Glaucoma image classification method based on deep learning

Abstract

The invention discloses an improved glaucoma diagnosis method based on a deep learning Adam optimization algorithm, and belongs to the technical field of medical image processing. The method comprises the following steps: acquiring a glaucoma data set image; dividing the data set into a training set, a test set and a verification set, and performing preprocessing operation; inputting the preprocessed images into a neural network model for feature extraction, and classifying the images by adopting an improved algorithm based on an Adam algorithm; outputting an image classification result; and performing performance evaluation on the classification result. According to the embodiment of the invention, image classification can be carried out without special equipment and special personnel, and meanwhile, the accuracy and efficiency of glaucoma image recognition can be improved.

Description

Technical Field

[0001] The present invention relates to fundus glaucoma detection and treatment in the medical field, and in particular to a glaucoma image classification method improved based on a deep learning Adam optimization algorithm. Background Art

[0002] In the medical field, glaucoma is an eye disease that is often associated with increased intraocular pressure, which can damage the optic nerve and lead to vision loss. In order to better treat patients with fundus glaucoma, medical researchers have been looking for more effective treatments. At the same time, optimization algorithms play a vital role in the field of machine learning and data science.

[0003] Although the Adam optimization algorithm is a commonly used optimization algorithm, its performance still needs to be improved in the application of the medical field. The Adam optimization algorithm faces challenges in terms of convergence speed and stability when processing complex data. Especially in the case of high-dimensional data or uneven data distribution, the Adam algorithm may show slow convergence speed, resulting in a long and unstable optimization process.

[0004] The treatment of glaucoma usually relies on the doctor's clinical experience and the analysis of medical imaging data. However, traditional treatment methods have some limitations, including the problem of inaccurate diagnostic results. With the widespread application of machine learning technology in medical image analysis and disease diagnosis, existing optimization algorithms still face some challenges when processing medical data, such as slow convergence speed and insufficient generalization ability. The present invention proposes a new optimization method to alleviate the problems existing in the prior art. Summary of the invention

[0005] The present invention provides a glaucoma image classification method based on the improved deep learning Adam optimization algorithm, which can effectively alleviate the poor convergence problem of the Adam algorithm and apply it to the medical field for fundus glaucoma inspection and treatment. The algorithm improves the gradient and learning rate of the Adam algorithm, and the network used in the design is trained using the MobileNetV2 model.

[0006] The present invention proposes an improved glaucoma image classification method based on a deep learning Adam optimization algorithm. The optimization is mainly performed to address the existing limitation of the poor convergence of the Adam optimizer. The design adopts gradient control and linear interpolation to adjust the learning rate for improved optimization.

[0007] First, the difference between the first-order moment estimate and the second-order moment estimate is defined as an operator λ. The values ​​of the first-order moment estimate and the second-order moment estimate are related to the degree of change of the gradient, and the degree of change of the gradient reflects the curvature of the parameter space to some extent. This strategy can be regarded as a measure of the gradient curvature. During the gradient update process, the size of the gradient will be adjusted according to the current λ value, thereby controlling the rate of gradient update.

[0008] Another optimization strategy provided by the present invention uses linear interpolation to perform adaptive adjustment of the learning rate, which can ensure that historical information is fully utilized during the training process and dynamically adjust the size of the learning rate according to the current situation.

[0009] The present invention proposes an improved glaucoma image classification method based on a deep learning Adam optimization algorithm, comprising: obtaining images of patients with fundus glaucoma and patients without the disease; the images are classified into two categories, including: 1, fundus glaucoma present (GlaucomaPresent) and 0, fundus glaucoma not present (Glaucoma not Present) of two types of patients; Data preprocessing: All full-eye fundus glaucoma images were subjected to data preprocessing operations, and the processed images were obtained by rotating 15°, center cropping, magnifying 1.25 times and regularizing; The processed images are input into the MobileNetV2 model for training, and the classified accuracy and loss value results are output.

[0010] The present invention has the following main advantages: 1. By adjusting the change in the size of λ, the parameters are updated more effectively and the possibility of converging to a local optimal solution is avoided. At the same time, using the minimum value to adjust the size of the gradient can more effectively control the oscillation and volatility of the gradient, thereby converging to the optimal solution faster; 2. By gradually fine-tuning the learning rate and using linear interpolation to find a more accurate value, this adaptive adjustment can effectively avoid performance degradation caused by too large or too small a learning rate, thereby improving the stability and efficiency of the optimization algorithm during training. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the specific implementation of the present invention, the following will briefly introduce the drawings required for use in the specific implementation. The drawings described below are some implementations of the present application: Figure 1 It is a flowchart of an improved glaucoma diagnosis method based on a deep learning Adam optimization algorithm provided by an embodiment of the present invention; Figure 2 is a diagram of a glaucoma image preprocessing process provided by an embodiment of the present invention; Figure 3 It is an accuracy image outputted after image classification by a neural network model according to an embodiment of the present invention; Figure 4 It is a loss value image output after image classification by a neural network model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0012] In order to solve the existing technical problems, the present invention provides an improved glaucoma image classification method based on deep learning Adam optimization algorithm for use in the medical field to detect whether a patient suffers from glaucoma. The following first introduces the improved glaucoma diagnosis method based on deep learning Adam optimization algorithm provided by the present invention, and the present invention includes the following steps: In one embodiment, if Figure 1 As shown, a glaucoma image classification method improved based on deep learning Adam optimization algorithm, the method comprises the following steps: S1. Obtain glaucoma dataset images; In one embodiment, after acquiring the glaucoma dataset image, the method further includes: S2, divide the data set into training set, test set and validation set, and perform preprocessing operations on the images; Figure 2 is a glaucoma image preprocessing process provided by one embodiment of the present invention, such as Figure 2 As shown, the glaucoma image is processed by rotating 15°, center cropping, magnifying 1.25 times, and regularizing; S3, input the preprocessed image into the MobileNetV2 neural network model for feature extraction, and classify the image using an algorithm improved based on the Adam algorithm; S4, output classification results; The accuracy of the classification results obtained after training with the neural network model MobileNetV2 is as follows Figure 3 As shown; The loss value of the classification result obtained after training the neural network model MobileNetV2 is as follows Figure 4 As shown; S5. Evaluate the output results, including classification accuracy and loss value, to assist clinicians in making accurate diagnoses.

[0013] Two optimization strategies of the present invention will be described in detail below. In order to make the purpose, optimization strategy and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific optimization strategies and specific implementation examples. It should be understood that the specific strategies described herein are only intended to explain the present invention. The following display of the embodiments is only to provide the advantages and good performance of the present invention, and cannot be understood as a limitation of the present invention.

[0014] The difference between the second-order moment estimate and the square of the first-order moment estimate is calculated and defined as an operator λ, which indirectly reflects the curvature information of the gradient in the parameter space. By controlling and adjusting this difference, the optimization algorithm can better adjust the size and direction of the gradient, thereby updating the parameters more effectively and avoiding the possibility of converging to a local optimal solution. The operator is expressed by the following formula:

[0015] In order to control the size of the operator, a correction factor (GR) is defined to limit it to a range (the range size is set to 1.0). Using the minimum function to adjust the size of the gradient can more effectively control the oscillation and volatility of the gradient. The purpose of this is to prevent the gradient from being too large or too small and maintain the balance of parameter updates. If the operator size exceeds this range, it is adjusted to 1.0; otherwise, the original value is retained. The correction factor is expressed by the following formula: GR=min(|Operator|,1.0) This smooth control of the gradient helps improve the stability of the optimization process, reduces oscillations during training, and thus converges to the optimal solution faster.

[0016] If the difference value is large, it means that the gradient changes quickly. At this time, the gradient value is 1.0, making the gradient update more gentle. On the contrary, if the difference value is small, it means that the gradient changes slowly. The value of the operator at this time is taken to accelerate the gradient update. Then the gradient tensor is multiplied by the correction coefficient GR and normalized to a unit vector. The formula is as follows: grad1=grad·GR Doing so keeps the gradient direction unchanged, but scales its length to 1, which helps control the magnitude of the gradient update, making it more stable and easier to adjust. This process can be seen as adjusting the gradient change to ensure that the appropriate gradient size is maintained when the parameters are updated; Among them, grad is the gradient before adjustment. After multiplying the gradient tensor grad with the correction coefficient GR, the gradient tensor before adjustment is overwritten to obtain grad1. Then grad1 is divided by its norm (i.e., modulus) and the gradient tensor is updated to obtain grad2.

[0017] A second optimization strategy is added to the above optimization method: the linear interpolation strategy. This strategy can better control the optimization process, making the adjustment of the learning rate smoother to avoid drastic fluctuations in gradient updates, while ensuring that historical information is fully utilized during the training process and dynamically adjusting the learning rate according to the current situation. This adaptive adjustment can effectively avoid performance degradation caused by excessive or too small learning rates, thereby improving the stability and efficiency of the optimization algorithm during training.

[0018] First, the global step number is used to take the remainder of the interpolation period to obtain the position t of the current step number relative to a period. The change of t causes the value of the learning rate to change at each step, achieving the effect of dynamically adjusting the learning rate. The formula is as follows: t=step%period Then use the linear interpolation method to get the value of the learning rate at this time according to the relative position of the current training step in the cycle. The learning rate formula at this time is as follows: lr=max_lr-(max_lr-min_lr)·(t / period) Finally, take a learning rate value between the maximum learning rate and the minimum learning rate to get the value of the dynamically adjusted learning rate. The formula is as follows: 1. new =max(min_lr,lr-max_lr_change) The purpose of this interpolation is to dynamically adjust the learning rate during training so that the model can better converge or adapt to the training data. As training progresses, t will change with the change of step, so that the learning rate will also change accordingly. At the end of each cycle, t will restart from 0. This can be used to implement a periodic learning rate adjustment strategy, that is, in each cycle, the learning rate changes in the same way. This adjustment method helps to improve the performance and stability of the optimization algorithm.

[0019] The specific improved algorithm process is shown in Algorithm 1: Improved algorithm process:

[0020] In order to better compare the performance of the improved algorithm with other optimization algorithms in different aspects, we conducted comparative experiments using the improved algorithm and other classic optimization algorithms, such as SGD, Adam, Adagrad, NAdam, etc. This allows for a clearer evaluation of the advantages and disadvantages of the improved algorithm relative to other algorithms.

[0021] The experimental results of the improved algorithm on the glaucoma dataset are shown in Table 1

[0022] The results show that the method and system of the present invention can effectively improve the accuracy of glaucoma image classification and have important practical application value for the automated analysis and diagnosis of medical images.

[0023] The above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. An improved glaucoma image classification method based on deep learning Adam optimization algorithm, characterized in that: include: S1. Obtain glaucoma dataset images; S2. Divide the data set into training set, test set and validation set and perform preprocessing operations; S3. Input the preprocessed image into the neural network model for feature extraction, and classify the image using an improved algorithm based on the Adam algorithm; S4. Output image classification results; S5. Perform performance evaluation on the classification results.

2. According to claim 1, a glaucoma image classification method improved based on deep learning Adam optimization algorithm is characterized in that: After acquiring the glaucoma dataset images, it also includes: After dividing the dataset into training set, test set and validation set, the glaucoma images were rotated 15°, center cropped, enlarged 1.25 times and regularized preprocessing operations were performed.

3. The improved glaucoma image classification method based on deep learning Adam optimization algorithm according to claim 1 is characterized in that: In step S3, the MobileNetV2 neural network model is used to extract features from the image. An improved algorithm based on the Adam algorithm is combined with gradient control and linear interpolation strategies to improve the generalization ability and training stability of the model. The deep learning framework Pytorch is used in conjunction with the improved algorithm for effective image classification.

4. The improved glaucoma image classification method based on deep learning Adam optimization algorithm according to claim 1, characterized in that: Step S4 displays the output classification results and provides necessary information.

5. The improved glaucoma image classification method based on deep learning Adam optimization algorithm according to claim 1 is characterized in that: The evaluation results in step S5 include classification accuracy and loss value to assist clinicians in making accurate diagnoses.

Images