Color eye fundus image quality evaluation method based on ensemble learning
By employing an ensemble learning approach that combines ResNet-18 and U-Net models, this study addresses the issues of low accuracy and misjudgment in existing color fundus image quality assessment methods. It enables real-time and efficient quality assessment of color fundus images, supporting automated ophthalmic diagnosis and on-site image quality judgment.
Patent Information
- Application Number
- CN202410944910.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-07-15
AI Technical Summary
Existing methods for evaluating the quality of color fundus images suffer from low accuracy and misjudgment when considering both traditional image features and pathological structural features. They are particularly difficult to achieve high accuracy on large-scale datasets, and existing methods ignore the special characteristics of fundus images.
An ensemble learning-based approach, combining ResNet-18 and U-Net models, is used to evaluate the quality of color fundus images through manually labeled features and automatic segmentation. Taking into account the general and structural parameters of the images, support vector machines, Naive Bayes classifiers, and AdaBoost classifiers are used for feature classification. The model performance is optimized by loss functions and Dice coefficients, and finally, image quality is assessed by setting a threshold.
It enables real-time and efficient quality assessment of color fundus images, allowing for on-site determination of image suitability for ophthalmic diagnosis, reducing repetitive imaging processes, improving diagnostic efficiency, alleviating the burden on examiners and patients, and providing reliable automated diagnostic support.
Smart Images

Figure CN118968266A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of electronic information technology, and in particular to a color fundus image quality evaluation method based on ensemble learning. Background Art
[0002] In the field of ophthalmology, digital fundus photography is being used for early and improved diagnosis of various ocular diseases. Fundus image diagnosis is performed by examining morphological changes in the optic disc, macula, and blood vessels. Advances in optics, computer sensors, and image processing have led to the invention of sophisticated imaging devices, such as the smartphone-based retinal imaging system D-EYE and the panoptic ophthalmoscope, which allow fundus images to be captured, stored, and shared in an affordable way with minimal effort. However, such ophthalmic imaging setups are more susceptible to quality impairment, which can lead to incorrect diagnoses. In addition, with the advent of digital imaging and computer-aided diagnosis (CAD) systems for various ocular diseases, the effects of the worsening doctor-to-patient ratio can also be addressed. CAD systems are very effective in reducing ophthalmologists' diagnostic oversights and false-negative rates. The indicative significance of fundus images depends on the ophthalmologist's perception of their visual quality. Whether it is manual or mechanical, the quality of fundus images must be guaranteed in order to ensure a reliable conclusion. However, subjective quality evaluation of fundus images is a rather tedious and hectic process while processing a large amount of patient data. In order to mimic the quality predictions of medical experts, it is necessary to conduct objective quality evaluation of fundus images.
[0003] As a non-invasive examination of the eye, CF imaging is considered an effective way to screen and diagnose several retina-related diseases, such as diabetic retinopathy, hard exudate detection, anemia detection, age-related macular degeneration, etc., so it is very important to study objective IQA algorithms for CF images. Unlike traditional IQA tasks, CF image quality also depends on the type of diagnosis and the visualization of key retinal structures. For example, an image with dark areas may be considered good for glaucoma detection but poor for diabetic retinopathy detection. Existing color fundus IQA methods are only a classification task, which classifies CF images into two or three categories for medical diagnosis. For example, "good" and "bad" or "good", "usable" and "bad". Based on the extracted features, existing methods can be divided into two parts: traditional image quality features such as illumination, clarity and contrast, and pathological features such as identification of pathological structures, field of view definition, and visibility of optic disc and macula. If all the required pathological structures are clearly visible, the quality of the medical image can be considered appropriate. The main structures visible in the fundus images of healthy eyes are optic disc, macula and blood vessels. The main factors behind the degradation of retinal image quality are: haze events, dust and dirt in the camera, dust in the camera lens, blinking and eyelash occlusion. Common distortions caused by the above factors include uneven illumination, low contrast, blur and light spots. Artifacts caused by dust and dirt are clearly visible light spots scattered throughout the image. Full blink events result in extremely poor fundus image quality. Improper gamma and flash adjustments can cause color distortions such as overexposure, underexposure and uneven lighting. In addition, there are other less common distortions caused by a few events. For example, improper cleaning of the camera lens can produce a circular halo in the center of the image.
[0004] According to the features extracted from CF images, they are divided into traditional image features and pathological features. The quality evaluation index of CF images should also include these two parts. In the past two decades, many studies on color fundus IQA have been published. Through these studies, it can be found that it is feasible to use traditional image features alone to objectively evaluate the quality of CF images, which is similar to the quality evaluation of natural images. This method can grade the quality of CF images to a certain extent, but it will not achieve a high accuracy on large-scale data sets because such methods ignore the particularity of medical images. Using the appearance of pathological structures in images as IQA indicators alone has a premise that poor display of pathological structures is caused by poor image quality. Although the final experimental results can prove that this common sense assumption is suitable for color fundus IQA. However, some tricky problems will be encountered, such as when the targeted pathological structure happens to be the location of the lesion, which may result in obvious misjudgment. Raj A et al. took into account both traditional image features and pathological structure features in the article “Multivariate Regression-Based Convolutional Neural Network Model for Fundus Image Quality Assessment” and established a fundus IQA model based on multivariate regression. The data set required for training this model requires careful scoring by professional doctors, which is a huge project.
[0005] Therefore, a color fundus image quality assessment method based on ensemble learning is proposed. Summary of the invention
[0006] The purpose of the present invention is to provide a color fundus image quality evaluation method based on ensemble learning in order to solve the above problems, aiming at real-time and efficient quality evaluation of CF images, that is, the proposed method can be used as a real-time automated CAD means in practical applications. The method can take into account the general parameters and structural parameters of CF images, conduct a comprehensive quality evaluation of CF images, and provide reliable fundus images for professional doctors and equipment.
[0007] The present invention achieves the above-mentioned purpose through the following technical scheme, a method for evaluating the quality of color fundus images based on ensemble learning, comprising the following steps:
[0008] S1. Dataset division. The dataset includes a color fundus IQA dataset and a segmentation dataset. 80% of the data in each dataset is randomly selected as the training set, and the remaining 20% of the data is the validation set. The ResNet-18 model and the U-Net model are trained separately without interfering with each other.
[0009] S2. Manually annotate 50% (adjust the proportion according to the amount of data) of the images in the training set of the color fundus IQA dataset with four features: contrast, focus, illumination, shadow, and reflection.
[0010] S3. For each feature in step S2, use manual annotation to train support vector machines, naive Bayes classifiers, classification trees, and AdaBoost classifiers. Then select the model with the best validation performance as the classifier for the feature. Use the 4 feature quality scores predicted for each CF image (in addition, there is a color fundus IQA dataset with binary annotations. 0 represents good quality and 1 represents poor quality, so the fine-tuned ResNet-18 finally gets a fully connected layer consisting of 5 output neurons) as input to train the ResNet-18 model, calculate the binary cross entropy loss for each output feature, and use the average value to merge them into the final overall loss. Update the model gradient through the loss function. The loss function expression is:
[0011]
[0012] In the formula, N=5, y n Represents the nth value of the model output, x n represents the corresponding true value;
[0013] S4. The quality classification performance of the ResNet-18 model was verified by the validation set of the color fundus IQA dataset. The quality classification performance of the ResNet-18 model was evaluated using accuracy, sensitivity, specificity, and F-measure. In order to describe these evaluation indicators, true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) need to be calculated. TP and TN represent the number of correctly matched samples between the predicted label and the actual label. FP and FN represent the number of oppositely matched samples between the predicted and actual labels. Therefore, the calculation process of the evaluation indicators is as follows:
[0014]
[0015]
[0016] S5. Use the training set of the CF image segmentation dataset to train U-Net for automatic blood vessel segmentation. The training part only requires the image to be adjusted to a fixed size, and no additional image normalization is required to standardize the final FD value. This step only trains the blood vessel segmentation capability of U-Net;
[0017] S6. Verify the U-Net segmentation performance in the validation set of the CF image segmentation dataset, and use the Dice coefficient to evaluate the blood vessel segmentation performance of the U-Net model. The formula for calculating the Dice coefficient is as follows: Dice = 2*(|A∩B|) / (|A|+|B|)
[0018] In the formula, |A| represents the number of pixels of A, |B| represents the number of pixels of B, and |A∩B| represents the number of pixels of the intersection of A and B. Here, A represents the segmentation result predicted by the model, and B represents the true label. The closer the Dice coefficient value is to 1, the better the performance of the U-Net model.
[0019] S7. Integrate the trained ResNet-18 model and U-Net model, and verify the performance of the integrated model on the DRIMDB dataset (a widely used benchmark fundus image dataset). Here, the original color images of the DRIMDB dataset need to be preprocessed (the specific steps are detailed in the application features section) and the FD value is standardized. For the final comprehensive score, different thresholds need to be debugged to divide "good" and "bad", and the threshold with the highest accuracy is retained.
[0020] The beneficial effects of the present invention are:
[0021] The CF image quality evaluation method based on ensemble learning proposed in this application meets the actual needs of ophthalmology clinics and can comprehensively evaluate CF images taken on site to determine whether the image is suitable for ophthalmologists to diagnose. In the case of poor quality, it allows for immediate adjustment or repetition of the imaging process, ideally while the patient is still on site. This saves time, reduces the burden on examiners and patients, and avoids unnecessary appointments. The method takes into account the traditional image features and pathological structural features of CF images, obtains visual cues from segmented vascular images, and analyzes structural degradation, which effectively represents the method for doctors to determine the quality of fundus images, making up for the shortcomings of traditional image feature analysis alone. The participation of traditional image feature scoring limits the impact of the limitations of vascular segmentation methods. Properly setting up and integrating this method into the workflow can enable efficient and automatic color fundus IQA before diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 This is a flow chart of the efficient IQA module training of the present invention;
[0023] Figure 2 This is a flowchart for automatic evaluation of CF radiographic images in actual medical treatment. DETAILED DESCRIPTION
[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.
[0025] A color fundus image quality evaluation method based on ensemble learning comprises the following steps:
[0026] S1. Dataset division. The dataset includes a color fundus IQA dataset and a segmentation dataset. 80% of the data in each dataset is randomly selected as the training set, and the remaining 20% of the data is the validation set. The ResNet-18 model and the U-Net model are trained separately without interfering with each other.
[0027] S2. Manually annotate 50% (adjust the proportion according to the amount of data) of the images in the training set of the color fundus IQA dataset with four features: contrast, focus, illumination, shadow, and reflection.
[0028] S3. For each feature in step S2, use manual annotation to train support vector machines, naive Bayes classifiers, classification trees, and AdaBoost classifiers. Then select the model with the best validation performance as the classifier for the feature. Use the 4 feature quality scores predicted for each CF image (in addition, there is a color fundus IQA dataset with binary annotations. 0 represents good quality and 1 represents poor quality, so the fine-tuned ResNet-18 finally gets a fully connected layer consisting of 5 output neurons) as input to train the ResNet-18 model, calculate the binary cross entropy loss for each output feature, and use the average value to merge them into the final overall loss. Update the model gradient through the loss function. The loss function expression is:
[0029]
[0030] In the formula, N=5, y n Represents the nth value of the model output, x n represents the corresponding true value;
[0031] S4. The quality classification performance of the ResNet-18 model was verified by the validation set of the color fundus IQA dataset. The quality classification performance of the ResNet-18 model was evaluated using accuracy, sensitivity, specificity, and F-measure. In order to describe these evaluation indicators, true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) need to be calculated. TP and TN represent the number of correctly matched samples between the predicted label and the actual label. FP and FN represent the number of oppositely matched samples between the predicted and actual labels. Therefore, the calculation process of the evaluation indicators is as follows:
[0032]
[0033]
[0034] S5. Use the training set of the CF image segmentation dataset to train U-Net for automatic blood vessel segmentation. The training part only requires the image to be adjusted to a fixed size, and no additional image normalization is required to standardize the final FD value. This step only trains the blood vessel segmentation capability of U-Net;
[0035] S6. Verify the U-Net segmentation performance in the validation set of the CF image segmentation dataset, and use the Dice coefficient to evaluate the blood vessel segmentation performance of the U-Net model. The formula for calculating the Dice coefficient is as follows: Dice = 2*(|A∩B|) / (|A|+|B|)
[0036] In the formula, |A| represents the number of pixels of A, |B| represents the number of pixels of B, and |A∩B| represents the number of pixels of the intersection of A and B. Here, A represents the segmentation result predicted by the model, and B represents the true label. The closer the Dice coefficient value is to 1, the better the performance of the U-Net model.
[0037] S7. Integrate the trained ResNet-18 model and U-Net model, and verify the performance of the integrated model on the DRIMDB dataset (a widely used benchmark fundus image dataset). Here, the original color images of the DRIMDB dataset need to be preprocessed (the specific steps are detailed in the application features section) and the FD value is standardized. For the final comprehensive score, different thresholds need to be debugged to divide "good" and "bad", and the threshold with the highest accuracy is retained.
[0038] The details are as follows:
[0039] 1. Selection and ensemble learning of common quality parameters
[0040] The contrast, focus, lighting, shadows, and reflections of CF images are important indicators for evaluating image quality and applicability. These factors can directly affect doctors' diagnosis and evaluation of fundus lesions. These four universal quality features are selected to evaluate the quality of CF images. For each universal quality feature, the CF image is input into multiple classifiers, and the classifier with the best classification result is selected as the quality classifier of the feature, and finally the four universal feature quality scores of each CF image are obtained. Finally, the four universal quality feature scores are used to train the final quality classifier.
[0041] 2. ResNet-18 Architecture
[0042] The final quality classifier follows the traditional ResNet-18 architecture, which consists of an initial convolution followed by 4 ResNet layers, a global average pooling and a fully connected layer, followed by a sigmoid function to form the model output. The initial convolution consists of a convolution layer, a batch normalization layer, a rectified linear unit (ReLu), and a max pooling operation. The ResNet layer consists of two convolution blocks, each of which consists of a convolution and a batch normalization layer, with an intermediate ReLu between the two blocks. In this application, the fine-tuned model ends up with a fully connected layer consisting of 5 output neurons. Each neuron provides a quality score between 0 and 1 for contrast, focus, lighting, shadows and reflections, and an overall quality score.
[0043] 3. Uncertainty Scoring
[0044] A dropout layer is added after each ResNet layer in the ResNet-18 architecture. The dropout layer remains activated to evaluate the input, and 16 inferences are performed on a single sample. The average of the predicted probabilities is used as the final probability score, and the variance is used as the corresponding uncertainty estimate.
[0045] 4. U-Net Architecture
[0046] The architecture of U-Net has an encoder-decoder structure, which can effectively capture the feature information in the image and perform image segmentation while retaining the spatial information. When processing the task of blood vessel segmentation in fundus images, U-Net can learn the shape characteristics of the blood vessels and output pixel-level blood vessel segmentation results. The present invention obtains 54 images on the annotated dataset RETA, enhances the 800 images to train the segmentation model, and verifies the blood vessel segmentation effect on three public datasets DRIVE, ARIA and STARE, to ensure that the blood vessel segmentation results will not have a negative effect on the final quality evaluation.
[0047] 5. Box Counting Method
[0048] The measurement of the vascular fractal dimension (FD) uses the box counting method to cover a box network of appropriate grid size with the segmented image and calculate the number of boxes required to cover all vascular pixels. The FD value is then calculated using the grid size and the number of grids covered. The smaller the grid size, the more retinal vascular details can be covered. The larger the FD value, the better the vascular visibility. Any image quality problem will reduce the visibility of the blood vessels, resulting in a smaller FD value.
[0049] 6. Comprehensive score
[0050] How to measure the contribution of the universal quality parameter score and FD value of the CF image to the final image quality score is very important. The uncertainty score is used to determine the impact of the two on the final score. The lower the uncertainty score, the higher the proportion of the image universal quality score, and the higher the uncertainty, the higher the proportion of the FD value.
[0051] 7. Threshold setting
[0052] In order to adapt to most current research work, it is necessary to set a reasonable threshold and classify images into "good" and "bad". In fact, it is more appropriate to leave this part of the work to professional doctors to flexibly choose in actual operation. Considering the integrity of the invention, this application will still give a recommended threshold. Here, if the CF image quality score is selected in percentage, 70 points is usually a reasonable threshold.
[0053] A study from the University of Wisconsin-Madison defined the following quality indicators for fundus images: focus and clarity, field clarity, visibility of the macula, optical disc, and blood vessels. These six indicators fully consider the generic and structural indicators of fundus images, but they are not the only criteria. There are many different quality parameters for evaluating fundus image quality, such as lighting, blur, color, contrast, etc. This application identifies contrast, focus, lighting, shadows, and reflections as generic quality indicators for fundus IQA, and these four indicators can better reflect the quality of CF images. Using the feature learning method proposed by Dias et al. in the article "Retinal image quality assessment using generic image quality indicators", quality assessment is performed using network ensembles. For each feature, custom manual features are extracted from the input image and used to train support vector machines, naive Bayes classifiers, classification trees, and AdaBoost classifiers. The model with the best validation performance is then selected as the classifier for the feature. In addition, the predicted quality scores of all features are used as input to train the last quality classifier. The use of machine learning and ensemble learning here greatly reduces the cost of manual annotation.
[0054] In this application, the uncertainty score determines the degree of influence of vascular FD on the quality score of a CF image. For image features between 0 and 1 output by the ResNet-18 architecture, there will be a high uncertainty. After trying to eliminate these uncertain samples, the accuracy of the model on the remaining data has been significantly improved. This shows that it is difficult for the model to accurately score such images. Obviously, it is not advisable to pre-eliminate images with high uncertainty in real work, and such images with high uncertainty cannot be simply classified as "abnormal". In order to solve this problem, a segmentation model is introduced to measure such images with high uncertainty with structural features, and the uncertainty affects the final quality score of the CF image. The specific calculation formula for the final comprehensive score is:
[0055] S=100(S ResNet (1-αN)+αFN) / (1+F max )
[0056] Where S ResNet represents the total quality score of ResNet-18 output, F represents the FD value, N is the uncertainty score, α=1.2, F max Indicates the maximum value of FD.
[0057] The formula retains the impact of the output scores of the ResNet-18 architecture on the final quality score of such images with high uncertainty, avoiding the limitation of evaluating CF images solely based on vascular FD. For example, proliferative diabetic retinopathy (PDR) can cause eye diseases with decreased vascular density, which will cause the vascular segmentation-based method to fail and result in misjudgment in the PDR disease dataset. CF images with general parameter quality scores close to 0 and 1 have a small uncertainty score, resulting in little effect of the vascular FD value on the final quality score result. This decision-making method limits the limitations of the vascular segmentation-based method.
[0058] A commonly used method for FD estimation in Euclidean space is to count the dimensions. Overlay a uniformly spaced grid of boxes of size λ with the segmented CF image and count the number of boxes required to cover all container pixels. By reducing λ, more details of the retinal vasculature can be seen from the overlay. Let N(λ) be the number of boxes covered by the vessel when the box side length is λ, and the formula for calculating the vascular FD value is:
[0059]
[0060] The issue of normalization of retinal FD measurements is also considered here. Fundus cameras can usually see a retinal area of 30°-60° (also called FOV). Fundus images taken by different cameras and operators may vary in image resolution, size, field of view shape and size. All of these situations will result in the calculated FD value being non-reference. Solving this problem requires image normalization before segmentation, removing black pixels outside the field of view area, and adjusting it to a fixed size. The specific process is to extract a binary field of view mask from the original color image, and then locate a square bounding box. The FOV mask corresponds to the circular area embedded in the box. Any missing circle segments will be filled with black pixels. The original color image is cropped using the upper left and lower right coordinates of the bounding box, and finally the cropped color image is adjusted to a fixed size.
[0061] To achieve automatic blood vessel segmentation, we set CF is a two-dimensional CF image, I S is the corresponding blood vessel image, which is manually annotated. Each pixel in the image has a Euclidean coordinate (x, y), x≤M, y≤N, where M and N are the height and width of the image respectively. The automatic segmentation of blood vessels is achieved using the segmentation module f(I CF ,W)→I S , f represents the segmentation module, and W represents the module weight set. The U-Net architecture is selected as the segmentation module. CF and I S Learning is performed between . f is iteratively updated by minimizing the binary cross entropy loss function L, which is defined as:
[0062]
[0063] The design of the comprehensive score has strong practical significance. Although the design of uncertainty is used to express the confidence of the model in the prediction results, it is found in the experiment that uncertainty is also related to the predicted image quality score. For predictions close to 0 (best quality) or 1 (worst quality), the uncertainty is low, while for predictions between the two, the uncertainty is high. Whether such images in between can be classified as "suitable" is difficult for the algorithm to decide. An experienced ophthalmologist may not always need high-quality images and can also effectively diagnose from relatively low-quality CF images. In the case of CAD, images of general quality may be classified as poor quality images, resulting in image recapture. Of course, this will lead to a waste of energy and time. Therefore, the FD value is used to correct the quality score of the general parameter, and the effect of the visualization of the pathological structure can largely determine whether such images of general quality are suitable for doctors' diagnosis.
[0064] For images with substandard quality, a common workflow is to retake another image that meets the IQA minimum image quality standard. However, frequent rejections can be discouraging if there is no convincing explanation for the low-quality decision. In addition, binary classification of image quality, such as "good" or "bad", does not seem to be the best choice for real-world IQA applications in ophthalmology clinics. It is better to use a continuous quality scale. Therefore, for applicability, IQA algorithms should be interpretable and grade image quality on a continuous scale. Second, subjective expectations should be different for CF images taken by different devices. For fundus images taken by handheld and smartphone fundus cameras, thresholds with smaller quality requirements must be considered because these images usually have a smaller field of view and image resolution. Therefore, using flexible threshold settings for quality classification meets practical needs. On the other hand, classification threshold settings are also necessary. It is understood that there is currently no public fundus IQA dataset with continuous scale grading, and there are only a few quality classification datasets, such as DRIMDB, DeepDR, and EyeQ datasets. Quality classification is necessary to train models using these datasets.
[0065] This application is a study on CF image quality evaluation methods based on ensemble learning, mainly to meet the actual needs of ophthalmic photography. It proposes a comprehensive and efficient CF image quality evaluation method, which can quickly and automatically evaluate the quality of CF images before diagnosis. Nowadays, the artificial intelligence-driven Automatic Retinal Image Analysis (ARIA) system can provide patients with ophthalmologist-like reports within a few seconds. The rapid and accurate response of the ARIA system requires high-quality CF images as a guarantee. The IOA module designed based on the method proposed in this application is integrated into the workflow to provide quality assurance for the images received by the ARIA system. In addition, the analysis of the structural degradation of CF images by the FD value in the method can effectively represent the method used by doctors to determine the quality of fundus images, further ensuring the diagnostic applicability of qualified images.
[0066] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A color fundus image quality evaluation method based on ensemble learning, characterized in that: The following steps are involved: S1. Dataset division; S2. Manually annotate 50% (adjust the proportion according to the amount of data) of the images in the training set of the color fundus IQA dataset with four features: contrast, focus, illumination, shadow, and reflection. S3, for each feature in step S2, train a support vector machine, a naive Bayes classifier, a classification tree, and an AdaBoost classifier using manual annotations, and then select the model with the best validation performance as the classifier for that feature; S4. The quality classification performance of the ResNet-18 model was partially verified by using the validation set of the color fundus IQA dataset; S5. Use the training set of the CF image segmentation dataset to train U-Net for automatic blood vessel segmentation. S6. Verify the U-Net segmentation performance on the validation set of the CF image segmentation dataset; S7. Integrate the trained ResNet-18 model and U-Net model, and verify the performance of the integrated model on the DRIMDB dataset (a widely used benchmark fundus image dataset).
2. The method for evaluating the quality of color fundus images based on ensemble learning according to claim 1, characterized in that: In step S1, the data set includes a color fundus IQA data set and a segmentation data set. 80% of the data in the data set are randomly selected as training sets, and the remaining 20% of the data are used as validation sets. The ResNet-18 model and the U-Net model are trained separately without interfering with each other.
3. The method for evaluating the quality of color fundus images based on ensemble learning according to claim 1, characterized in that: In step S3, the model with the best validation performance is selected as the classifier of the feature, and the four feature quality scores predicted for each CF image are used as input to train the ResNet-18 model (in addition, the color fundus IQA dataset is labeled with binary values, 0 represents good quality and 1 represents poor quality, so the fine-tuned ResNet-18 finally obtains a fully connected layer consisting of 5 output neurons), and the binary cross entropy loss of each output feature is calculated, and the average value is used to merge them into the final overall loss, and the model gradient is updated by the loss function, and the loss function expression is: In the formula, N=5, y n Represents the nth value of the model output, x n Represents the corresponding true value.
4. The method for evaluating color fundus image quality based on ensemble learning according to claim 1, characterized in that: In step S4, the quality classification performance of the ResNet-18 model is evaluated using accuracy, sensitivity, specificity and F-measure. In order to describe these evaluation indicators, true positive (TP), true negative (TN), false positive (FP) and false negative (FN) need to be calculated. TP and TN represent the number of correctly matched samples between the predicted label and the actual label, and FP and FN represent the number of oppositely matched samples between the predicted and actual labels. The calculation process of the evaluation indicators is as follows:
5. The method for evaluating the quality of color fundus images based on ensemble learning according to claim 1, characterized in that: In step S5, the training part only requires adjusting the image to a fixed size, and no additional image normalization is required to regulate the final FD value. This step only trains the blood vessel segmentation capability of U-Net.
6. The method for evaluating the quality of color fundus images based on ensemble learning according to claim 1, characterized in that: In step S6, the Dice coefficient is used to evaluate the blood vessel segmentation performance of the U-Net model. The formula for calculating the Dice coefficient is as follows: Dice=2*(|A∩B|) / (|A|+|B|) Among them, |A| represents the number of pixels of A, |B| represents the number of pixels of B, |A∩B| represents the number of pixels of the intersection of A and B, A represents the segmentation result predicted by the model, B represents the true label, and the closer the Dice coefficient value is to 1, the better the performance of the U-Net model.
7. The method for evaluating color fundus image quality based on ensemble learning according to claim 1, characterized in that: In step S7, the original color image of the DRIMDB dataset is preprocessed, the FD value is standardized, and different thresholds need to be debugged for the final comprehensive score to divide "good" and "bad", and the threshold with the highest accuracy is retained.
Citation Information
Patent Citations
Knowledge data and artificial intelligence driven ophthalmology multi-disease identification system
CN113962311A
Eye fundus image blood vessel segmentation method and system
CN115661185A
Retina OCTA image fundus segmentation method based on deep learning
CN117437685A
Semi-supervised fundus image quality assessment method using IR tracking
WO2022180227A1
