Evaluation method of diabetic retinopathy deep learning classification model

Abstract

The invention provides a diabetic retinopathy deep learning classification model evaluation method, and the method comprises the steps: firstly, describing a set of human eye model and fundus camera combined optical system; according to the optical system, a human eye model and a fundus camera model are combined, and diabetic retinopathy fundus examination simulation with high reduction degree and high resolution is achieved. In addition, a diabetic retinopathy simulation parallax image generation method based on computational optics is provided based on the method, a parallax simulation system is designed through ZEMAX, and a simulation parallax image is calculated through a point spread function matrix. Based on the generated simulated parallax image, a further deep learning classification model generalization performance test is provided, and the evaluation method can effectively evaluate the classification performance of the classification model on the target simulated parallax image.

Description

Technical Field

[0001] The present invention relates to the evaluation of medical-assisted deep learning classification models, and more specifically, to an evaluation method for a deep learning classification model of diabetic retinopathy, and to a method for generating a simulated parallax image of a target diabetic retinopathy by combining an optical system of a human eye model and a fundus camera. The simulated parallax image generated by the present invention can be used to evaluate the performance of a deep learning classification model of diabetic retinopathy in an environment with a parallax image. Background Art

[0002] At present, with the development of deep learning models, medical auxiliary diagnosis models based on deep learning are playing an increasingly important role. However, in the field of medical metrology, there is a lack of a systematic and effective evaluation method for deep learning medical auxiliary diagnosis models.

[0003] The existing deep learning medical auxiliary diagnosis classification model for diabetic retinopathy receives the fundus images taken by patients as input when applied, and outputs the probability of the image being classified into five severity levels of diabetic retinopathy, providing doctors with auxiliary discrimination basis for classification diagnosis. In the field of deep learning classification models, deep neural networks are trained through standard data sets, and then have the ability to recognize input images with similar features to the data set images. The model can show good classification performance in the same or similar environment as the training set, but when the shooting environment or target features change, there will often be a significant performance drop. This typical problem is often described as the domain generalization problem of deep learning models. However, most of the current diabetic retinopathy datasets only shoot conventional fundus lesion images, that is, patients do not have other obvious visual diseases except diabetic retinopathy, and the images taken for examination have no other parallax. In actual situations, patients often suffer from other ophthalmic diseases of varying severity, such as myopia, presbyopia, astigmatism, glaucoma and cataracts. These ophthalmic diseases will affect the shooting quality of the photos and change the recognition characteristics of the dataset. And due to patient resource limitations, there is a lack of large data sets with a single other target parallax. Therefore, the currently proposed deep learning classification model for diabetic retinopathy has not been trained for other parallaxes. When faced with fundus images of other parallaxes, there will be a significant performance degradation, which affects the credibility of the model under actual deployment. There is currently a lack of methods to evaluate this situation in order to evaluate the actual application performance of the model.

[0004] In view of this, one or more embodiments of the present specification simultaneously relate to a basic structure of a human eye model and a fundus camera simulating an optical system, a target parallax simulation structure of multiple human eye models and fundus camera combined optical systems, and a method for generating a parallax image simulating a diabetic retinopathy image, so as to solve the problem in the prior art that the performance degradation of the diabetic retinopathy deep learning classification model cannot be systematically and effectively evaluated when facing parallax caused by environmental changes and other diseases in the patient's eyes. Summary of the invention

[0005] The present invention provides a method for evaluating a deep learning classification model for diabetic retinopathy, characterized in that the method comprises a set of human eye models and fundus camera combined optical systems, a method for generating a simulated parallax image of diabetic retinopathy based on computational optics, and a subsequent generalization performance test of the deep learning classification model. The method comprises the following steps: Step 1: According to the human eye model and the fundus camera structure parameters, establish the basic structure of the optical system of the human eye model and the fundus camera combination; Step 2: Establish a series of simulation structures of the combined optical system based on the target parallax optical characteristics and pathological features for the deep learning classification model evaluation of diabetic retinopathy; Step 3: According to the established simulation structure, the point spread function of the entire system is calculated, and a certain proportion of the original data set is subjected to convolution image processing to obtain a simulated parallax image of the original data set at the target parallax; Step 4: Input the mixed data set consisting of the original data set images and the simulated disparity images into the model to be evaluated for deep learning prediction and classification; Step 5: Calculate various deep learning model evaluation indicators based on the classification results; Step 6: Based on the calculation indicators, obtain the generalization performance evaluation results of the deep learning model on the target disparity.

[0006] The human eye model and fundus camera combined optical system, its basic structure includes a simulated human eye model, a retinal objective lens group and an imaging lens group. The simulated human eye model, the retinal objective lens group and the imaging lens group are arranged on the same optical axis from left to right, and the simulated human eye model includes a retinal surface, a rear surface of the lens, an anterior surface of the lens, a pupil stop surface, a rear surface of the cornea, and an anterior surface of the cornea arranged from left to right. The retinal surface bulges to the left, and its curvature radius ranges from greater than or equal to 11.50 mm to less than or equal to 12.50 mm; the rear surface of the lens bulges to the left, and its curvature radius ranges from greater than or equal to 5.50 mm to less than or equal to 6.50 mm; the anterior surface of the lens bulges to the right, and its curvature radius ranges from greater than or equal to 9.50 mm to less than or equal to 10.50 mm. The pupil surface only acts as a stop and is set as an ideal plane. The posterior surface of the cornea bulges to the right, and its curvature radius ranges from greater than or equal to 6.00 mm to less than or equal to 6.50 mm. The anterior surface of the cornea bulges to the right, and its curvature radius ranges from greater than or equal to 7.50 mm to less than or equal to 8.00 mm.

[0007] The retina objective lens group includes a first lens surface, a second lens surface, a third lens surface, a fourth lens surface, a fifth lens surface and a sixth lens surface which are arranged in sequence from left to right; the first lens surface is convex to the left, and the curvature radius range is greater than or equal to 26.15 mm and less than or equal to 26.65 mm; the second lens surface is convex to the right, and the curvature radius range is greater than or equal to 16.50 mm and less than or equal to 17.50 mm; the third lens surface and the second lens surface are arranged coplanar; the fourth lens surface is convex to the left, and the curvature radius range is greater than or equal to 57.00 mm and less than or equal to 57.50 mm; the fifth lens surface is convex to the left, and the curvature radius range is greater than or equal to 23.00 mm and less than or equal to 23.50 mm; the sixth lens surface is convex to the left, and the curvature radius range is greater than or equal to 54.50 mm and less than or equal to 55.50 mm.

[0008] The imaging lens group includes a seventh lens surface, an eighth lens surface, a ninth lens surface, a tenth lens surface, an eleventh lens surface, a twelfth lens surface, a thirteenth lens surface, a fourteenth lens surface, a fifteenth lens surface and a sixteenth lens surface arranged in sequence from left to right; the seventh lens surface is convex to the left, and the curvature radius range is greater than or equal to 21.75mm and less than or equal to 22.50mm; the eighth lens surface is convex to the right, and the curvature radius range is greater than or equal to 23.50mm and less than or equal to 24.50mm; the ninth lens surface is arranged coplanar with the eighth lens surface; the tenth lens surface is convex to the left, and the curvature radius range is greater than or equal to 96.00mm and less than or equal to 97.00mm; the eleventh lens surface The surface is convex to the left, and the curvature radius range is greater than or equal to 12.20mm and less than or equal to 12.70mm; the twelfth lens surface is convex to the right, and the curvature radius range is greater than or equal to 35.20mm and less than or equal to 35.70mm; the thirteenth lens surface and the twelfth lens surface are coplanar; the fourteenth lens surface is convex to the left, and the curvature radius range is greater than or equal to 7.50mm and less than or equal to 8.50mm; the fifteenth lens surface is convex to the left, and the curvature radius range is greater than or equal to 12.30mm and less than or equal to 12.80mm; the sixteenth lens surface is convex to the left, and the curvature radius range is greater than or equal to 18.50mm and less than or equal to 19.00mm.

[0009] Furthermore, in the simulated human eye model, the thickness range of the independent lens composed of the retinal surface and the rear surface of the lens is [16.10mm, 17.10mm], the thickness range of the independent lens composed of the rear surface of the lens and the front surface of the lens is [3.50mm, 4.50mm], the thickness range of the independent lens composed of the front surface of the lens and the rear surface of the cornea is [3.20mm, 4.20mm], and the thickness range of the independent lens composed of the rear surface of the cornea and the front surface of the cornea is [0.42mm, 1.42mm].

[0010] Furthermore, the thickness range of the independent lens composed of the first lens surface and the second lens surface is [11.76mm, 12.76mm], the thickness range of the independent lens composed of the third lens surface and the fourth lens surface is [8.45mm, 9.45mm], and the thickness range of the independent lens composed of the fifth lens surface and the sixth lens surface is [14.11mm, 15.11mm].

[0011] Furthermore, the thickness range of the independent lens composed of the seventh lens surface and the eighth lens surface is [5.20 mm, 6.20 mm], the thickness range of the independent lens composed of the ninth lens surface and the tenth lens surface is [5.25 mm, 6.25 mm], the thickness range of the independent lens composed of the eleventh lens surface and the twelfth lens surface is [5.48 mm, 6.48 mm], the thickness range of the independent lens composed of the thirteenth lens surface and the fourteenth lens surface is [4.32 mm, 5.32 mm], and the thickness range of the independent lens composed of the fifteenth lens surface and the sixteenth lens surface is [5.20 mm, 6.20 mm].

[0012] The human eye model group comprises a vitreous body simulation lens composed of a retinal surface and a rear surface of a lens, a lens simulation lens and a cornea simulation lens.

[0013] The retina objective lens group comprises a double cemented lens and a plano-convex lens.

[0014] The imaging lens group comprises a front imaging lens group and a rear imaging lens group, the front imaging lens group is a double-cemented lens, and the rear imaging lens group comprises a double-cemented lens and a positive lens.

[0015] Furthermore, the basic structure of the human eye model optical system is characterized in that: the glass material of the independent lens composed of the retinal surface and the rear surface of the lens is VITREOUS; the glass material of the independent lens composed of the rear surface of the lens and the front surface of the lens is LENS; the glass material of the independent lens composed of the front surface of the lens and the rear surface of the cornea is AQUEOUS; the glass material of the independent lens composed of the rear surface of the cornea and the front surface of the cornea is CORNEA. The glass material of the independent lens composed of the first lens surface and the second lens surface is H-LAK2; the glass material of the independent lens composed of the third lens surface and the fourth lens surface is F2; ​​the glass material of the independent lens composed of the fifth lens surface and the sixth lens surface is H-LAK2; the glass material of the independent lens composed of the seventh lens surface and the eighth lens surface is H-LAK3; the glass material of the independent lens composed of the ninth lens surface and the tenth lens surface is ZF6; the glass material of the independent lens composed of the eleventh lens surface and the twelfth lens surface is H-LAK3; the glass material of the independent lens composed of the thirteenth lens surface and the fourteenth lens surface is ZF6; the glass material of the independent lens composed of the fifteenth lens surface and the sixteenth lens surface is H-LAK3.

[0016] The basic structure of the human eye model and fundus camera combined optical system described above is highly realistic as a basic method for fundus examination of diabetic retinopathy. It is also the physical basis for the subsequent generation of parallax images of diabetic retinopathy image simulation based on computational optics, and can realize effective computational simulation of the shooting process of fundus examination of diabetic retinopathy.

[0017] The method for generating a parallax image simulating a retinal diabetic retinopathy image comprises the following steps: 1) According to the optical characteristics and pathological mechanism of the target diabetic retinopathy simulated parallax image, the optical system is modified accordingly to obtain the target parallax simulation structure of the human eye model and fundus camera combined optical system. The simulated structural changes include corresponding changes to the curvature parameters, material parameters and light transmission characteristic parameters of the lens lens and corneal lens in the human eye model, adding additional optical structures to the human eye model, and corresponding changes to the retinal objective lens group and imaging lens group to obtain a simulated lesion simulation model with the corresponding optical characteristics of the target diabetic retinopathy image; 2) According to the simulated lesion simulation model optical system described in step 1), the point spread function from each point on the object plane of the optical system to the image plane is calculated, wherein the object plane is the retinal plane in the simulated lesion simulation model optical system. The calculated field of view angle of the object plane should be no less than 40°. The calculation method is based on the ray tracing method provided by the optical design software, and the light within the field of view of the aperture stop is traced and calculated. The point spread function obtained is a matrix related to the object plane sampling rate and pixel size, and contains the propagation matrix information from each point on the object plane to the image plane; 3) Based on the obtained point spread function, the target diabetic retinopathy image simulation parallax image is generated by convolving the source bitmap file with the point spread function array through image simulation according to the original source bitmap image file of diabetic retinopathy. The influences considered in this method include diffraction, aberration, distortion, relative illumination, image direction and polarization existing in the optical system of the lesion simulation model. The original source bitmap image file of diabetic retinopathy should have a clear field of view, clear classification, clear and visible basis for lesion classification, and no obvious abnormal aberration. The generated simulated parallax image has the optical characteristics of the target simulated parallax, such as diffraction, aberration, distortion, relative illumination, image direction and polarization, reflecting the imaging characteristics and pathological mechanism of the target parallax simulation structure of the optical system of the human eye model and fundus camera combination.

[0018] The deep learning classification model generalization performance test is characterized in that the method comprises the following steps: 1) Based on the original diabetic retinopathy image and the generated target diabetic retinopathy image simulated disparity image, a certain proportion of random mixing is used as the diabetic retinal model target disparity evaluation mixed dataset (referred to as mixed dataset) for testing. The characteristics of the mixed dataset are that, in addition to the fundus and lesion features of the original diabetic retinopathy image, it should have the basic annotations required by the deep learning classification model, that is, the five labels corresponding to each level of diabetic retinopathy, and the optical features corresponding to the target simulated disparity and the severity of the corresponding simulated disparity; 2) Based on the mixed data set, a test process is performed on the generalization ability of the diabetic retinopathy deep learning model to be evaluated (hereinafter referred to as the model to be evaluated) for the target lesion. In addition to its own deep learning network structure and corresponding data enhancement preprocessing method, the model to be evaluated should also have weight information based on its own data set training, and the weight information can be obtained by inference training of any parameters on any training set; 3) The test process includes mixed data set input and reading, classification prediction of the model to be evaluated, calculation of evaluation indicators based on the model classification results, and evaluation of model generalization performance based on the evaluation indicators.

[0019] Furthermore, the mixed data set input and reading step should read the mixed data set image, the label corresponding to the image, and the simulated parallax type and severity corresponding to the image.

[0020] Furthermore, the classification prediction process of the model to be evaluated, that is, the deep learning model outputs a prediction probability matrix of each input image in five different categories through feature recognition and model reasoning for each input mixed data set image.

[0021] Furthermore, the evaluation index calculation process calculates the accuracy, area under the ROC curve (AUC), and F1 score indicators based on the prediction probability output in the previous step.

[0022] Furthermore, the precision index is calculated by extracting the maximum value of the prediction probability matrix output by the model to obtain the prediction result, comparing the prediction result matrix with the label matrix, and taking the ratio of the prediction result to the label result as the precision.

[0023] Further, the area under the ROC curve index (Area under curve). The ROC curve is a receiver operating characteristic curve (receiver operating characteristic curve), which is an index invented by electronic engineers and radar engineers to detect enemy vehicles (aircraft, ships) on the battlefield, and belongs to signal detection theory. The abscissa of the ROC curve is the false positive rate (also called false positive rate, False Positive Rate), which is the probability of being judged as a positive example but not a true example, that is, the probability of being judged as a positive example in a true negative example, and the ordinate is the true positive rate (true positive rate, True Positive Rate), which is the probability of being judged as a positive example and also a true example, that is, the probability of being judged as a positive example in a true example (also known as the positive example recall rate). The AUC index can be used as an index of the quality of the model when comparing different classification models. The larger the area value below the ROC curve, the better the model classifier effect is, and the more application value it has.

[0024] Furthermore, the F1 score is an indicator used in statistics to measure the accuracy of a classification model. It is defined as the harmonic mean of precision and recall, taking into account both the precision and recall of the classification model. The closer the value is to 1, the higher the model classification performance.

[0025] Furthermore, in the process of evaluating the generalization performance of the model based on the evaluation index, after completing the calculation of the index, the target disparity and disparity severity of each output index and the input data set are counted as the performance of the model corresponding to the target disparity and disparity severity, and the generalization performance of the model to be evaluated when facing the target disparity is obtained according to the change of the performance under different severity levels. Specifically, for a model with higher generalization performance, in addition to showing normal classification performance under the standard data set, each index should maintain a high stability when facing a disparity different from the standard data set.

[0026] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art: 1) The modeling process of this method uses optical design software, which can conduct targeted analysis and design for the selection of optical elements in the optical system. The proposed human eye model and fundus camera combined optical system target parallax simulation structure can realistically simulate the fundus examination process of diabetic retinopathy; 2) The image simulation process of this method can generate target diabetic retinopathy image simulation disparity images that are missing from existing public datasets based on the target disparity simulation structure of the combined optical system of the human eye model and the fundus camera, effectively making up for the shortcomings of the existing datasets and having the possibility of generating more data; 3) The deep learning classification model evaluation process of this method can test the generalization performance and actual performance of the model to be evaluated by simulating disparity images of the generated target diabetic retinopathy images, which makes up for the lack of evaluation methods for deep learning models of medical-assisted classification in the current metrology field. It can effectively evaluate the performance of the model to be evaluated on standard data sets and various target disparity data sets, and provide an effective and reliable basis for the subsequent deployment in the field of medical-assisted diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a light path schematic diagram of a simulated structure of a human eye model and fundus camera combined optical system provided by an embodiment of the present specification.

[0028] Figure 2 It is a schematic diagram of a spherical parallax simulation lens 400 in a simulation structure of a human eye model and a fundus camera combined optical system provided in three embodiments of this specification.

[0029] Figure 3 It is a schematic diagram of a cylindrical parallax simulation lens 400 in a simulation structure of a human eye model and a fundus camera combined optical system provided in three embodiments of this specification.

[0030] Figure 4 It is a schematic diagram of the generalization performance test flow of the deep learning classification model for diabetic retinopathy provided by an embodiment of the present specification.

[0031] Figure 5 This is a table of evaluation results of a deep learning classification model for diabetic retinopathy provided by an embodiment of this specification.

[0032] Figure 6 This is a curve chart of the evaluation results of the deep learning classification model for diabetic retinopathy provided by an embodiment of the present specification. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and this specification can be implemented in many other ways different from those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited by the specific implementation disclosed below.

[0034] Figure 1 It is the basic structure of the human eye model and fundus camera simulating optical system of the embodiment of the present invention. The fundus that needs to be detected when the optical system is working is replaced by the human eye model 100 in this embodiment. At the same time, a detector 500 is set on the optical image plane for detecting the image information.

[0035] like Figure 1 As shown, a basic structure of a human eye model and a fundus camera simulated optical system includes a human eye model 100, a retina objective lens group 200 and an imaging lens group 300. The human eye model 100, the retina objective lens group 200 and the imaging lens group 300 are arranged on the same optical axis from left to right, and the simulated human eye model includes a retinal surface 101, a rear surface of the lens 102, an anterior surface of the lens 103, a pupil stop surface 104, a rear surface of the cornea 105, and an anterior surface of the cornea 106. The retinal surface 101 is convex to the left, and its curvature radius range is greater than or equal to 11.50 mm and less than or equal to 12.50 mm; the rear surface of the lens 102 is convex to the left, and its curvature radius range is greater than or equal to 5.50 mm and less than or equal to 6.50 mm; the anterior surface of the lens 103 is convex to the right, and its curvature radius range is greater than or equal to 9.50 mm and less than or equal to 10.50 mm. The pupil surface 104 only acts as a stop and is set as an ideal plane. The corneal posterior surface 105 bulges to the right, and its curvature radius ranges from 6.00 mm to 6.50 mm. The corneal anterior surface 106 bulges to the right, and its curvature radius ranges from 7.50 mm to 8.00 mm.

[0036] The retina objective lens group 200 includes a first lens surface 201, a second lens surface 202, a third lens surface 203, a fourth lens surface 204, a fifth lens surface 205 and a sixth lens surface 206 arranged in sequence from left to right; the first lens surface 201 is convex toward the left direction, and its curvature radius range is greater than or equal to 26.15mm and less than or equal to 26.65mm; the second lens surface 202 is convex toward the right direction, and its curvature radius range is greater than or equal to 16.50mm and less than or equal to 17.50mm The third lens surface 203 and the second lens surface 204 are coplanarly arranged; the fourth lens surface 204 is convex toward the left direction, and the curvature radius range is greater than or equal to 57.00 mm and less than or equal to 57.50 mm; the fifth lens surface 205 is convex toward the left direction, and the curvature radius range is greater than or equal to 23.00 mm and less than or equal to 23.50 mm; the sixth lens surface 206 is convex toward the left direction, and the curvature radius range is greater than or equal to 54.50 mm and less than or equal to 55.50 mm.

[0037] The imaging lens group 300 includes a seventh lens surface 307, an eighth lens surface 308, a ninth lens surface 309, a tenth lens surface 310, an eleventh lens surface 311, a twelfth lens surface 312, a thirteenth lens surface 313, a fourteenth lens surface 314, a fifteenth lens surface 315 and a sixteenth lens surface 316 which are arranged in sequence from left to right; the seventh lens surface 307 is convex toward the left direction, and the curvature radius thereof is in the range of greater than or equal to 21.75 mm and less than or equal to 22.50 mm; the eighth lens surface 308 is convex toward the right direction, and the curvature radius thereof is in the range of greater than or equal to 23.50 mm and less than or equal to 24.50 mm; the ninth lens surface 309 is arranged coplanar with the eighth lens surface 308; the tenth lens surface 310 is convex toward the left direction, and the curvature radius thereof is in the range of greater than or equal to 96.00 mm and less than or equal to 9 7.00mm; the eleventh lens surface 311 is convex toward the left direction, and the curvature radius range is greater than or equal to 12.20mm and less than or equal to 12.70mm; the twelfth lens surface 312 is convex toward the right direction, and the curvature radius range is greater than or equal to 35.20mm and less than or equal to 35.70mm; the thirteenth lens surface 313 and the twelfth lens surface 312 are arranged on the same plane; the fourteenth lens surface 314 is convex toward the left direction, and the curvature radius range is greater than or equal to 7.50mm and less than or equal to 8.50mm; the fifteenth lens surface 315 is convex toward the left direction, and the curvature radius range is greater than or equal to 12.30mm and less than or equal to 12.80mm; the sixteenth lens surface 316 is convex toward the left direction, and the curvature radius range is greater than or equal to 18.50mm and less than or equal to 19.00mm.

[0038] The thickness range of the independent lens composed of the retinal surface 101 and the rear surface of the lens 102 in the simulated human eye model is [16.10mm, 17.10mm], the thickness range of the independent lens composed of the rear surface of the lens 102 and the front surface of the lens 103 is [3.50mm, 4.50mm], the thickness range of the independent lens composed of the front surface of the lens 103 and the rear surface of the cornea 105 is [3.20mm, 4.20mm], and the thickness range of the independent lens composed of the rear surface of the cornea 105 and the front surface of the cornea 106 is [0.42mm, 1.42mm].

[0039] The thickness range of the independent lens composed of the first lens surface 201 and the second lens surface 202 is [11.76mm, 12.76mm], the thickness range of the independent lens composed of the third lens surface 203 and the fourth lens surface 204 is [8.45mm, 9.45mm], and the thickness range of the independent lens composed of the fifth lens surface 205 and the sixth lens surface 206 is [14.11mm, 15.11mm].

[0040] The thickness range of the independent lens composed of the seventh lens surface 307 and the eighth lens surface 308 is [5.20mm, 6.20mm], the thickness range of the independent lens composed of the ninth lens surface 309 and the tenth lens surface 310 is [5.25mm, 6.25mm], the thickness range of the independent lens composed of the eleventh lens surface 311 and the twelfth lens surface 312 is [5.48mm, 6.48mm], the thickness range of the independent lens composed of the thirteenth lens surface 313 and the fourteenth lens surface 314 is [4.32mm, 5.32mm], and the thickness range of the independent lens composed of the fifteenth lens surface 315 and the sixteenth lens surface 316 is [5.20mm, 6.20mm].

[0041] The human eye model group comprises a vitreous body simulation lens composed of a retinal surface and a rear surface of a lens, a lens simulation lens and a cornea simulation lens.

[0042] The retina objective lens group comprises a double cemented lens and a plano-convex lens.

[0043] The imaging lens group comprises a front imaging lens group and a rear imaging lens group, the front imaging lens group is a double-cemented lens, and the rear imaging lens group comprises a double-cemented lens and a positive lens.

[0044] The basic structure of the human eye model optical system is that the glass material of the independent lens composed of the retinal surface 101 and the rear surface of the lens 102 is VITREOUS; the glass material of the independent lens composed of the rear surface of the lens 102 and the front surface of the lens 103 is LENS; the glass material of the independent lens composed of the front surface of the lens 103 and the rear surface of the cornea 105 is AQUEOUS; the glass material of the independent lens composed of the rear surface of the cornea 105 and the front surface of the cornea 106 is CORNEA. The glass material of the independent lens composed of the first lens surface 201 and the second lens surface 202 is H-LAK2; the glass material of the independent lens composed of the third lens surface 203 and the fourth lens surface 204 is F2; ​​the glass material of the independent lens composed of the fifth lens surface 205 and the sixth lens surface 206 is H-LAK2; the glass material of the independent lens composed of the seventh lens surface 307 and the eighth lens surface 308 is H-LAK3; the glass material of the independent lens composed of the ninth lens surface 309 and the tenth lens surface 310 is ZF6; the glass material of the independent lens composed of the eleventh lens surface 311 and the twelfth lens surface 312 is H-LAK3; the glass material of the independent lens composed of the thirteenth lens surface 313 and the fourteenth lens surface 314 is ZF6; the glass material of the independent lens composed of the fifteenth lens surface 315 and the sixteenth lens surface 316 is H-LAK3.

[0045] Furthermore, in this embodiment, the most common basic target parallax, namely, defocus parallax and astigmatism parallax, is simulated. Figure 1 The human eye model and fundus camera simulated optical system basic structure is added with a parallax simulation lens 400. The parallax simulation independent lens 400 is composed of a parallax simulation lens front surface and a parallax simulation lens rear surface, and its glass material is H-LAK2.

[0046] Furthermore, the parallax simulation lens has different designs under different parallax simulation severity levels. In this embodiment, the parallax simulation lens has different designs under different parallax simulation severity levels. Figure 2 The three simulation structures are shown. Figure 3 Three simulation structures are shown.

[0047] The first simulation structure is as follows Figure 2 (1). Among them, the parallax simulation lens rear surface sagittal height map is as follows Figure 2 (1)(a) shows a standard spherical surface, convex to the left, with a radius of curvature of 2760 mm. The front surface sagittal height diagram of the parallax simulation lens is shown in Figure 2 As shown in (1)(b), it is a spherical standard surface, convex to the right, with a curvature radius of 2760 mm. The parallax simulation lens formed by it provides a defocus diopter of 0.5D for the parallax simulation optical system.

[0048] The second simulation structure is as follows Figure 2 (2) shown in Figure 2. Among them, the parallax simulation lens rear surface sagittal height map is as follows: Figure 2 (2)(a) shows a standard spherical surface, convex to the left, with a radius of curvature of 1380 mm. Figure 2 (2)(b) is a spherical standard surface, convex to the right, with a curvature radius of 1380 mm. The parallax simulation lens formed by it provides a 1.0D defocus diopter for the parallax simulation optical system.

[0049] The third simulation structure is as follows Figure 2 (3) shown in Figure 3. Among them, the parallax simulation lens rear surface sagittal height map is as follows: Figure 2 (3)(a) shows a standard spherical surface, convex to the left, with a radius of curvature of 920 mm. The front surface sagittal height diagram of the parallax simulation lens is shown in Figure 2 (3)(b) shows a spherical standard surface, which is convex to the right and has a curvature radius of 920 mm. The parallax simulation lens formed by the surface provides a 1.5D defocus diopter for the parallax simulation optical system.

[0050] The fourth simulation structure is as follows Figure 3 (1). Among them, the parallax simulation lens rear surface sagittal height map is as follows Figure 3(1)(a) shows a standard cylindrical surface, convex to the left, with a radius of curvature of 2760 mm. The front surface sagittal height diagram of the parallax simulation lens is shown in Figure 3 (1)(b) shows a cylindrical standard surface, which is convex to the right and has a curvature radius of 2760 mm. The parallax simulation lens formed by the cylindrical standard surface provides a 0.5D astigmatism diopter for the parallax simulation optical system.

[0051] The fifth simulation structure is as follows Figure 3 (2) shown in Figure 2. Among them, the parallax simulation lens rear surface sagittal height map is as follows: Figure 3 (2)(a) shows a standard cylindrical surface, convex to the left, with a radius of curvature of 1380 mm. The front surface sagittal height diagram of the parallax simulation lens is shown in Figure 3 (2)(b) shows a cylindrical standard surface, which is convex to the right and has a curvature radius of 1380 mm. The parallax simulation lens formed by the cylindrical standard surface provides a 1.0D astigmatism diopter for the parallax simulation optical system.

[0052] The sixth simulation structure is as follows Figure 3 (3) shown in Figure 3. Among them, the parallax simulation lens rear surface sagittal height map is as follows: Figure 3 (3)(a) shows the standard cylindrical surface, which is convex to the left and has a curvature radius of 920 mm. Figure 3 (3)(b) shows a cylindrical standard surface, which is convex to the right and has a curvature radius of 920 mm. The parallax simulation lens formed by the cylindrical standard surface provides a 1.5D astigmatism diopter for the parallax simulation optical system.

[0053] Based on the six simulation structures described above, the point spread function from each point on the object plane of the optical system to the image plane is calculated, wherein the object plane is the retinal plane in the simulation structure. The calculated field of view of the object plane should be no less than 40°. The calculation method is based on the ray tracing method provided by the optical design software, and the light rays within the field of view of the aperture stop are traced and calculated. The point spread function obtained is a matrix related to the object plane sampling rate and pixel size, and contains the propagation matrix information from each point on the object plane to the image plane. In this example, the sampling resolution of the point spread function on the object side is 30 microns, and the sampling resolution on the image plane is 4 microns. That is, for a retinal image with a field of view angle of 40° on the object plane, the diffusion of each point light source on the image plane with a sampling rate of 4 microns is calculated at a sampling rate of 30 microns. The accuracy required for diabetic retinopathy examination is met.

[0054] Furthermore, the point spread function of the central point light source of the three defocus simulation structures is as follows: Figure 2 As shown in (1-3)(c), it can be seen that under the defocus diopter of (1) 0.5D (2) 1.0D (3) 1.5D, the diffusion of the point light source is reflected as a gradually increasing defocus spot. The diffusion of the point light source overlaps each other under convolution, thereby affecting the final image quality.

[0055] Furthermore, the point spread function of the central point light source of the three astigmatism simulation structures is as follows: Figure 3 As shown in (1-3)(c), it can be seen that under the astigmatism diopter of (1) 0.5D (2) 1.0D (3) 1.5D, the diffusion of the point light source is reflected as a gradually increasing astigmatism spot. Unlike defocus, the point light source has the greatest impact on the astigmatism axis and the least impact in the direction perpendicular to the astigmatism axis. The diffusion conditions overlap each other under convolution, thereby affecting the final image quality.

[0056] Based on the obtained point spread function, the target diabetic retinopathy image simulation parallax image is generated by convolving the source bitmap file with the point spread function array through image simulation according to the original source bitmap image file of diabetic retinopathy. The influences considered in this method include diffraction, aberration, distortion, relative illumination, image direction and polarization existing in the simulated structure. The original source bitmap image file of diabetic retinopathy should have a clear field of view, clear classification, clear and visible basis for lesion classification, and no obvious abnormal aberration. The generated simulated parallax image has the optical characteristics of the target simulated parallax, such as diffraction, aberration, distortion, relative illumination, image direction and polarization, reflecting the imaging characteristics and lesion mechanism of the simulated structure.

[0057] The generalization performance test process of the deep learning classification model in this embodiment is as follows: Figure 4 As shown, the following steps are included: 1) Based on the original data set 10 of diabetic retinopathy and the target diabetic retinopathy image simulated disparity image generated by the combined model simulation structure 20 including the human eye model 21, the disparity lens 22, and the camera model 23, a certain proportion of random mixing is used as the diabetic retinal model target disparity evaluation mixed data set 30 (referred to as the mixed data set) for testing. The characteristics of the mixed data set are that, in addition to the fundus and lesion characteristics of the original image of diabetic retinopathy, it should have the basic annotations required by the deep learning classification model, that is, five labels corresponding to each level of diabetic retinopathy, and have the optical characteristics corresponding to the target simulated disparity and the severity of the corresponding simulated disparity. In this embodiment, six groups of public data sets are selected, including APTOS, DDR, DEEPDR, FGADR, IDRID and Messidor, wherein each group uses 20% of the pictures as the original images of diabetic retinopathy, and the generated simulated disparity images have different degrees of defocus diopter and astigmatism diopter imposed by the six structures on the basis of the diabetic retinopathy characteristics of the original image. Each target disparity is labeled with 0.5D, 1.0D, and 1.5D of diopter as the target disparity severity; 2) Based on the mixed data set, the deep learning model 40 of diabetic retinopathy to be evaluated (hereinafter referred to as the model to be evaluated) is subjected to a test process for the generalization ability of the model to be evaluated for the target lesion. In addition to retaining its own deep learning network structure and the corresponding data enhancement preprocessing method, the model to be evaluated should also have weight information after training based on its own data set, and the weight information can be obtained by training any parameters on any training set. In this embodiment, in order to verify the reliability of the method in the application scenario, four advanced public diabetic retinopathy classification models are trained on the public data set. The public models include GDR-net, mixup, DeepDR, and PCM. The training method maintains the network structure and pre-training method provided by the public model unchanged, and the same standard classification training is carried out on the training set of the above public data set. The classification training is carried out according to the training method provided by the model, and the same original images and corresponding labels are input, and the same batch size and epoch rounds are used for training. In this embodiment, the batch size is 32 and the epoch rounds are 30. After the model training, good classification performance is achieved on the standard test set; 3) The test process includes inputting and reading a mixed data set 30, classification prediction results of the model 50, calculation of evaluation indicators 60 based on the classification results of the model, and evaluation of model generalization performance 70 based on the evaluation indicators.

[0058] Furthermore, the mixed data set input and reading step inputs and reads the mixed data set image 30, the label corresponding to the image, and the simulated parallax type and severity corresponding to the image. In this embodiment, each image in the mixed data set has a simulated parallax type such as no parallax, defocus, and astigmatism lesions in addition to the category label of the lesion. The simulated parallax severity is such as 0.5D, 1.0D, 1.5D.

[0059] Furthermore, the classification prediction process of the model, that is, for each input mixed data set image, outputs a prediction probability matrix of each input image in five different categories through feature recognition and model reasoning. In this embodiment, the model outputs corresponding prediction probabilities for 5 categories for the input 32 image information in each batch, and the matrix size is [32,5]. In each test, all classification prediction results 50 of the input images are saved for subsequent processing.

[0060] Furthermore, the evaluation index calculation process performs evaluation index 60 based on the prediction result 50 outputted in the previous step, including calculation of accuracy, area under ROC curve (AUC), and F1 score.

[0061] Furthermore, the precision index is calculated by extracting the maximum value of the prediction probability matrix output by the model to obtain the prediction result, and taking the ratio of the prediction result matrix to the label matrix result as the precision.

[0062] Further, the AUC index is defined as the area under the ROC curve and the coordinate axis. The abscissa of the ROC curve is the false positive rate, which is the probability of being judged as a positive example but not a true positive example, that is, the probability of being judged as a positive example among true negative examples, and the ordinate is the true positive rate, which is the probability of being judged as a positive example and also a true positive example, that is, the probability of being judged as a positive example among true negative examples. The AUC index can be used as an indicator of the quality of the model when comparing different classification models. The larger the area value under the ROC curve, the better the model classifier effect is and the more application value it has.

[0063] Furthermore, the F1 score is an indicator used in statistics to measure the accuracy of a classification model. It is defined as the harmonic mean of precision and recall, taking into account both the precision and recall of the classification model. The closer the value is to 1, the higher the model classification performance.

[0064] Furthermore, in the model generalization performance evaluation process based on the evaluation index, after completing the calculation of the index 60, the target disparity and disparity severity of each output index and the input data set are counted as the performance of the model corresponding to the target disparity and disparity severity, and according to the change of the performance under different severity levels, the generalization performance evaluation result 70 of the model to be evaluated when facing the target disparity is obtained. Specifically, for models with higher generalization performance, in addition to showing normal classification performance under the standard data set, when facing the disparity different from the standard data set, each indicator should maintain a high stability, reflecting that it has corresponding credibility in real applications. In this embodiment, the performance of the four selected classification models GDR-net, mixup, DeepDR, and PCM on the six mixed data sets APTOS, DDR, DEEPDR, FGADR, IDRID, and Messidor is shown as follows. Figure 5 As shown in the figure, it can be seen that the proposed indicators of the four models show an overall performance decline trend as the degree of defocus increases. In the following section, as an example, we will use the performance of the DeepDR model on the FGADR dataset. Figure 5 The underlined data were used for the main analysis.

[0065] Specifically, due to Figure 5All the indicators shown are capped at 1, so the units are selected as (%). The closer the value is to 100, the higher the model performance. In more detail, ACC, as an accuracy rate, reflects the direct accuracy rate of the model for classification. When the DeepDR model faces different degrees of defocus in the FGADR dataset, the accuracy rate drops from the initial 88.5 to 56.0, with an average decrease of 10.8 per 0.5D, which will have a significant impact on the performance of auxiliary diagnosis.

[0066] Furthermore, the AUC indicator, as the area under the ROC curve, shows the prediction results under different confidence thresholds. The higher the value, the higher the classification performance of the model. As shown in the figure, when the DeepDR model faces the FGADR dataset, the AUC indicator decreases from 99.0 to 86.9, with an average decrease of 4.03 per 0.5D, indicating that the receiver operating characteristic curve of the model is significantly affected.

[0067] Furthermore, the F1 score index calculates the harmonic mean of precision and recall, taking into account the precision and recall of the classification model, and jointly reflects the prediction accuracy and case detection coverage of the model. When the DeepDR model faces different degrees of defocus in the FGADR dataset, the F1 score index drops from 88.4 to 52.3, with an average drop of 12.03 per 0.5D. This indicates that the comprehensive performance of the model's prediction accuracy and case detection coverage has declined.

[0068] Furthermore, in order to more intuitively represent the test results, Figure 5 The results are expressed in the form of a curve: Figure 6 , Figure 6 The three pictures show the accuracy of the results, the AUC curve and the F1 index decrease with the severity of the target disparity. The different curves in the pictures represent different classification models. As can be seen from the figure, the performance of the classification model has dropped significantly under different degrees of defocus, indicating that the generated hybrid dataset can effectively simulate the target disparity and has the function of evaluating the performance of the model at the corresponding disparity.

Claims

1. A method for evaluating a deep learning classification model for diabetic retinopathy, characterized in that: The method includes a human eye model and a fundus camera combined optical system, a method for generating a parallax image of simulated diabetic retinopathy based on computational optics, and a subsequent generalization performance test of a deep learning classification model. The method includes the following steps: Step 1: According to the structural parameters of the human eye model and the fundus camera, the basic structure of the optical system of the human eye model and the fundus camera is established; Step 2: Establish a series of simulation structures of the combined optical system based on the target parallax optical characteristics and pathological features for the deep learning classification model evaluation of diabetic retinopathy; Step 3: According to the established simulation structure, the point spread function of the entire system is calculated, and a certain proportion of the original data set is subjected to convolution image processing to obtain a simulated parallax image of the original data set at the target parallax; Step 4: Input the mixed data set consisting of the original data set images and the simulated disparity images into the model to be evaluated for deep learning prediction and classification; Step 5: Calculate various deep learning model evaluation indicators based on the classification results; Step 6: Based on the calculation indicators, obtain the generalization performance evaluation results of the deep learning model on the target disparity.

2. The human eye model and fundus camera combined optical system according to claim 1, wherein the basic structure comprises a simulated human eye model, a retina objective lens group and an imaging lens group. The simulated human eye model, the retina objective lens group and the imaging lens group are arranged on the same optical axis from left to right, and the characteristics are as follows: The simulated human eye model includes a retinal surface, a rear surface of the lens, an anterior surface of the lens, a pupil stop surface, a rear surface of the cornea, and an anterior surface of the cornea, which are arranged in sequence from left to right. The retinal surface bulges to the left, and its radius of curvature ranges from greater than or equal to 11.50 mm to less than or equal to 12.50 mm; the rear surface of the lens bulges to the left, and its radius of curvature ranges from greater than or equal to 5.50 mm to less than or equal to 6.50 mm; the anterior surface of the lens bulges to the right, and its radius of curvature ranges from greater than or equal to 9.50 mm to less than or equal to 10.50 mm. The pupil surface only acts as an aperture and is set as an ideal plane. The rear surface of the cornea bulges to the right, and its radius of curvature ranges from greater than or equal to 6.00 mm to less than or equal to 6.50 mm; the anterior surface of the cornea bulges to the right, and its radius of curvature ranges from greater than or equal to 7.50 mm to less than or equal to 8.00 mm; The retina objective lens group includes a first lens surface, a second lens surface, a third lens surface, a fourth lens surface, a fifth lens surface and a sixth lens surface which are arranged in sequence from left to right; the first lens surface is convex toward the left direction, and the curvature radius thereof is in the range of greater than or equal to 26.15mm and less than or equal to 26.65mm; the second lens surface is convex toward the right direction, and the curvature radius thereof is in the range of greater than or equal to 16.50mm and less than or equal to 17.50mm; the third lens surface and the second lens surface are arranged coplanarly; the fourth lens surface is convex toward the left direction, and the curvature radius thereof is in the range of greater than or equal to 57.00mm and less than or equal to 57.50mm; the fifth lens surface is convex toward the left direction, and the curvature radius thereof is in the range of greater than or equal to 23.00mm and less than or equal to 23.50mm; the sixth lens surface is convex toward the left direction, and the curvature radius thereof is in the range of greater than or equal to 54.50mm and less than or equal to 55.50mm; The imaging lens group includes a seventh lens surface, an eighth lens surface, a ninth lens surface, a tenth lens surface, an eleventh lens surface, a twelfth lens surface, a thirteenth lens surface, a fourteenth lens surface, a fifteenth lens surface and a sixteenth lens surface arranged in sequence from left to right; the seventh lens surface is convex toward the left, and the curvature radius thereof is greater than or equal to 21.75 mm and less than or equal to 22.50 mm; The eighth lens surface is convex toward the right, and its curvature radius is greater than or equal to 23.50 mm and less than or equal to 24.50 mm; The ninth lens surface is coplanar with the eighth lens surface; the tenth lens surface is convex to the left, and the curvature radius thereof is greater than or equal to 96.00 mm and less than or equal to 97.00 mm; the eleventh lens surface is convex to the left, and the curvature radius thereof is greater than or equal to 12.20 mm and less than or equal to 12.70 mm; The twelfth lens surface is convex toward the right, and the curvature radius thereof is greater than or equal to 35.20 mm and less than or equal to 35.70 mm; The thirteenth lens surface and the twelfth lens surface are arranged on the same plane; the fourteenth lens surface is convex toward the left, and the curvature radius thereof is greater than or equal to 7.50 mm and less than or equal to 8.50 mm; The fifteenth lens surface is convex toward the left, and the curvature radius thereof is greater than or equal to 12.30 mm and less than or equal to 12.80 mm; The sixteenth lens surface is convex toward the left, and the curvature radius thereof is greater than or equal to 18.50 mm and less than or equal to 19.00 mm; The human eye model group comprises a vitreous body simulation lens composed of a retinal surface and a rear surface of a lens, a lens simulation lens and a cornea simulation lens; The retina objective lens group includes a doublet lens and a plano-convex lens; The imaging lens group comprises a front imaging lens group and a rear imaging lens group, the front imaging lens group is a double-cemented lens, and the rear imaging lens group comprises a double-cemented lens and a positive lens.

3. The human eye model and fundus camera combined optical system according to claim 1, characterized in that: The thickness range of the independent lens composed of the first lens surface and the second lens surface is [11.50mm, 12.50mm], the thickness range of the independent lens composed of the third lens surface and the fourth lens surface is [8.20mm, 9.20mm], and the thickness range of the independent lens composed of the fifth lens surface and the sixth lens surface is [14.10mm, 15.20mm].

4. The portable fundus camera optical system according to claim 1, characterized in that: The thickness range of the independent lens composed of the seventh lens surface and the eighth lens surface is [5.20mm, 6.20mm], the thickness range of the independent lens composed of the ninth lens surface and the tenth lens surface is [5.20mm, 6.20mm], the thickness range of the independent lens composed of the eleventh lens surface and the twelfth lens surface is [5.15mm, 6.15mm], the thickness range of the independent lens composed of the thirteenth lens surface and the fourteenth lens surface is [4.25mm, 5.25mm], and the thickness range of the independent lens composed of the fifteenth lens surface and the sixteenth lens surface is [5.25mm, 6.25mm].

5. The human eye model optical system according to claim 1, characterized in that: The glass material of the independent lens composed of the retinal surface and the rear surface of the lens is VITREOUS; the glass material of the independent lens composed of the rear surface of the lens and the front surface of the lens is LENS; the glass material of the independent lens composed of the front surface of the lens and the rear surface of the cornea is AQUEOUS; the glass material of the independent lens composed of the front surface of the cornea and the rear surface of the cornea is CORNEA.

6. The fundus camera optical system according to claim 1, characterized in that: The glass material of the independent lens composed of the first lens surface and the second lens surface is H-LAK2; the glass material of the independent lens composed of the third lens surface and the fourth lens surface is F2; ​​the glass material of the independent lens composed of the fifth lens surface and the sixth lens surface is H-LAK2; the glass material of the independent lens composed of the seventh lens surface and the eighth lens surface is H-LAK3; the glass material of the independent lens composed of the ninth lens surface and the tenth lens surface is ZF6; the glass material of the independent lens composed of the eleventh lens surface and the twelfth lens surface is H-LAK3; the glass material of the independent lens composed of the thirteenth lens surface and the fourteenth lens surface is ZF6; the glass material of the independent lens composed of the fifteenth lens surface and the sixteenth lens surface is H-LAK3.

7. A method for generating a parallax image based on computational optics for simulating diabetic retinopathy images according to claim 1, characterized in that: The method comprises the following steps: 1) Based on the basic structure of a human eye model and fundus camera combined optical system as described in claim 2 (hereinafter referred to as the basic structure), according to the optical characteristics and pathological mechanism of the target diabetic retinopathy simulated parallax image, the optical system is correspondingly changed in structure to obtain the target parallax simulation structure of the human eye model and fundus camera combined optical system (hereinafter referred to as the simulation structure). The changes in the simulation structure include making corresponding changes to the curvature parameters, material parameters and light transmission characteristic parameters of the lens lens and corneal lens in the human eye model as described in claim 1, adding additional optical structures to the human eye model as described in claim 1, and making corresponding changes to the retinal objective lens group and imaging lens group as described in claim 1, so as to obtain a simulated lesion simulation model having the corresponding optical characteristics of the target diabetic retinopathy image; 2) According to the simulation structure described in step 1), the point spread function from each point on the object plane of the optical system to the image plane is calculated, wherein the object plane is the retinal plane in the simulation structure. The calculated field of view of the object plane should be no less than 40°. The calculation method is based on the ray tracing method provided by the optical design software, and the light within the aperture stop field of view is traced and calculated. The point spread function obtained is a matrix related to the object plane sampling rate and pixel size, and contains the propagation matrix information from each point on the object plane to the image plane; 3) Based on the obtained point spread function, the target diabetic retinopathy image simulated disparity image is generated by convolving the source bitmap file with the point spread function array according to the original source bitmap image file of diabetic retinopathy. The influences considered in this method include diffraction, aberration, distortion, relative illumination, image direction and polarization in the simulated structure. The original source bitmap image file of diabetic retinopathy should have a clear field of view, clear classification, clear and visible basis for lesion classification, and no obvious abnormal aberration. The generated simulated disparity image has the optical characteristics of the target simulated disparity, such as diffraction, aberration, distortion, relative illumination, image direction and polarization, reflecting the imaging characteristics and lesion mechanism of the simulated structure.

8. The generalization performance test of the deep learning classification model is characterized by: The method comprises the following steps: 1) Based on the simulated disparity images of the original diabetic retinopathy images and the generated target diabetic retinopathy images, a certain proportion of random mixing is used as the target disparity evaluation mixed dataset of the diabetic retinal model for testing (referred to as the mixed dataset). The characteristics of the mixed dataset are that, in addition to the fundus and lesion features of the original diabetic retinopathy images, it should have the basic annotations required by the deep learning classification model, that is, the five labels corresponding to each level of diabetic retinopathy, and the optical features corresponding to the target simulated disparity and the severity of the corresponding simulated disparity; 2) Based on the mixed data set, a test process is performed on the deep learning model of diabetic retinopathy to be evaluated (hereinafter referred to as the model to be evaluated) to test the generalization ability of the model to be evaluated for the target lesion. In addition to its own deep learning network structure and corresponding data enhancement preprocessing method, the model to be evaluated should also have weight information based on its own data set training, and the weight information can be obtained by inference training of any parameters on any training set; 3) The test process includes mixed data set input and reading, classification prediction of the model to be evaluated, calculation of evaluation indicators based on the model classification results, and evaluation of model generalization performance based on the evaluation indicators.

9. Furthermore, the mixed data set input and reading step inputs and reads the mixed data set image, the label corresponding to the image, and the simulated parallax type and severity corresponding to the image; Furthermore, the classification prediction process of the model to be evaluated, that is, the deep learning model outputs a prediction probability matrix of each input image in five different categories through feature recognition and model reasoning for each input mixed data set image; Furthermore, the evaluation index calculation process calculates the accuracy, area under the ROC curve (AUC), and F1 score indicators based on the prediction probability output in the previous step; Furthermore, the precision index is calculated by extracting the prediction probability matrix output by the model through the maximum value to obtain the prediction result, and taking the ratio of the prediction result matrix to the label matrix result as the precision; Furthermore, the AUC indicator is defined as the area under the ROC curve and the coordinate axis. The ROC curve is the receiver operating characteristic curve, which was invented by electronic engineers and radar engineers to detect enemy vehicles (aircraft, ships) on the battlefield. It belongs to the signal detection theory. The horizontal axis of the ROC curve is the false positive rate (also called the false positive rate, False Positive Rate), which is the probability that a positive example is not a true positive example, that is, the probability of a true negative example being judged as a positive example. The vertical axis is the true positive rate (true positive rate, True Positive Rate), which is the probability that a positive example is also a true positive example, that is, the probability of a true positive example being judged as a positive example (also known as the positive recall rate). The AUC indicator can be used as an indicator of the quality of the model when comparing different classification models. The larger the area value under the ROC curve, the better the model classifier effect and the more application value it has; Furthermore, the F1 score is an indicator used in statistics to measure the accuracy of classification models. It is defined as the harmonic mean of precision and recall, taking into account both the precision and recall of the classification model. The closer the value is to 1, the higher the classification performance of the model. Furthermore, in the process of evaluating the generalization performance of the model based on the evaluation index, after completing the calculation of the index, the target disparity and disparity severity of each output index and the input data set are counted as the performance of the model corresponding to the target disparity and disparity severity, and the generalization performance of the model to be evaluated when facing the target disparity is obtained according to the change of the performance under different severity levels. Specifically, for a model with higher generalization performance, in addition to showing normal classification performance under the standard data set, each index should maintain a high stability when facing a disparity different from the standard data set, reflecting its corresponding credibility in real applications.

Images