OCT image intelligent classification processing method and system based on deep learning
Through the deep learning-based intelligent classification processing method of OCT images, the improved convolutional network and MobileNet network are used to solve the problems of low efficiency and low accuracy of OCT images classification, and the rapid and accurate image classification is achieved, and the diagnosis quality is improved. .
Patent Information
- Application Number
- CN202411051907.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the prior art, OCT image classification mainly relies on manual experience, is low efficiency and low accuracy, which affects the diagnosis quality.
The intelligent classification processing method of OCT images based on deep learning is adopted, and noise reduction is performed through the improved convolutional network model, the ROI region is extracted, and the PCA algorithm is used to reduce the feature dimensions, and finally image classification is performed based on the improved MobileNet network.
It realizes rapid and accurate classification of OCT images, reduces the work burden of medical staff, and improves the quality of diagnosis.
Smart Images

Figure CN119091187A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing technology, and in particular to an OCT image intelligent classification processing method and system based on deep learning. Background Art
[0002] OCT images, or optical coherence tomography, are a new type of ophthalmic medical diagnostic method with non-contact and high-resolution features. They have now become an important reference for doctors to diagnose ophthalmic diseases, especially for early detection and clinical diagnosis of retinal diseases. However, there are many types of OCT images, and currently OCT image classification mainly relies on manual experience, which is not only time-consuming and labor-intensive, but also brings a heavy workload to doctors. At the same time, the accuracy of its classification cannot be guaranteed, thus affecting the quality of diagnosis. Summary of the invention
[0003] In order to solve at least one of the technical problems mentioned above, the present invention provides an OCT image intelligent classification processing method and system based on deep learning.
[0004] In a first aspect, the present invention provides an OCT image intelligent classification processing method based on deep learning, the method comprising:
[0005] Acquire an OCT image of the target object, and perform noise reduction processing on the OCT image;
[0006] Extracting an original ROI region from the denoised OCT image, performing data bias field correction processing on the original ROI region, and obtaining a target ROI region;
[0007] Extract local features of each type of feature in the target ROI area, and use the PCA algorithm to reduce the feature dimension of the local features of each extracted feature to generate target features;
[0008] The target features are input into a trained OCT image classification model to generate a classification result.
[0009] Preferably, performing noise reduction processing on the OCT image includes:
[0010] Denoising the OCT image using an improved convolutional network model, wherein the improved convolutional network model includes a sparse local attention module, a depth extraction module, a global attention module, and a residual module;
[0011] The sparse local attention module includes a hole convolution layer, a batch normalization layer, a ReLU activation layer and a convolution layer;
[0012] The depth extraction module adopts a three-layer convolution extraction network, each layer of which includes a batch normalization layer, a ReLU activation layer and a convolution layer;
[0013] The global attention module uses a convolution kernel to compress the features extracted by the sparse local attention module and the deep extraction module to generate a weight vector; the weight vector is used to adjust the network weights of the sparse local attention module and the deep extraction module.
[0014] Preferably, after the OCT image is denoised using the improved convolutional network model, the method further comprises:
[0015] The noise reduction effect is evaluated using evaluation indicators, which are:
[0016]
[0017] In the formula, f PSNR represents the peak signal-to-noise ratio, max(I) represents the maximum grayscale value of the pixel in image I, and f MsE is the mean square error between the denoised image and the true image; W and H represent the width and height of the image respectively, and I d (x, y) represents the grayscale value of the denoised image, I c (x,y) represents the gray value of the true image, and (x,y) represents the coordinates of the image pixel.
[0018] Preferably, extracting the original ROI area from the denoised OCT image includes:
[0019] Based on the denoised OCT image, the blank edges are removed and the blank areas of the image are filled using the average value of the background pixels in the OCT image;
[0020] Performing global threshold processing on the filled image to generate the first retinal structure image;
[0021] Perform median filtering on the threshold processed image to remove black spots in the retina and generate a second retinal structure image;
[0022] Performing quadratic polynomial fitting on the second retinal structure image to generate a fitting curve;
[0023] The images of different retinal structures are aligned, and after the alignment, each column of pixels is translated upward or downward by a number of pixels so that the pixel points on the fitting curve are located on the same horizontal line, and then the original ROI area is cut out.
[0024] Preferably, performing local feature extraction on each type of feature in the target ROI region includes:
[0025] Divide the target ROI area into a plurality of rectangular blocks, and calculate a quantization coefficient of each pixel in each of the rectangular blocks;
[0026] Generating a histogram of quantization coefficients on each rectangular block, merging all the histograms to obtain a local feature of each feature;
[0027] Wherein, the expression of the local feature is:
[0028]
[0029] Where (x, y) is the pixel coordinate, W u represents the basis vectors of the two-dimensional discrete Fourier transform with frequency u, Represents the basis vector W u The transpose of , j represents the jth component, f x N x Medium 2 The vector composed of the grayscale values of pixels, N x is a M×M rectangular neighborhood.
[0030] Preferably, the OCT image classification model is obtained by training a MobileNet network; the MobileNet network comprises:
[0031] A multi-feature fusion module is used to perform multi-feature fusion on the target features to generate a fusion feature map; the multi-feature fusion includes channel layer feature fusion and space layer feature fusion;
[0032] A channel attention module is used to average pool the fused feature map, convolve the compressed image after average pooling twice, activate it through a Sigmoid function, multiply the activated feature map with the input fused feature map, and obtain a channel output feature map;
[0033] The spatial attention module is used to perform maximum pooling and average pooling on the channel output feature map in sequence, input the processed feature map into a convolution layer with a convolution kernel of 8×8, and multiply it with the channel output feature map after activation by a Sigmoid function to generate a final classification feature map.
[0034] Preferably, the expression of the target feature is:
[0035]
[0036] In the formula, F add is the target feature, K i represents the convolution kernel, * represents convolution, D i Retinal flatness, Z i V is the degree of depression of the center of the macula,i is the width of the retinal nerve fiber, i = 1, 2, 3…n.
[0037] In a second aspect, the present invention further provides an OCT image intelligent classification and processing system based on deep learning, the system comprising:
[0038] A noise reduction unit, used for acquiring an OCT image of a target object and performing noise reduction processing on the OCT image;
[0039] The ROI region generating unit is used to extract the original ROI region from the noise-reduced OCT image, and perform data bias field correction processing on the original ROI region to obtain the target ROI region;
[0040] A target feature generation unit is used to extract local features of each type of feature in the target ROI area, and use a PCA algorithm to perform feature dimension reduction on the local features of each extracted feature to generate target features;
[0041] The OCT image classification unit is used to input the target features into a trained OCT image classification model to generate a classification result.
[0042] In a third aspect, the present invention further provides an electronic device comprising: a processor and a memory, wherein the memory is used to store computer program code, wherein the computer program code comprises computer instructions, and when the processor executes the computer instructions, the electronic device executes the method as described in the first aspect above and any possible implementation thereof.
[0043] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, wherein the computer program includes program instructions, and when the program instructions are executed by a processor of an electronic device, the processor executes the method as described in the first aspect above and any possible implementation thereof.
[0044] Compared with the prior art, the present invention has the following beneficial effects:
[0045] The present invention provides an OCT image intelligent classification processing method based on deep learning. First, an OCT image of a target object is obtained, and the OCT image is subjected to denoising. In the denoising process, an improved convolutional network model is mainly used for unsupervised denoising. The model includes a sparse local attention module, a deep extraction module, a global attention module and a residual module. It can still achieve good denoising performance when the data set is insufficient, has strong generalization ability, and provides image samples with better quality for subsequent processing. When extracting the ROI area, the original ROI area is generated by pixel filling, global threshold processing, median filtering, polynomial fitting and image cutting, and the target ROI area is further obtained by bias field correction; when extracting features, the amount of calculation can be greatly reduced by extracting local features of various types of features, and the feature dimension reduction realized by the PCA algorithm, so as to improve the recognition efficiency. Finally, multiple features are fused, and OCT image classification is performed based on the improved MobileNet network, so that the OCT image can be completed quickly and accurately. Compared with manual experience classification, it has the advantages of high efficiency and high accuracy.
[0046] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the background technology, the drawings required for use in the embodiments of the present invention or the background technology will be described below.
[0048] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used to illustrate the technical solutions of the present disclosure together with the specification.
[0049] Figure 1 A schematic diagram of a flow chart of an OCT image intelligent classification processing method based on deep learning provided by an embodiment of the present invention;
[0050] Figure 2 A structural schematic diagram of an OCT image intelligent classification and processing system based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0051] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0052] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.
[0053] The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can represent including any one or more elements selected from the set consisting of A, B, and C.
[0054] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0055] In addition, in order to better illustrate the present invention, numerous specific details are provided in the following specific embodiments. It should be understood by those skilled in the art that the present invention can be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present invention.
[0056] At present, OCT image classification mainly relies on manual experience, which is not only inefficient but also prone to errors. To this end, the present invention provides an OCT image intelligent classification processing method based on deep learning, which can effectively reduce the interference of invalid features through noise reduction and ROI area extraction, and can greatly reduce the amount of calculation through local feature extraction and feature dimension reduction. The OCT image classification model is obtained through improved MobileNet network training, which can realize fast and accurate classification of OCT images, thereby reducing the workload of medical staff and helping to improve the quality of diagnosis.
[0057] See also Figure 1 , Figure 1The present invention provides a flowchart of an OCT image intelligent classification processing method based on deep learning. Figure 1 As shown, a method for intelligent classification and processing of OCT images based on deep learning includes the following steps:
[0058] S10, acquiring an OCT image of the target object, and performing noise reduction processing on the OCT image.
[0059] S20, extracting an original ROI region from the denoised OCT image, and performing data bias field correction processing on the original ROI region to obtain a target ROI region.
[0060] S30, extracting local features of each type of feature in the target ROI region, and performing feature dimension reduction on the local features of each extracted feature using a PCA algorithm to generate target features.
[0061] S40, inputting the target features into a trained OCT image classification model to generate a classification result.
[0062] In this embodiment, the target object generally refers to a patient undergoing an ophthalmic examination. First, an OCT image of the target object is acquired, and then noise reduction processing is performed. After noise reduction, interference noise can be filtered out, and then the region of interest, that is, the original ROI region, is extracted therefrom. In order to optimize the image quality of the ROI region, the original ROI region needs to be biased and corrected to generate a target ROI region. When classifying an image, it is necessary to extract features based on the target ROI region. In this embodiment, in order to reduce the amount of calculation and improve the classification effect, a local feature extraction method is adopted, that is, for each type of feature, its local features are extracted for identification to avoid interference from invalid features. Finally, through multi-feature fusion recognition, by inputting multiple features into the OCT image classification model trained using the MobileNet network, the classification results can be obtained quickly and accurately.
[0063] In some embodiments, performing noise reduction processing on the OCT image includes:
[0064] Denoising the OCT image using an improved convolutional network model, wherein the improved convolutional network model includes a sparse local attention module, a depth extraction module, a global attention module, and a residual module;
[0065] The sparse local attention module includes a hole convolution layer, a batch normalization layer, a ReLU activation layer and a convolution layer;
[0066] The depth extraction module adopts a three-layer convolution extraction network, each layer of which includes a batch normalization layer, a ReLU activation layer and a convolution layer;
[0067] The global attention module uses a convolution kernel to compress the features extracted by the sparse local attention module and the deep extraction module to generate a weight vector; the weight vector is used to adjust the network weights of the sparse local attention module and the deep extraction module.
[0068] In this embodiment, the original feature map is first extracted by a Conv and ReLU activation function in the image input network, and used as the input of the sparse local attention module. The sparse local attention module adopts a strategy of using ordinary convolution and dilated convolution to form a sparse block, which reduces the use of parameters while expanding the receptive field and avoids the "grid effect" of dilated convolution. The module mainly includes: dilated Conv+BN+ReLU, Conv+BN+ReLU layer, Conv layer, Concat layer. Among them, BN is a batch normalization layer and ReLU is an activation layer.
[0069] The dilation rate of the dilated convolution kernel is set to 2, and the size of the convolution kernel is set to 3×3. The Concat layer is a feature fusion layer that uses residual connections to fuse features before and after sparse convolution, making full use of information at different scales to improve the denoising effect. The σ activation function is selected to increase the nonlinear expression ability of the module, and the attention mechanism is applied to guide the convolutional neural network to train the denoising model. The learning results of the previous stage are used to guide the noise learning at the current stage, so that the network automatically devotes more computing resources to the important features of the task to improve efficiency.
[0070] As the network depth increases, the deeper network may be affected by the weakening of the shallow layer on the deeper layer, and the use of dilated convolution will lose the noise reduction accuracy. Therefore, this embodiment provides a deep extraction module, which uses a three-layer convolution extraction network, namely Conv+BN+ReLU network, to perform deep extraction through a long path to improve the noise reduction effect.
[0071] The global attention module is used to redistribute attention to global features. It uses the last Conv with a convolution kernel size of 1×1 to compress the obtained features into a vector and use it as the weight to adjust the previous stage. The global attention module combines local attention with global attention, making full use of local and global information to mine more robust features, and applies the attention mechanism at multiple scales. While extracting local information, it effectively pays attention to global semantic information to achieve in-depth mining of data. Finally, the RB layer is used to remove the real image from the noisy image in the hidden layer, so that the network can be trained using noise information, which can speed up the convergence speed without changing the denoising performance.
[0072] In some embodiments, after the OCT image is denoised using the improved convolutional network model, the method further includes evaluating the denoising effect using an evaluation index, wherein the evaluation index is:
[0073]
[0074] In the formula, f PSNR represents the peak signal-to-noise ratio, max(I) represents the maximum grayscale value of the pixel in image I, and f MSE is the mean square error between the denoised image and the true image; W and H represent the width and height of the image respectively, and I d (x, y) represents the grayscale value of the denoised image, I c (x,y) represents the gray value of the true image, and (x,y) represents the coordinates of the image pixel.
[0075] Therefore, this embodiment uses an improved convolutional neural network to reduce noise in OCT images, and uses evaluation indicators to evaluate the noise reduction effect. When the noise reduction effect does not meet the standard, the network can be iteratively trained until the noise reduction effect meets the preset conditions. Through this unsupervised learning noise reduction method, high-quality noise reduction of OCT images is achieved while having strong generalization ability.
[0076] In some embodiments, extracting the original ROI area from the denoised OCT image includes:
[0077] 2.1) Based on the denoised OCT image, the blank edges are removed and the blank areas of the image are filled using the average value of the background pixels in the OCT image;
[0078] 2.2) performing global threshold processing on the filled image to generate the first retinal structure image;
[0079] 2.3) Perform median filtering on the threshold processed image to remove black spots in the retina and generate a second retinal structure image;
[0080] 2.4) performing quadratic polynomial fitting on the second retinal structure image to generate a fitting curve;
[0081] 2.5) Aligning the images of different retinal structures, after alignment, shifting each column of pixels upward or downward by a number of pixels so that the pixel points on the fitting curve are located on the same horizontal line, and then cutting out the original ROI area.
[0082] It should be noted that spots in retinal OCT images have a serious impact on later images. In addition, during the OCT imaging process, the morphology and position of the retina will change significantly, which makes it impossible to align all the retinas to a relatively uniform position, and the OCT image will contain a large number of irrelevant parts, which will affect the accuracy of image feature extraction and classification. Therefore, before extracting features, it is necessary to extract the ROI area from the image.
[0083] First, the region of interest needs to be detected to obtain the entire retinal morphology, which mainly includes steps 2.1)-2.3). The Otsu algorithm is used in the global threshold processing process. After obtaining the first retinal structure image, the image is median filtered to remove the black spots around the retina; then the morphological closing operation is used to remove the large black spots inside the retina, and the morphological opening operation is used to remove the white spots outside the retina, and finally the second retinal structure image is obtained.
[0084] Furthermore, in the fitting stage of step 2.4), in order to align the retinal areas in different images, the present embodiment extracts two sets of data points from the image obtained after the morphological opening operation: the middle data point (in each column of pixels, the coordinates of the middle data point are obtained by averaging the coordinates of the white pixels) and the bottom data point (in each column of pixels, the bottom white pixel is taken as the bottom data point). Then, a quadratic polynomial is fitted to the middle data point to obtain a fitting curve. If the fitting opening is upward, the middle data point is selected, otherwise the bottom data point is selected; when the middle data point is selected, the data point is continued to be used for linear fitting, and the correlation coefficients of the two fitting point sets and the original middle data point are compared, and finally the fitting method with a high correlation coefficient is selected. When the bottom data point set is selected, a quadratic polynomial is fitted to the bottom data point. If the opening is upward, the data point is continued to be used for linear fitting, and then a fitting method with a large correlation coefficient is selected; if the opening is downward, the linear fitting method is directly selected.
[0085] Finally, in step 2.5), the ROI region is cut. According to the fitting result of step 2.4), the retina is aligned, and then the part that is irrelevant to the identification of the retinal type is cut off. The specific process is: each column of pixels is translated up or down by a number of pixels so that the pixel points on the fitting curve are located on a horizontal line, and then the part of interest of the image is cut out according to the highest and lowest points of the aligned white pixels. In this way, the retinal morphological structure that is conducive to classification can be retained, and the irrelevant interference parts can be removed.
[0086] In some embodiments, after the original ROI area is cut out, data bias field correction processing and normalization processing are also performed to obtain the target ROI area.
[0087] It should be noted that the bias field can cause inconsistent brightness and contrast on the image. This phenomenon is caused by the inhomogeneity of the magnetic field during the scanning process. If the bias field correction is not performed, it may cause errors in the subsequent image analysis and interpretation, thus affecting the accuracy of classification.
[0088] The filter-based method and the histogram-based method are two common bias field correction algorithms. The filter-based method regards the bias field as a low-frequency artifact and uses a low-pass filter to remove it from the high-frequency component of the image. The histogram-based method performs correction based on the intensity distribution of the image. Preferably, the present embodiment uses the histogram-based method to perform bias field correction. After the correction is completed, normalization processing is performed to obtain the target ROI area.
[0089] In some embodiments, performing local feature extraction on each type of feature in the target ROI region includes:
[0090] Divide the target ROI area into a plurality of rectangular blocks, and calculate a quantization coefficient of each pixel in each of the rectangular blocks;
[0091] A histogram of quantization coefficients is generated on each rectangular block, and all the histograms are combined to obtain a local feature of each feature.
[0092] Wherein, the expression of the local feature is:
[0093]
[0094] Where (x, y) is the pixel coordinate, W u represents the basis vectors of the two-dimensional discrete Fourier transform with frequency u, Represents the basis vector W u The transpose of , j represents the jth component, f x N x Medium 2 The vector composed of the grayscale values of pixels, N x is a M×M rectangular neighborhood.
[0095] In this embodiment, local features are usually only considered at four frequency points u1 = [a, 0] T ,u2=[0,a] T ,u3=[a,a] T ,u4=[a,-a] T The Fourier coefficients on , where a is a sufficiently small constant. Let:
[0096] F x =[F(u1,x), F(u2,x), F(u3,x), F(u4,x)];
[0097] The phase information in the Fourier coefficients is given by F x The real and imaginary parts of each component in can be quantized by the following formula:
[0098]
[0099] In the formula, q j (x) is G x =[Re{F x},lm{F x The jth component in the vector is represented by a binary code as an integer in the range [0, 255], specifically:
[0100]
[0101] Where b is the quantization coefficient, Re{F x},lm{F x} respectively represent F x The real and imaginary parts of .
[0102] After obtaining the local features, the PCA algorithm is used to perform feature dimension reduction on the local features of each feature extracted to generate the target feature. Preferably, the expression of the target feature is:
[0103]
[0104] In the formula, F add is the target feature, K i represents the convolution kernel, * represents convolution, D i Retinal flatness, Z i V is the degree of depression of the center of the macula, i is the width of the retinal nerve fiber, i = 1, 2, 3…n.
[0105] It should be noted that OCT images are usually represented in the form of a matrix, and the matrix is composed of multiple vectors, each of which represents certain feature information in the image, such as edges, textures, etc. In multi-feature fusion, the pathological features of the OCT image can be extracted through related operations of convolution, activation function and pooling layer. In the fully connected layer, all feature vectors are flattened into a vector to represent the features of the entire image. Finally, multi-feature fusion is achieved by element-by-element multiplication.
[0106] In some embodiments, the OCT image classification model is obtained by training a MobileNet network; the MobileNet network includes:
[0107] A multi-feature fusion module is used to perform multi-feature fusion on the target features to generate a fusion feature map; the multi-feature fusion includes channel layer feature fusion and space layer feature fusion;
[0108] A channel attention module is used to average pool the fused feature map, convolve the compressed image after average pooling twice, activate it through a Sigmoid function, multiply the activated feature map with the input fused feature map, and obtain a channel output feature map;
[0109] The spatial attention module is used to perform maximum pooling and average pooling on the channel output feature map in sequence, input the processed feature map into a convolution layer with a convolution kernel of 8×8, and multiply it with the channel output feature map after activation by a Sigmoid function to generate a final classification feature map.
[0110] Specifically, when training the MobileNet network, the following steps are included:
[0111] 4.1) Dataset enhancement: First, obtain the original training data set, which can be retrieved from the database. Then, based on the original data set, rotate, flip, and normalize the pixels of the data set to increase the number of data sets and provide more sufficient data samples for training the network.
[0112] 4.2) Image preprocessing: Use an improved convolutional neural network to eliminate noise and improve the clarity of OCT image samples;
[0113] 4.3) Model construction: Improve the current MobileNetV2 model, combine the gated attention mechanism and multi-feature fusion mechanism, improve and optimize the convolutional layer, and finally form an OCT image classification model;
[0114] 4.4) Network training: Input the target features into the model to achieve multi-feature fusion and systematically learn feature information;
[0115] 4.5) Network testing: Input the OCT images in the test set into the OCT image classification model to verify the training accuracy; if the accuracy is not satisfactory, it is necessary to adjust the original data set to iteratively train the network until the classification accuracy meets the preset value, and then generate the final OCT image classification model.
[0116] Finally, when applying it, you only need to obtain the OCT image of the target object and input it into the OCT image classification model to quickly and accurately output the classification results.
[0117] In summary, the embodiment of the present invention provides an intelligent classification and processing method for OCT images based on deep learning. First, the OCT image of the target object is obtained, and the OCT image is subjected to denoising. In the denoising process, the improved convolutional network model is mainly used for unsupervised denoising. The model includes a sparse local attention module, a deep extraction module, a global attention module and a residual module. It can still achieve good denoising performance when the data set is insufficient, has strong generalization ability, and provides image samples with better quality for subsequent processing. When extracting the ROI area, the original ROI area is generated by pixel filling, global threshold processing, median filtering, polynomial fitting and image cutting, and the target ROI area is further obtained by bias field correction; when extracting features, by extracting local features of various types of features and using the PCA algorithm to achieve feature dimensionality reduction, the amount of calculation can be greatly reduced and the recognition efficiency can be improved. Finally, multiple features are fused, and OCT image classification is performed based on the improved MobileNet network, which can quickly and accurately complete the OCT image. Compared with manual experience classification, it has the advantages of high efficiency and high accuracy.
[0118] See also Figure 2 The present invention also provides an OCT image intelligent classification processing system based on deep learning, the system comprising:
[0119] The noise reduction unit 100 is used to obtain an OCT image of a target object and perform noise reduction processing on the OCT image;
[0120] The ROI region generating unit 200 is used to extract the original ROI region from the denoised OCT image, and perform data bias field correction processing on the original ROI region to obtain the target ROI region;
[0121] A target feature generating unit 300 is used to extract local features of each type of feature in the target ROI region, and use a PCA algorithm to perform feature dimension reduction on the local features of each extracted feature to generate target features;
[0122] The OCT image classification unit 400 is used to input the target features into a trained OCT image classification model to generate a classification result.
[0123] It can be understood that the functions or modules included in the system provided in this embodiment can be used to execute the method described in the above method embodiment. Its specific implementation can refer to the description of the above method embodiment. For the sake of brevity, it will not be repeated here.
[0124] The present invention also provides an electronic device, comprising: a processor, a sending device, an input device, an output device and a memory, wherein the memory is used to store computer program code, and the computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method as described in any possible implementation manner.
[0125] The present invention also provides a computer-readable storage medium, in which a computer program is stored. The computer program includes program instructions. When the program instructions are executed by a processor of an electronic device, the processor executes a method as described in any possible implementation manner.
[0126] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0127] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. Those skilled in the art can also clearly understand that the descriptions of the various embodiments of the present invention have different focuses. For the convenience and brevity of description, the same or similar parts may not be repeated in different embodiments. Therefore, for parts not described or not described in detail in a certain embodiment, refer to the records of other embodiments.
[0128] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.
[0129] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0130] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
[0131] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a digital versatile disc (DVD)), or a semiconductor medium (eg, a solid state disk (SSD)).
[0132] A person skilled in the art can understand that to implement all or part of the processes in the above-mentioned embodiments, the processes can be completed by a computer program to instruct the relevant hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned method embodiments. The aforementioned storage medium includes: a read-only memory (ROM) or a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.
Claims
1. A method for intelligent classification and processing of OCT images based on deep learning, characterized in that: The method comprises: Acquire an OCT image of the target object, and perform noise reduction processing on the OCT image; Extracting an original ROI region from the denoised OCT image, performing data bias field correction processing on the original ROI region, and obtaining a target ROI region; Extract local features of each type of feature in the target ROI area, and use the PCA algorithm to reduce the feature dimension of the local features of each extracted feature to generate target features; The target features are input into a trained OCT image classification model to generate a classification result.
2. The OCT image intelligent classification processing method based on deep learning according to claim 1 is characterized in that: The OCT image is subjected to noise reduction processing, comprising: Denoising the OCT image using an improved convolutional network model, wherein the improved convolutional network model includes a sparse local attention module, a depth extraction module, a global attention module, and a residual module; The sparse local attention module includes a hole convolution layer, a batch normalization layer, a ReLU activation layer and a convolution layer; The depth extraction module adopts a three-layer convolution extraction network, each layer of which includes a batch normalization layer, a ReLU activation layer and a convolution layer; The global attention module uses a convolution kernel to compress the features extracted by the sparse local attention module and the deep extraction module to generate a weight vector; the weight vector is used to adjust the network weights of the sparse local attention module and the deep extraction module.
3. The OCT image intelligent classification processing method based on deep learning according to claim 2 is characterized in that: After the OCT image is denoised using the improved convolutional network model, the method further includes: The noise reduction effect is evaluated using evaluation indicators, which are: In the formula, f PSNR represents the peak signal-to-noise ratio, max(I) represents the maximum grayscale value of the pixel in image I, and f MSE is the mean square error between the denoised image and the true image; W and H represent the width and height of the image respectively, and I d (x, y) represents the grayscale value of the denoised image, I c (x,y) represents the gray value of the true image, and (x,y) represents the coordinates of the image pixel.
4. The OCT image intelligent classification processing method based on deep learning according to claim 1 is characterized in that: The extracting the original ROI area from the denoised OCT image comprises: Based on the denoised OCT image, the blank edges are removed and the blank areas of the image are filled using the average value of the background pixels in the OCT image; Performing global threshold processing on the filled image to generate the first retinal structure image; Perform median filtering on the threshold processed image to remove black spots in the retina and generate a second retinal structure image; Performing quadratic polynomial fitting on the second retinal structure image to generate a fitting curve; The images of different retinal structures are aligned, and after the alignment, each column of pixels is translated upward or downward by a number of pixels so that the pixel points on the fitting curve are located on the same horizontal line, and then the original ROI area is cut out.
5. The OCT image intelligent classification processing method based on deep learning according to claim 1 is characterized in that: Performing local feature extraction on each type of feature in the target ROI region includes: Divide the target ROI area into a plurality of rectangular blocks, and calculate a quantization coefficient of each pixel in each of the rectangular blocks; Generating a histogram of quantization coefficients on each rectangular block, merging all the histograms to obtain a local feature of each feature; Wherein, the expression of the local feature is: Where (x, y) is the pixel coordinate, W u represents the basis vectors of the two-dimensional discrete Fourier transform with frequency u, Represents the basis vector W u The transpose of , j represents the jth component, f x N x Medium 2 The vector composed of the grayscale values of pixels, N x is a M×M rectangular neighborhood.
6. The OCT image intelligent classification processing method based on deep learning according to claim 1 is characterized in that: The OCT image classification model is obtained by training the MobileNet network; The MobileNet network includes: A multi-feature fusion module is used to perform multi-feature fusion on the target features to generate a fusion feature map; the multi-feature fusion includes channel layer feature fusion and space layer feature fusion; A channel attention module is used to average pool the fused feature map, convolve the compressed image after average pooling twice, activate it through a Sigmoid function, multiply the activated feature map with the input fused feature map, and obtain a channel output feature map; The spatial attention module is used to perform maximum pooling and average pooling on the channel output feature map in sequence, input the processed feature map into a convolution layer with a convolution kernel of 8×8, and multiply it with the channel output feature map after activation by a Sigmoid function to generate a final classification feature map.
7. The OCT image intelligent classification processing method based on deep learning according to claim 1 is characterized in that: The expression of the target feature is: In the formula, F add is the target feature, K i represents the convolution kernel, * represents convolution, D i Retinal flatness, Z i V is the degree of depression of the center of the macula, i is the width of the retinal nerve fiber, i = 1, 2, 3…n.
8. An OCT image intelligent classification processing system based on deep learning, characterized in that: The system comprises: A noise reduction unit, used for acquiring an OCT image of a target object and performing noise reduction processing on the OCT image; The ROI region generating unit is used to extract the original ROI region from the noise-reduced OCT image, and perform data bias field correction processing on the original ROI region to obtain the target ROI region; A target feature generation unit is used to extract local features of each type of feature in the target ROI area, and use a PCA algorithm to perform feature dimension reduction on the local features of each extracted feature to generate target features; The OCT image classification unit is used to input the target features into a trained OCT image classification model to generate a classification result.
9. An electronic device, characterized in that: include: Processor and memory; The memory is used to store computer program code, which includes computer instructions; when the processor executes the computer instructions, the electronic device executes the deep learning-based OCT image intelligent classification processing method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: A computer program is stored in the computer-readable storage medium; the computer program includes program instructions, and when the program instructions are executed by a processor of an electronic device, the processor executes the deep learning-based OCT image intelligent classification processing method as described in any one of claims 1 to 7.
Citation Information
Patent Citations
Visual representation learning for brain tumor classification
CN107851194A
Fundus lesion data processing method and device, equipment and storage medium
CN117831749A
Retina OCT disease image classification method based on improved neural network
CN118038182A
Road sheet
KR102561191B1
Supervised machine learning based multi-task artificial intelligence classification of retinopathies
US20220151568A1
Cited By
AI-driving-based auxiliary diagnosis method, device, equipment, medium and product for gynecological 3D detection mirror
CN120356649A
Computer program product, nasal septum deflection identification device and equipment
CN120411647A
Computer program product, device and apparatus for deviated nasal septum recognition
CN120411647B
