOU indicator is also [0,1]. Define the overlap error OE as another indicator for evaluating the experiment, OE = 1-IOU, the smaller the OE, the better the segmentation effect.
[0079] Furthermore, in step S6, CDR and NRR are calculated by segmenting the optic disc and optic cup areas. These clinical indicators can accurately assess the degree of optic nerve damage in patients and scientifically and comprehensively judge the progression of glaucoma in patients. The cup-to-disc ratio (CDR) is one of the important criteria for measuring whether glaucoma is present or not. CDR can be obtained by calculating the area ratio of the optic cup area to the optic disc area. The formula is:
[0080]
[0081] Among them, s cup Represents the area within the entire optic cup outline, s disc Represents the area within the entire optic disc outline including the optic cup area.
[0082] The retinal edge ratio (NRR) is used to measure whether the fundus under test follows the ISNT rule. The formula is as follows:
[0083]
[0084] Among them, S Superior , S Inferior , S Nasal , S Temporal Represents the area of the superior, inferior, nasal and temporal regions of the disc respectively.
[0085] Generally, the NRR of a normal eye is > 1. Once the disc edge of the fundus violates the ISNT rule, the NRR will be ≤ 1. At this time, it can be determined that there is a high risk of suspected glaucoma in the fundus. If the CDR is > 0.67 at the same time, the patient can be clinically diagnosed as having strong glaucoma.
[0086] Example 2
[0087] This embodiment tests the DRU-net network segmentation effect and detection results, and the test steps are as follows:
[0088] Step 1) Select the RIM-ONE V3 dataset as the preprocessing dataset. This dataset can be used for retinal fundus structure segmentation and glaucoma classification. It not only provides expert segmentation annotation maps of the optic cup and optic disc, but also provides glaucoma classification labels. By performing operations such as up and down translation, scaling, clockwise rotation of 15 degrees, horizontal and vertical flipping on each sample image, the total number of samples in the RIM-ONE V3 dataset is expanded by 32 times.
[0089] Step 2) Configure the test environment. Randomly select 30% of the samples as the validation set and the remaining samples as the training set. The computer hardware configuration for training the network is: RTX2060 graphics card, 16G running memory and i7 10750H processor.
[0090] Step 3) Train the DRU-net model, select the Adam optimizer to optimize the network parameters, set the learning rate Ir to 0.0001, the dropout to 50%, the epoch to 200, and continuously adjust the parameters during training.
[0091] Step 4) Figure 4 As shown in the figure, the overall process of DRU-net segmentation of optic cup and optic disc is as follows:
[0092] First, the validation set samples are taken as input, and the preprocessed fundus images are sent to the first DRU-net network training to extract features and output the preliminary segmentation map of the optic disc, and the final segmentation result map of the optic disc is generated through the ellipse fitting technology; then the optic cup segmentation is performed, and the preprocessed fundus images are further processed to obtain the optic disc ROI image, and the optic disc ROI image is sent to the second DRU_net network training, and the optic cup segmentation result map is obtained through the same operation.
[0093] Step 5): Evaluate the segmentation effect of DRU_net. Figure 5 As shown in the figure, when the number of convolution kernels in the convolution layer is reduced, compared with the unimproved classic U-net, the Dice coefficient of optic disc segmentation decreases by 0.4%, and the OE index increases by 0.5%; while the Dice coefficient of optic cup segmentation decreases by 0.5%, and the OE index increases by 0.6%.
[0094] Step 6): After performing pixel difference calculation on the segmented optic disc and optic cup, the retinal disc edge of the segmented image is obtained. Based on the centroid of the disc edge, the original disc edge is cropped into a 512x512 image containing the disc edge. According to the ISNT rule, mask images of the four regions of the lower, upper, nasal and temporal sides of the retinal disc edge are prepared respectively. The image size is consistent with the size of the cropped disc edge, as shown in Figure 6 , 7 As shown. Figure 6 .a is the upper side; Figure 6.b is the lower side; Figure 6 .c is the nasal side; Figure 6 .d is the temporal side.
[0095] Step 7): After performing a score operation on the processed disk edge image and the mask images of the four regions, the number of white pixels in each mask region is calculated, and then the CDR and NRR values of the image are calculated according to the formula.
[0096] Step 8): Evaluate the diagnostic performance of the glaucoma images. Figure 8 As shown in the figure, compared with the other four methods, the CDR and NRR indicators designed in the test in this embodiment are 1.67% higher than the wavelet energy characteristics and cup-to-disc ratio in average accuracy, and the accuracy is improved by 3%-8% compared with other methods.
[0097] The above embodiments are only preferred implementations of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. A glaucoma image detection method based on deep learning, characterized in that: The following steps are involved: S1: Preprocessing of fundus images. For optic cup segmentation, use the smallest possible ROI image containing the optic disc as input to reduce external interference. Locate the optic disc on the grayscale fundus image based on the two features of circular shape and high brightness, extract the optic disc ROI area, and finally use the restricted contrast adaptive histogram equalization method to enhance the contrast and brightness of the optic disc ROI image, and resize it to 256×256 to obtain the preprocessed fundus image. S2: Establish an improved DRU-net network model. On the basis of simplifying the U-net network, the variable convolution and residual network are integrated to establish the DRU-net network model, which is used to solve the problem of low segmentation accuracy of the segmentation network and unstable segmentation effect of different fundus lesions; S3: Establishing a fundus image data set, making the fundus image preprocessed in step S1 into a data set, and generating a fundus image data set that can be recognized and utilized by the DRU-net network model; S4: feature extraction, training the DRU-net network model to extract features based on the fundus image data set obtained in step S3, and selecting the Adam optimizer to continuously optimize and adjust the network parameters to obtain an accurate segmentation map of the image optic cup and optic disc; S5: ellipse fitting, using the optic disc segmentation map obtained in step S4 and extracted features through DRU-net network training and output as the input of the ellipse fitting technology to generate the final optic disc segmentation result map; S6: Calculate, by fitting the disc edge of the optic disc segmentation result image obtained in step S5 to calculate the cup-to-disc ratio and the retinal edge ratio as clinical detection indicators for glaucoma.
2. The deep learning-based glaucoma image detection method according to claim 1, characterized in that: In step S2, the DRU-net network model adopts a residual convolution block, which uses residual mapping to cascade and superimpose the output results of several convolutions. The first two 3×3 convolution layers extract multi-dimensional features, and the last 1×1 convolution layer extracts features from the previous output again to reduce the amount of network calculation; a LeakyReLU activation function and a batch normalization layer are added before each convolution layer.
3. The glaucoma image detection method based on deep learning according to claim 2, characterized in that: In step S5, after the optic cup and optic disc of the fundus image are segmented, the Dice coefficient and intersection-over-union ratio are used to measure the performance of the algorithm in segmentation. The Dice coefficient calculates the similarity between two samples based on the principle of set similarity and is widely used in the field of image semantic segmentation. The specific formula is as follows: Where X represents the experimental segmentation result of the optic cup and optic disc, and Y represents the corresponding manual segmentation annotation; |X and |Y| represent the sum of the corresponding image pixels, respectively, and |X∩Y| is the overlapped part of the two images. From the formula, it can be seen that the value range of the Dice coefficient is [0, 1]; Since the Dice coefficient can intuitively reflect the performance of the network, LOSSDice is used as the loss function of this network. The formula is as follows: LOSSDic e =-log(x,y) (3) The I0U indicator is a classic indicator for measuring the performance of image semantic segmentation and target detection algorithms. The formula is as follows: A represents the experimental segmentation result of the optic cup and optic disc, B represents the corresponding manual segmentation annotation, and A ∩ B represents the overlapped part of the two. A∪B Represents the union image of the two.
4. The deep learning-based glaucoma image detection method according to claim 3, characterized in that: In step S6, CDR and NRR are calculated by segmenting the optic disc and optic cup regions, and CDR is obtained by calculating the area ratio of the optic cup region to the optic disc region, and the formula is: Among them, S cup represents the area within the entire optic cup outline, S disc It represents the area within the entire optic disc outline including the optic cup area; NRR is used to measure whether the fundus under test follows the ISNT rule. The formula is as follows: Among them, S Superioc , S Inferior , S Nasal , S Temporal They represent the areas of the superior, inferior, nasal, and temporal regions of the disc rim, respectively; When NRR>1, the eye is detected as normal; when NRR≤1, the fundus is detected to have a high risk of suspected glaucoma. If CDR>0.67 at the same time, the patient is detected as having strong glaucoma.