A method for evaluating a diabetic retinopathy deep learning classification model

Abstract

The application provides a diabetic retinopathy deep learning classification model evaluation method, first, a human eye model and fundus camera combined optical system are described, the optical system combines the human eye model and the fundus camera model, realizes high restoration degree and high resolution of the diabetic retinopathy fundus examination simulation, and further based on the method, a diabetic retinopathy simulation parallax image generation method based on computational optics is provided, the method designs a parallax simulation system through ZEMAX, and calculates a simulation parallax image through a point spread function matrix. Based on the generated simulation parallax image, a further deep learning classification model generalization performance test is proposed, and the evaluation method can effectively evaluate the classification performance of the classification model on the target simulation parallax image.

Description

Technical Field

[0001] This invention relates to the evaluation of medical-aided deep learning classification models, and more specifically, to an evaluation method for a deep learning classification model of diabetic retinopathy. It involves a method for generating simulated disparity images of the target diabetic retinopathy using a combined optical system of a human eye model and a fundus camera. The simulated disparity images generated by this invention can be used to evaluate the performance of a deep learning classification model for diabetic retinopathy in the presence of disparity images. Background Technology

[0002] Currently, with the development of deep learning models, deep learning-based medical diagnostic models 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-based medical diagnostic models.

[0003] Existing deep learning-based medical-aided diagnostic classification models for diabetic retinopathy (DR) receive fundus images taken by patients as input and output the probability of classifying the images into five severity levels of DR, providing doctors with auxiliary diagnostic criteria. In the field of deep learning classification models, deep neural networks are trained on standard datasets, enabling them to identify input images with similar features to those in the dataset. The model exhibits good classification performance in environments identical to or similar to the training set, but significant performance degradation often occurs when the shooting environment or target features change. This typical problem is frequently described as the domain generalization problem of deep learning models. However, most current DR datasets only contain images of routine fundus lesions; that is, patients do not have other obvious visual diseases besides DR, and the images taken for examination show no other parallax. In reality, patients often suffer from other ophthalmic diseases of varying severity, such as myopia, presbyopia, astigmatism, glaucoma, and cataracts. These ophthalmic diseases affect the image quality and alter the recognition features of the dataset. Furthermore, due to limited patient resources, there is a lack of large-scale datasets with single, other target parallaxes. Therefore, the proposed deep learning classification model for diabetic retinopathy has not been trained for other parallaxes. When faced with fundus images of other parallaxes, it will experience a significant performance drop, affecting the reliability of the model in actual deployment. Furthermore, there is currently a lack of methods to evaluate the model's performance in practical applications.

[0004] In view of this, one or more embodiments of this specification simultaneously relate to a basic structure of a human eye model and fundus camera simulation optical system, a target parallax simulation structure of multiple human eye model and fundus camera combined optical systems, and a method for generating parallax images simulating diabetic retinopathy, in order to solve the problem in the prior art that the performance degradation of deep learning classification models for diabetic retinopathy cannot be systematically and effectively evaluated when faced with parallax caused by environmental changes and other eye diseases in patients. Summary of the Invention

[0005] This invention provides a method for evaluating a deep learning classification model for diabetic retinopathy. The method comprises a combined optical system of a human eye model and a fundus camera, a method for generating simulated parallax images of diabetic retinopathy based on computational optics, and subsequent testing of the generalization performance of the deep learning classification model. The method includes the following steps:

[0006] Step 1: Based on the structural parameters of the human eye model and the fundus camera, establish the basic structure of the combined optical system of the human eye model and the fundus camera;

[0007] Step 2: Based on the target parallax optical characteristics and pathological features of the deep learning classification model for diabetic retinopathy, establish a series of simulated structures for the combined optical system;

[0008] Step 3: Based on the established simulation structure, calculate the point spread function of the entire system, perform convolutional image processing on a certain proportion of the original dataset, and obtain the simulated disparity image of the original dataset on the target disparity.

[0009] Step 4: Input the hybrid dataset consisting of the original dataset images and simulated parallax images into the model to be evaluated for deep learning prediction and classification;

[0010] Step 5: Calculate various evaluation metrics for deep learning models based on the classification results;

[0011] Step 6: Based on the calculated metrics, obtain the evaluation results of the generalization performance of the deep learning model on the target disparity.

[0012] The described optical system combining a human eye model and a fundus camera includes a simulated human eye model, a retinal objective lens group, and an imaging lens group. The simulated human eye model, retinal objective lens group, and imaging lens group are arranged along the same optical axis from left to right. The simulated human eye model includes, from left to right, the retinal surface, the posterior surface of the lens, the anterior surface of the lens, the pupillary aperture surface, the posterior surface of the cornea, and the anterior surface of the cornea. The retinal surface bulges to the left, with a radius of curvature greater than or equal to 11.50 mm and less than or equal to 12.50 mm; the posterior surface of the lens bulges to the left, with a radius of curvature greater than or equal to 5.50 mm and less than or equal to 6.50 mm; and the anterior surface of the lens bulges to the right, with a radius of curvature greater than or equal to 9.50 mm and less than or equal to 10.50 mm. The pupillary aperture surface only functions as an aperture and is designed as an ideal plane. The posterior surface of the cornea bulges to the right with a radius of curvature greater than or equal to 6.00 mm and less than or equal to 6.50 mm, and the anterior surface of the cornea bulges to the right with a radius of curvature greater than or equal to 7.50 mm and less than or equal to 8.00 mm.

[0013] The retinal objective lens assembly 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 arranged sequentially from left to right. The first lens surface convexes to the left, with a radius of curvature ranging from greater than or equal to 26.15 mm to less than or equal to 26.65 mm. The second lens surface convexes to the right, with a radius of curvature ranging from greater than or equal to 16.50 mm to less than or equal to 17.50 mm. The third and second lens surfaces are coplanar. The fourth lens surface convexes to the left, with a radius of curvature ranging from greater than or equal to 57.00 mm to less than or equal to 57.50 mm. The fifth lens surface convexes to the left, with a radius of curvature ranging from greater than or equal to 23.00 mm to less than or equal to 23.50 mm. The sixth lens surface convexes to the left, with a radius of curvature ranging from greater than or equal to 54.50 mm to less than or equal to 55.50 mm.

[0014] The imaging lens assembly includes, from left to right, 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; the seventh lens surface convexes to the left, with a radius of curvature greater than or equal to 21.75 mm and less than or equal to 22.50 mm; the eighth lens surface convexes to the right, with a radius of curvature 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 convexes to the left, with a radius of curvature greater than or equal to 96.00 mm and less than or equal to 97.00 mm; the eleventh lens surface... The 12th lens surface bulges to the left, with a radius of curvature ranging from ≥12.20 mm to ≤12.70 mm; the 13th lens surface bulges to the right, with a radius of curvature ranging from ≥35.20 mm to ≤35.70 mm; the 14th lens surface bulges to the left, with a radius of curvature ranging from ≥7.50 mm to ≤8.50 mm; the 15th lens surface bulges to the left, with a radius of curvature ranging from ≥12.30 mm to ≤12.80 mm; the 16th lens surface bulges to the left, with a radius of curvature ranging from ≥18.50 mm to ≤19.00 mm.

[0015] Furthermore, in the simulated human eye model, the thickness range of the independent lens formed by the retinal surface and the posterior surface of the lens is [16.10mm, 17.10mm], the thickness range of the independent lens formed by the posterior surface of the lens and the anterior surface of the lens is [3.50mm, 4.50mm], the thickness range of the independent lens formed by the anterior surface of the lens and the posterior surface of the cornea is [3.20mm, 4.20mm], and the thickness range of the independent lens formed by the posterior surface of the cornea and the anterior surface of the cornea is [0.42mm, 1.42mm].

[0016] Furthermore, the thickness range of the independent lens formed by the first lens surface and the second lens surface is [11.76mm, 12.76mm], the thickness range of the independent lens formed by the third lens surface and the fourth lens surface is [8.45mm, 9.45mm], and the thickness range of the independent lens formed by the fifth lens surface and the sixth lens surface is [14.11mm, 15.11mm].

[0017] Furthermore, the thickness range of the independent lens formed by the seventh and eighth lens surfaces is [5.20mm, 6.20mm], the thickness range of the independent lens formed by the ninth and tenth lens surfaces is [5.25mm, 6.25mm], the thickness range of the independent lens formed by the eleventh and twelfth lens surfaces is [5.48mm, 6.48mm], the thickness range of the independent lens formed by the thirteenth and fourteenth lens surfaces is [4.32mm, 5.32mm], and the thickness range of the independent lens formed by the fifteenth and sixteenth lens surfaces is [5.20mm, 6.20mm].

[0018] The human eye model group includes a vitreous simulated lens composed of the retinal surface and the posterior surface of the lens, a lens simulated lens, and a corneal simulated lens.

[0019] The retinal objective lens group includes one cemented doublet lens and one plano-convex lens.

[0020] The imaging lens group includes a front imaging lens group and a rear imaging lens group. The front imaging lens group is a cemented doublet lens, and the rear imaging lens group includes a cemented doublet lens and a positive lens.

[0021] 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 posterior surface of the lens is VITREOUS; the glass material of the independent lens composed of the posterior surface of the lens and the anterior surface of the lens is LENS; the glass material of the independent lens composed of the anterior surface of the lens and the posterior surface of the cornea is AQUEOUS; and the glass material of the independent lens composed of the posterior surface of the cornea and the anterior surface of the cornea is CORNEA. The glass material of the independent lens composed of the first and second lens surfaces is H-LAK2; the glass material of the independent lens composed of the third and fourth lens surfaces is F2; ​​the glass material of the independent lens composed of the fifth and sixth lens surfaces is H-LAK2; the glass material of the independent lens composed of the seventh and eighth lens surfaces is H-LAK3; the glass material of the independent lens composed of the ninth and tenth lens surfaces is ZF6; the glass material of the independent lens composed of the eleventh and twelfth lens surfaces is H-LAK3; the glass material of the independent lens composed of the thirteenth and fourteenth lens surfaces is ZF6; and the glass material of the independent lens composed of the fifteenth and sixteenth lens surfaces is H-LAK3.

[0022] The aforementioned optical system combining the human eye model and fundus camera, with its basic structure, serves as a highly realistic method for fundus examination of diabetic retinopathy. As the physical basis for subsequent parallax image generation based on computational optics for diabetic retinopathy image simulation, it enables effective computational simulation of the fundus examination imaging process for diabetic retinopathy.

[0023] The method for generating parallax images simulating diabetic retinopathy includes the following steps:

[0024] 1) Based on the optical characteristics and pathological mechanisms of the simulated parallax image of the target diabetic retinopathy, corresponding structural modifications are made to the optical system to obtain the target parallax simulation structure of the combined optical system of the human eye model and fundus camera. The modifications include altering the curvature, material, and light transmission parameters of the lens and corneal lens in the human eye model, adding additional optical structures to the human eye model, and making corresponding modifications to the retinal objective lens group and imaging lens group, thereby obtaining a simulated pathological model with the optical characteristics corresponding to the target diabetic retinopathy image.

[0025] 2) Based on the simulated lesion model optical system described in step 1), calculate the point spread function from each point on the object plane of the optical system to the image plane, where the object plane is the retinal plane in the simulated lesion model optical system. 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, which performs ray tracing calculation within the field of view of the aperture stop. The obtained point spread function is a matrix related to the object plane sampling rate and pixel size, containing the propagation matrix information from each point on the object plane to the image plane.

[0026] 3) Based on the obtained point spread function, a simulated disparity image of the target diabetic retinopathy is generated by convolving the source bitmap image file with the point spread function array through image simulation. The methods consider factors including diffraction, aberrations, distortion, relative illumination, image orientation, and polarization present in the optical system of the lesion simulation model. The original source bitmap image file of the diabetic retinopathy should have a clear field of view, well-defined classification, clearly visible lesion classification criteria, and no obvious abnormal aberrations. The generated simulated disparity image possesses the optical characteristics of the target simulated disparity, including diffraction, aberrations, distortion, relative illumination, image orientation, and polarization, reflecting the imaging characteristics and lesion mechanism of the target disparity simulation structure of the combined optical system of the human eye model and fundus camera.

[0027] The generalization performance test of the deep learning classification model is characterized by the following steps:

[0028] 1) Based on the original image of diabetic retinopathy and the generated target image of diabetic retinopathy, simulated disparity images are randomly mixed in a certain proportion to form a mixed dataset for evaluating the target disparity of the diabetic retinopathy model (hereinafter referred to as the mixed dataset). The mixed dataset is characterized by having fundus and lesion features of the original image of diabetic retinopathy, as well as the basic annotations required by the deep learning classification model, namely the five labels corresponding to each level of diabetic retinopathy, and the optical features corresponding to the simulated disparity of the target and the severity of the simulated disparity.

[0029] 2) Based on the aforementioned hybrid dataset, a testing procedure is implemented to assess the generalization ability of the deep learning model for diabetic retinopathy (hereinafter referred to as the model to be evaluated) on the target lesion. In addition to its own deep learning network structure and corresponding data augmentation preprocessing method, the model to be evaluated should also possess weight information trained on its own dataset. This weight information can be obtained through inference training on any training set using arbitrary parameters.

[0030] 3) The testing process includes inputting and reading mixed datasets, classifying and predicting the model to be evaluated, calculating evaluation metrics based on the model classification results, and evaluating the model's generalization performance based on the evaluation metrics.

[0031] Furthermore, the mixed dataset input and reading steps should read the mixed dataset image, the label corresponding to the image, and the simulated parallax type and severity corresponding to the image.

[0032] Furthermore, the classification prediction process of the model to be evaluated, namely, the deep learning model outputs a prediction probability matrix for each input image in five different categories through feature recognition and model inference for each input mixed dataset image.

[0033] Furthermore, the evaluation index calculation process, based on the predicted probability output in the previous step, calculates the accuracy, area under the ROC curve (AUC), and F1 score.

[0034] Furthermore, the accuracy metric is calculated by extracting the maximum value from the predicted 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 accuracy.

[0035] Furthermore, the area under the ROC curve (AUC) is mentioned. The ROC curve, or receiver operating characteristic curve, is an indicator invented by electronics and radar engineers to detect enemy vehicles (aircraft, ships) on the battlefield, belonging to signal detection theory. The horizontal axis of the ROC curve represents the false positive rate (also called the false positive class rate), the probability of a case being classified as positive but not a true positive (i.e., the probability of a true negative case being classified as positive). The vertical axis represents the true positive rate (the probability of a case being classified as positive and also a true positive (i.e., the positive recall rate)). The AUC metric can be used as an indicator of model performance when comparing different classification models. The larger the area under the ROC curve, the better the model classifier's performance and the greater its application value.

[0036] Furthermore, the F1 score is a statistical metric used 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 F1 score is to 1, the higher the model's classification performance.

[0037] Furthermore, the model generalization performance evaluation process based on evaluation metrics involves, after calculating the metrics, statistically analyzing the target disparity and disparity severity of each output metric and the input dataset. This serves as the model's performance at the corresponding target disparity and disparity severity. Based on the changes in performance under different severity levels, the generalization performance of the model under evaluation when facing target disparities is determined. Specifically, a model with high generalization performance, in addition to exhibiting normal classification performance on standard datasets, should maintain high stability across all metrics when facing disparities different from those on standard datasets.

[0038] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0039] 1) The modeling process of this method utilizes optical design software, which allows for targeted analysis and design of the selection of optical components in the optical system. The proposed target parallax simulation structure of the combined optical system of human eye model and fundus camera can realistically simulate the fundus examination process of diabetic retinopathy.

[0040] 2) The image simulation process of this method can generate target diabetic retinopathy image simulation parallax images that are lacking in existing public datasets based on the target parallax simulation structure of the combined optical system of human eye model and fundus camera. This effectively makes up for the deficiencies of existing datasets and has the potential to generate more data.

[0041] 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 on the generated target diabetic retinopathy image simulated parallax image. It makes up for the lack of evaluation methods for medical auxiliary classification deep learning models in the current field of metrology. It can effectively evaluate the performance of the model to be evaluated on standard datasets and multiple target parallax datasets, and provides an effective and reliable basis for the subsequent deployment in the field of medical auxiliary diagnosis. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the optical path of a simulated structure of a human eye model and a fundus camera combined optical system, provided in one embodiment of this specification.

[0043] Figure 2 This is a schematic diagram of the spherical parallax simulation lens 400 in the simulated structure of the human eye model and fundus camera combined optical system provided in three embodiments of this specification.

[0044] Figure 3 This is a schematic diagram of the cylindrical parallax simulation lens 400 in the simulated structure of the human eye model and fundus camera combined optical system provided in three embodiments of this specification.

[0045] Figure 4 This is a schematic diagram of the generalization performance test process for a deep learning classification model for diabetic retinopathy provided in one embodiment of this specification.

[0046] Figure 5 This is a table showing the evaluation results of a deep learning classification model for diabetic retinopathy, provided in one embodiment of this specification.

[0047] Figure 6 This is a graph showing the evaluation results of a deep learning classification model for diabetic retinopathy provided in one embodiment of this specification. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention, and this specification can be implemented in many ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this specification; therefore, this specification is not limited to the specific embodiments disclosed below.

[0049] Figure 1 This is the basic structure of the human eye model and fundus camera simulation optical system in this embodiment of the invention. In this embodiment, the fundus that needs to be detected by the optical system is replaced by the human eye model 100. At the same time, a detector 500 is set on the image plane of the optical system to detect the information of the obtained image.

[0050] like Figure 1 As shown, a basic structure of a human eye model and fundus camera simulation optical system includes a human eye model 100, a retinal objective lens group 200, and an imaging lens group 300. The human eye model 100, the retinal objective lens group 200, and the imaging lens group 300 are arranged coaxially from left to right. The simulated human eye model includes, from left to right, a retinal surface 101, a posterior lens surface 102, an anterior lens surface 103, a pupillary aperture 104, a posterior corneal surface 105, and an anterior corneal surface 106. The retinal surface 101 bulges to the left, with a radius of curvature greater than or equal to 11.50 mm and less than or equal to 12.50 mm; the posterior lens surface 102 bulges to the left, with a radius of curvature greater than or equal to 5.50 mm and less than or equal to 6.50 mm; and the anterior lens surface 103 bulges to the right, with a radius of curvature greater than or equal to 9.50 mm and less than or equal to 10.50 mm. The pupillary surface 104 serves only as an aperture and is designed as an ideal plane. The posterior corneal surface 105 protrudes to the right with a radius of curvature greater than or equal to 6.00 mm and less than or equal to 6.50 mm. The anterior corneal surface 106 protrudes to the right with a radius of curvature greater than or equal to 7.50 mm and less than or equal to 8.00 mm.

[0051] The retinal objective lens assembly 200 includes, from left to right, 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; the first lens surface 201 convexes to the left, and its radius of curvature is greater than or equal to 26.15 mm and less than or equal to 26.65 mm; the second lens surface 202 convexes to the right, and its radius of curvature is greater than or equal to 16.50 mm and less than or equal to 17.50 mm. The third lens surface 203 and the second lens surface 204 are coplanar; the fourth lens surface 204 protrudes to the left, and its radius of curvature is greater than or equal to 57.00 mm and less than or equal to 57.50 mm; the fifth lens surface 205 protrudes to the left, and its radius of curvature is greater than or equal to 23.00 mm and less than or equal to 23.50 mm; the sixth lens surface 206 protrudes to the left, and its radius of curvature is greater than or equal to 54.50 mm and less than or equal to 55.50 mm.

[0052] The imaging lens group 300 includes, from left to right, 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. The seventh lens surface 307 bulges to the left, with a radius of curvature greater than or equal to 21.75 mm and less than or equal to 22.50 mm. The eighth lens surface 308 bulges to the right, with a radius of curvature greater than or equal to 23.50 mm and less than or equal to 24.50 mm. The ninth lens surface 309 is coplanar with the eighth lens surface 308. The tenth lens surface 310 bulges to the left, with a radius of curvature greater than or equal to 96.00 mm and less than or equal to 96.00 mm. 7.00mm; the eleventh lens surface 311 protrudes to the left, with a radius of curvature ranging from greater than or equal to 12.20mm to less than or equal to 12.70mm; the twelfth lens surface 312 protrudes to the right, with a radius of curvature ranging from greater than or equal to 35.20mm to less than or equal to 35.70mm; the thirteenth lens surface 313 and the twelfth lens surface 312 are coplanar; the fourteenth lens surface 314 protrudes to the left, with a radius of curvature ranging from greater than or equal to 7.50mm to less than or equal to 8.50mm; the fifteenth lens surface 315 protrudes to the left, with a radius of curvature ranging from greater than or equal to 12.30mm to less than or equal to 12.80mm; the sixteenth lens surface 316 protrudes to the left, with a radius of curvature ranging from greater than or equal to 18.50mm to less than or equal to 19.00mm.

[0053] In the simulated human eye model, the thickness range of the independent lens formed by the retinal surface 101 and the posterior lens surface 102 is [16.10mm, 17.10mm], the thickness range of the independent lens formed by the posterior lens surface 102 and the anterior lens surface 103 is [3.50mm, 4.50mm], the thickness range of the independent lens formed by the anterior lens surface 103 and the posterior corneal surface 105 is [3.20mm, 4.20mm], and the thickness range of the independent lens formed by the posterior corneal surface 105 and the anterior corneal surface 106 is [0.42mm, 1.42mm].

[0054] The thickness range of the independent lens formed by the first lens surface 201 and the second lens surface 202 is [11.76mm, 12.76mm], the thickness range of the independent lens formed by the third lens surface 203 and the fourth lens surface 204 is [8.45mm, 9.45mm], and the thickness range of the independent lens formed by the fifth lens surface 205 and the sixth lens surface 206 is [14.11mm, 15.11mm].

[0055] The thickness range of the independent lens formed by the seventh lens surface 307 and the eighth lens surface 308 is [5.20mm, 6.20mm], the thickness range of the independent lens formed by the ninth lens surface 309 and the tenth lens surface 310 is [5.25mm, 6.25mm], the thickness range of the independent lens formed by the eleventh lens surface 311 and the twelfth lens surface 312 is [5.48mm, 6.48mm], the thickness range of the independent lens formed by the thirteenth lens surface 313 and the fourteenth lens surface 314 is [4.32mm, 5.32mm], and the thickness range of the independent lens formed by the fifteenth lens surface 315 and the sixteenth lens surface 316 is [5.20mm, 6.20mm].

[0056] The human eye model group includes a vitreous simulated lens composed of the retinal surface and the posterior surface of the lens, a lens simulated lens, and a corneal simulated lens.

[0057] The retinal objective lens group includes one cemented doublet lens and one plano-convex lens.

[0058] The imaging lens group includes a front imaging lens group and a rear imaging lens group. The front imaging lens group is a cemented doublet lens, and the rear imaging lens group includes a cemented doublet lens and a positive lens.

[0059] The basic structure of the optical system of the human eye model, the glass material of the independent lens composed of the retinal surface 101 and the posterior surface of the lens 102 is VITREOUS; the glass material of the independent lens composed of the posterior surface of the lens 102 and the anterior surface of the lens 103 is LENS; the glass material of the independent lens composed of the anterior surface of the lens 103 and the posterior surface of the cornea 105 is AQUEOUS; and the glass material of the independent lens composed of the posterior surface of the cornea 105 and the anterior surface of the cornea 106 is CORNEA. The glass material of the independent lens formed by the first lens surface 201 and the second lens surface 202 is H-LAK2; the glass material of the independent lens formed by the third lens surface 203 and the fourth lens surface 204 is F2; ​​the glass material of the independent lens formed by the fifth lens surface 205 and the sixth lens surface 206 is H-LAK2; the glass material of the independent lens formed by the seventh lens surface 307 and the eighth lens surface 308 is H-LAK3; the glass material of the independent lens formed by the ninth lens surface 309 and the tenth lens surface 310 is ZF6; the glass material of the independent lens formed by the eleventh lens surface 311 and the twelfth lens surface 312 is H-LAK3; the glass material of the independent lens formed by the thirteenth lens surface 313 and the fourteenth lens surface 314 is ZF6; and the glass material of the independent lens formed by the fifteenth lens surface 315 and the sixteenth lens surface 316 is H-LAK3.

[0060] Furthermore, this embodiment simulates the most common basic target parallaxes, namely defocus parallax and astigmatic parallax. Figure 1 The human eye model and fundus camera simulation optical system shown have been augmented with a parallax simulation lens 400. The parallax simulation lens 400 consists of a front surface and a rear surface of the parallax simulation lens, and its glass material is H-LAK2.

[0061] Furthermore, the parallax simulation lens has different designs for different degrees of parallax simulation severity. In this embodiment, it is designed to handle defocus parallax. Figure 2 The three simulated structures shown are for cases with astigmatic parallax. Figure 3 The three simulation structures are shown.

[0062] The first type of simulation structure is as follows: Figure 2 As shown in (1). The parallax simulation lens rear surface elevation diagram is shown in... Figure 2 As shown in (1)(a), the standard spherical surface convexes to the left, with a radius of curvature of 2760 mm. The sagittal diagram of the front surface of the parallax simulation lens is shown below. Figure 2 As shown in (1)(b), it is a spherical standard surface that convexes to the right with a radius of curvature of 2760mm. The parallax simulation lens composed of it provides 0.5D defocus diopter for the parallax simulation optical system.

[0063] The second simulation structure is as follows: Figure 2 (2) is shown. The parallax simulation lens rear surface elevation diagram is shown below. Figure 2 As shown in (2)(a), the standard spherical surface convexes to the left, with a radius of curvature of 1380 mm. Figure 2 As shown in (2)(b), it is a spherical standard surface that convexes to the right with a radius of curvature of 1380 mm. The parallax simulation lens composed of it provides 1.0 D of defocus refractive power for the parallax simulation optical system.

[0064] The third simulation structure is as follows: Figure 2 (3) is shown. The parallax simulation lens rear surface elevation diagram is shown below. Figure 2 As shown in (3)(a), the standard spherical surface convexes to the left, with a radius of curvature of 920 mm. The sagittal diagram of the front surface of the parallax simulation lens is shown below. Figure 2 As shown in (3)(b), it is a spherical standard surface that bulges to the right with a radius of curvature of 920 mm. The parallax simulation lens composed of it provides 1.5D defocus diopter for the parallax simulation optical system.

[0065] The fourth simulation structure is as follows: Figure 3 As shown in (1). The parallax simulation lens rear surface elevation diagram is shown in... Figure 3As shown in (1)(a), the standard cylindrical surface convexes to the left, with a radius of curvature of 2760 mm. Figure 3 As shown in (1)(b), the cylindrical standard surface convexes to the right with a radius of curvature of 2760 mm. The parallax simulation lens composed of it provides 0.5D astigmatic refractive power for the parallax simulation optical system.

[0066] The fifth simulation structure is as follows Figure 3 (2) is shown. The parallax simulation lens rear surface elevation diagram is shown below. Figure 3 As shown in (2)(a), the standard cylindrical surface convexes to the left, with a radius of curvature of 1380 mm. The sagittal diagram of the front surface of the parallax simulation lens is shown below. Figure 3 As shown in (2)(b), it is a cylindrical standard surface that bulges to the right with a radius of curvature of 1380 mm. The parallax simulation lens composed of it provides 1.0 D of astigmatic refractive power for the parallax simulation optical system.

[0067] The sixth simulation structure is as follows Figure 3 (3) is shown. The parallax simulation lens rear surface elevation diagram is shown below. Figure 3 As shown in (3)(a), the standard cylindrical surface convexes to the left, with a radius of curvature of 920 mm. The sagittal diagram of the front surface of the parallax simulation lens is shown below. Figure 3 As shown in (3)(b), the cylindrical standard surface convexes to the right with a radius of curvature of 920 mm. The parallax simulation lens composed of it provides 1.5D astigmatic refractive power for the parallax simulation optical system.

[0068] Based on the six simulation structures described above, the point spread function (PSF) from each point on the object plane of the optical system to the image plane is calculated. The object plane is the retinal plane in the simulation structure. The calculated field of view for 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. Rays within the aperture stop's field of view are traced and calculated. The resulting PSF is a matrix related to the object plane sampling rate and pixel size, containing the propagation matrix information from each point on the object plane to the image plane. In this example, the sampling resolution of the PSF on the object plane is 30 micrometers, and the sampling resolution on the image plane is 4 micrometers. That is, for a retinal image with a 40° field of view on the object plane, the diffusion of each point source on the image plane with a 4-micrometer sampling rate is calculated using a 30-micrometer sampling rate. This meets the accuracy requirements for diabetic retinopathy examination.

[0069] Furthermore, the point spread function of the central point light source in the three defocus simulation structures is as follows: Figure 2 As shown in (1-3)(c), it can be seen that under defocus diopter of (1) 0.5D (2) 1.0D (3) 1.5D, the diffusion of the point light source is manifested as a gradually increasing defocus spot. The diffusion of the point light source overlaps with each other under convolution, thus affecting the final image quality.

[0070] Furthermore, the point spread function of the central point source of the three astigmatism simulation structures is as follows: Figure 3 As shown in (1-3)(c), it can be seen that under astigmatic diopter of (1) 0.5D (2) 1.0D (3) 1.5D, the diffusion of the point light source is manifested as a gradually increasing astigmatic spot. Unlike defocus, the point light source has the greatest impact on the astigmatic axis and the least impact in the direction perpendicular to the astigmatic axis. The diffusion patterns overlap under convolution, thus affecting the final image quality.

[0071] Based on the obtained point spread function, a simulated parallax image of the target diabetic retinopathy image is generated by convolving the source bitmap image file with the point spread function array through image simulation. The methods consider factors including diffraction, aberrations, distortion, relative illumination, image orientation, and polarization present in the simulated structure. The original source bitmap image file of diabetic retinopathy should have a clear field of view, well-defined classification, clearly visible lesion classification criteria, and no obvious abnormal aberrations. The generated simulated parallax image possesses the optical characteristics of the target simulated parallax, including diffraction, aberrations, distortion, relative illumination, image orientation, and polarization, reflecting the imaging characteristics and pathological mechanisms of the simulated structure.

[0072] The generalization performance testing process of the deep learning classification model in this embodiment is as follows: Figure 4 As shown, it includes the following steps:

[0073] 1) Based on the original diabetic retinopathy dataset 10 and the combined model simulation structure 20 including a human eye model 21, a disparity lens 22, and a camera model 23, simulated disparity images of the target diabetic retinopathy image are generated and randomly mixed in a certain proportion to form a mixed dataset 30 (hereinafter referred to as the mixed dataset) for testing the target disparity evaluation of the diabetic retinopathy model. The mixed dataset is characterized by having fundus and lesion features of the original diabetic retinopathy image, as well as the basic annotations required by the deep learning classification model, namely, 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 simulated disparity. In this embodiment, six sets of public datasets were selected, including APTOS, DDR, DEEPDR, FGADR, IDRID, and Messidor. In each set, 20% of the images are used as the original diabetic retinopathy images. The generated simulated disparity images, based on the diabetic retinopathy features of the original images, also have different degrees of defocus refractive power and astigmatic refractive power applied by the six structures. Each type of target parallax is labeled with a diopter of 0.5D, 1.0D, and 1.5D as the severity of the target parallax.

[0074] 2) Based on the aforementioned mixed dataset, a testing process is conducted to evaluate the generalization ability of the deep learning model 40 for diabetic retinopathy (hereinafter referred to as the model to be evaluated) on the target lesion. In addition to retaining its own deep learning network structure and corresponding data augmentation preprocessing method, the model to be evaluated should also have weight information trained on its own dataset. This weight information can be obtained by training any parameters on any training set. In this embodiment, to verify the reliability of the method in the application scenario, four advanced publicly available diabetic retinopathy classification models were trained on the publicly available dataset. These publicly available models include GDR-net, mixup, DeepDR, and PCM. The training method maintains the network structure and pre-training method provided by the publicly available models. Classification training was performed on the training set of the aforementioned publicly available dataset using the same standard. The classification training followed the training method provided by the model, inputting identical original images and corresponding labels, and training with the same batch size and epoch number. In this embodiment, the batch size was 32 and the epoch number was 30. After training, all models achieved good classification performance on the standard test set.

[0075] 3) The test process includes inputting and reading a mixed dataset 30, classifying and predicting the model's classification results 50, calculating the evaluation metrics 60 based on the model's classification results, and evaluating the model's generalization performance 70 based on the evaluation metrics.

[0076] Furthermore, the mixed dataset input and reading steps involve inputting and reading the mixed dataset image 30, the corresponding label of the image, and the simulated parallax type and severity of the image. In this embodiment, each image in the mixed dataset, in addition to its lesion category label, also has a simulated parallax type such as: no parallax, defocus, astigmatic lesion. The simulated parallax severity is such as: 0.5D, 1.0D, 1.5D.

[0077] Furthermore, the model's classification prediction process involves, for each input mixed dataset image, using feature recognition and model inference, outputting a prediction probability matrix for each input image across five different categories. In this embodiment, the model outputs corresponding prediction probabilities for each of the 32 input images across five categories in each batch, with a matrix size of [32, 5]. In each test, all classification prediction results 50 for the input images are saved for subsequent processing.

[0078] Furthermore, the evaluation index calculation process, based on the prediction result 50 output in the previous step, performs evaluation index 60, including the calculation of accuracy, area under the ROC curve (AUC), and F1 score.

[0079] Furthermore, the accuracy metric is calculated by extracting the maximum value from the predicted probability matrix output by the model to obtain the prediction result, and taking the ratio of the prediction result matrix to the label matrix as the accuracy.

[0080] Furthermore, the AUC metric is defined as the area under the ROC curve and the coordinate axis. The horizontal axis of the ROC curve represents the false positive rate, the probability of a case being classified as a positive example but not a true example (i.e., the probability of a true negative example being classified as a positive example). The vertical axis represents the true positive rate, the probability of a case being classified as a positive example and also a true example (i.e., the probability of a true example being classified as a positive example). The AUC metric can be used as an indicator of the performance of different classification models. The larger the area under the ROC curve, the better the model classifier's performance and the greater its application value.

[0081] Furthermore, the F1 score is a statistical metric used 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 F1 score is to 1, the higher the model's classification performance.

[0082] Furthermore, the model generalization performance evaluation process based on evaluation metrics, after calculating the metrics 60, statistically analyzes the target disparity and disparity severity of each output metric and the input dataset, using this as the model's performance at the corresponding target disparity and disparity severity. Based on the changes in performance under different severity levels, the generalization performance evaluation result 70 of the model under evaluation when facing target disparities is obtained. Specifically, a model with high generalization performance, in addition to exhibiting normal classification performance on standard datasets, should maintain high stability in each metric when facing disparities different from standard datasets, reflecting its corresponding credibility in real-world applications. In this embodiment, the performance of the four selected classification models GDR-net, mixup, DeepDR, and PCM on six mixed datasets APTOS, DDR, DEEPDR, FGADR, IDRID, and Messidor is as follows: Figure 5 As shown, it can be seen that the proposed metrics of all four models exhibit an overall performance decline trend as the defocus level increases. The following section, as an example, illustrates the performance of the DeepDR model on the FGADR dataset. Figure 5 The underlined data will be analyzed in detail.

[0083] Specifically, due to Figure 5All metrics presented are capped at 1, so the unit is expressed as (%). Values ​​closer to 100 indicate higher model performance. More specifically, ACC, as accuracy, reflects the direct accuracy rate of the model in classification. When faced with different degrees of defocus in the FGADR dataset, the DeepDR model's accuracy dropped from an initial 88.5% to 56.0%, a decrease of 10.8% per 0.5D on average, significantly impacting its diagnostic performance.

[0084] Furthermore, the AUC metric, as the area under the ROC curve, shows the prediction results under different confidence thresholds. The higher the value, the higher the model's classification performance. As shown in the figure, the AUC metric of the DeepDR model decreased from 99.0 to 86.9 when facing the FGADR dataset, with an average decrease of 4.03 per 0.5D, indicating that the receiver operating characteristic curve of the model was significantly affected.

[0085] Furthermore, the F1 score metric is calculated as the harmonic mean of precision and recall, taking into account both the precision and recall of the classification model, and jointly reflecting the model's predictive accuracy and case detection coverage. When faced with different degrees of defocus in the FGADR dataset, the DeepDR model's F1 score dropped from 88.4 to 52.3, a decrease of 12.03 per 0.5D on average. This indicates a decline in the model's overall performance in terms of predictive accuracy and case detection coverage.

[0086] Furthermore, to present the test results more intuitively, Figure 5 The result is expressed in curve form. Figure 6 , Figure 6 The three images represent the accuracy of the results, the AUC curve, and the F1 score as the severity of target disparity decreases. Different curves in the images represent different classification models. As can be seen in the figures, the performance of the classification model significantly decreases under different levels of defocus severity, indicating that the generated mixed dataset can effectively simulate target disparity and has the function of evaluating the model's performance on corresponding disparities.

Claims

1. An evaluation method for a deep learning classification model of diabetic retinopathy, characterized in that, The method includes a human eye model and a fundus camera combined optical system, a method for generating simulated parallax images of diabetic retinopathy based on computational optics, and a generalization performance test of a deep learning classification model. Specifically, it includes the following steps: Step 1: Establish the basic structure of the optical system combining the human eye model and fundus camera; the basic structure includes a simulated human eye model, a retinal objective lens group, and an imaging lens group; the simulated human eye model includes a vitreous simulated lens composed of the retinal surface and the posterior surface of the lens, a simulated lens of the lens, and a simulated corneal lens; the retinal objective lens group includes a cemented doublet lens and a plano-convex lens, and the imaging lens group includes an anterior imaging lens group and a posterior imaging lens group; Step 2: Based on the target parallax optical characteristics and pathological features of the deep learning classification model for diabetic retinopathy, the curvature parameters, material parameters and light transmission characteristics of the lens and corneal simulated lenses in the human eye model are modified accordingly. A parallax simulated lens optical structure is added to the structure, and the parameters of the retinal objective lens group and imaging lens group are modified to obtain a combined optical system simulation structure with the optical features corresponding to the target diabetic retinopathy image. Step 3: Based on the established simulation structure, using the retinal surface of the human eye model as the object surface, calculate the point spread function of all object points through the entire optical system with a field of view of not less than 40°. Based on the calculated point spread function, perform convolution image processing on a certain proportion of the original dataset to obtain the simulated parallax image of the original dataset on the target parallax. Step 4: Input the mixed dataset consisting of the original dataset images and simulated parallax images into the deep learning classification model to be evaluated for deep learning prediction and classification; Step 5: Calculate the evaluation metrics for the deep learning model based on the classification results; Step 6: Based on the calculated metrics, obtain the evaluation results of the generalization performance of the deep learning model on target disparity.

2. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 1, characterized in that, The simulated human eye model, retinal objective lens group, and imaging lens group are arranged along the same optical axis from left to right: The simulated human eye model includes, from left to right, the retinal surface, the posterior surface of the lens, the anterior surface of the lens, the pupillary aperture, the posterior surface of the cornea, and the anterior surface of the cornea; The retinal surface bulges to the left, with a radius of curvature greater than or equal to 11.50 mm and less than or equal to 12.50 mm; the posterior surface of the lens bulges to the left, with a radius of curvature greater than or equal to 5.50 mm and less than or equal to 6.50 mm; the anterior surface of the lens bulges to the right, with a radius of curvature greater than or equal to 9.50 mm and less than or equal to 10.50 mm; the pupillary aperture surface only functions as an aperture and is set as an ideal plane; the posterior surface of the cornea bulges to the right, with a radius of curvature greater than or equal to 6.00 mm and less than or equal to 6.50 mm; the anterior surface of the cornea bulges to the right, with a radius of curvature greater than or equal to 7.50 mm and less than or equal to 8.00 mm; The retinal objective lens assembly 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 arranged sequentially from left to right. The first lens surface convexes to the left, with a radius of curvature ranging from greater than or equal to 26.15 mm to less than or equal to 26.65 mm. The second lens surface convexes to the right, with a radius of curvature ranging from greater than or equal to 16.50 mm to less than or equal to 17.50 mm. The third and second lens surfaces are coplanar. The fourth lens surface convexes to the left, with a radius of curvature ranging from greater than or equal to 57.00 mm to less than or equal to 57.50 mm. The fifth lens surface convexes to the left, with a radius of curvature ranging from greater than or equal to 23.00 mm to less than or equal to 23.50 mm. The sixth lens surface bulges to the left, and its radius of curvature is greater than or equal to 54.50 mm and less than or equal to 55.50 mm. 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 sequentially from left to right; the seventh lens surface protrudes to the left, and its radius of curvature is greater than or equal to 21.75 mm and less than or equal to 22.50 mm. The eighth lens surface bulges to the right, and its radius of curvature 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 bulges to the left, and its radius of curvature is greater than or equal to 96.00 mm and less than or equal to 97.00 mm; the eleventh lens surface bulges to the left, and its radius of curvature is greater than or equal to 12.20 mm and less than or equal to 12.70 mm. The twelfth lens surface bulges to the right, and its radius of curvature ranges from greater than or equal to 35.20 mm to less than or equal to 35.70 mm. The thirteenth lens surface and the twelfth lens surface are coplanar; the fourteenth lens surface protrudes to the left, and its radius of curvature is greater than or equal to 7.50 mm and less than or equal to 8.50 mm. The fifteenth lens surface bulges to the left, and its radius of curvature is greater than or equal to 12.30 mm and less than or equal to 12.80 mm. The sixteenth lens surface bulges to the left, and its radius of curvature ranges from greater than or equal to 18.50 mm to less than or equal to 19.00 mm.

3. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 2, characterized in that: The thickness range of the independent lens formed by the first lens surface and the second lens surface is [11.50mm, 12.50mm], the thickness range of the independent lens formed by the third lens surface and the fourth lens surface is [8.20mm, 9.20mm], and the thickness range of the independent lens formed by the fifth lens surface and the sixth lens surface is [14.10mm, 15.20mm].

4. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 2, characterized in that: The thickness range of the independent lens formed by the seventh and eighth lens surfaces is [5.20mm, 6.20mm], the thickness range of the independent lens formed by the ninth and tenth lens surfaces is [5.20mm, 6.20mm], the thickness range of the independent lens formed by the eleventh and twelfth lens surfaces is [5.15mm, 6.15mm], the thickness range of the independent lens formed by the thirteenth and fourteenth lens surfaces is [4.25mm, 5.25mm], and the thickness range of the independent lens formed by the fifteenth and sixteenth lens surfaces is [5.25mm, 6.25mm].

5. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 2, characterized in that: The glass material of the independent lens composed of the retinal surface and the posterior surface of the lens is VITREOUS; the glass material of the independent lens composed of the posterior surface of the lens and the anterior surface of the lens is LENS; the glass material of the independent lens composed of the anterior surface of the lens and the posterior surface of the cornea is AQUEOUS; and the glass material of the independent lens composed of the anterior surface of the cornea and the posterior surface of the cornea is CORNEA.

6. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 5, characterized in that: The glass material of the independent lens composed of the first and second lens surfaces is H-LAK2; the glass material of the independent lens composed of the third and fourth lens surfaces is F2; ​​the glass material of the independent lens composed of the fifth and sixth lens surfaces is H-LAK2; the glass material of the independent lens composed of the seventh and eighth lens surfaces is H-LAK3; the glass material of the independent lens composed of the ninth and tenth lens surfaces is ZF6; the glass material of the independent lens composed of the eleventh and twelfth lens surfaces is H-LAK3; the glass material of the independent lens composed of the thirteenth and fourteenth lens surfaces is ZF6; and the glass material of the independent lens composed of the fifteenth and sixteenth lens surfaces is H-LAK3.

7. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 1, characterized in that, The method for generating parallax images simulating diabetic retinopathy based on computational optics includes the following steps: 1) Based on the basic structure of the optical system combining the human eye model and the fundus camera, and according to the optical characteristics and pathological mechanisms of the simulated parallax image of the target diabetic retinopathy, the optical system is modified accordingly to obtain the target parallax simulation structure of the optical system combining the human eye model and the fundus camera. The modification of the simulation structure includes changing the curvature parameters, material parameters and light transmission characteristics of the lens and corneal lens in the human eye model, adding a parallax simulation lens optical structure to the structure, and changing the parameters of the retinal objective lens group and the imaging lens group. 2) Based on the target parallax simulation structure described in step 1), the point spread function from each point on the object surface of the optical system to the image surface is calculated. The object surface is the retinal surface in the simulation structure, and the calculated field of view of the object surface should be no less than 40°. The point spread function is calculated based on the ray tracing method of the optical simulation software. Using different object points as initial points, the rays within the field of view of the aperture stop are traced and calculated. The obtained point spread function is a matrix related to the object surface sampling rate and pixel size, which contains the propagation matrix information from each point on the object surface to the image surface. 3) Based on the obtained point spread function, a simulated parallax image of the target diabetic retinopathy image is generated by convolving the source bitmap image file with the point spread function array. The effects considered include diffraction, aberrations, distortion, relative illumination, image orientation, and polarization present in the simulated structure. The source bitmap image file of the diabetic retinopathy has a clear field of view, clear classification, and the lesion classification criteria are clearly visible, with no abnormal aberrations. The generated simulated parallax image has the diffraction, aberration, distortion, relative illumination, image orientation, and polarization optical characteristics of the target simulated parallax.

8. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 1, characterized in that, The generalization performance test of the deep learning classification model includes the following steps: 1) Based on the original image of diabetic retinopathy and the generated target image of diabetic retinopathy, simulated disparity images are randomly mixed in a certain proportion to form a mixed dataset for evaluating the target disparity of the diabetic retinopathy model for testing; the mixed dataset is characterized by having fundus and lesion features of the original image of diabetic retinopathy, as well as the basic annotations required by the deep learning classification model, namely the five labels corresponding to the five levels of diabetic retinopathy. Furthermore, it should have the optical features corresponding to the target simulated disparity corresponding to the simulated structure and the severity of the corresponding simulated disparity. 2) Based on the aforementioned hybrid dataset, conduct a test procedure to evaluate the generalization ability of the deep learning model for diabetic retinopathy to the target disparity. 3) The testing process includes inputting and reading mixed datasets, classifying and predicting the model to be evaluated, calculating evaluation metrics based on the model classification results, and evaluating the model's generalization performance based on the evaluation metrics.

9. The evaluation method for a deep learning classification model of diabetic retinopathy according to claim 8, characterized in that, The mixed dataset input and reading steps include inputting and reading mixed dataset images, images corresponding to labels, and images corresponding to simulated disparity types and severity. The classification prediction process of the model to be evaluated is that the deep learning model outputs a prediction probability matrix for each input image in five different categories through feature recognition and model inference for each input mixed dataset image. The evaluation index calculation process is based on the predicted probability output in the previous step, and calculates the accuracy, area under the ROC curve (AUC), and F1 score. The accuracy metric is calculated by extracting the maximum value from the predicted probability matrix output by the model to obtain the prediction result, and taking the ratio of the same prediction result matrix to the label matrix as the accuracy. The F1 score mentioned above is a statistical indicator used to measure the accuracy of a classification model. It is defined as the harmonic mean of precision and recall, which takes into account both the precision and recall of the classification model. The closer the value is to 1, the higher the model's classification performance. The model generalization performance evaluation process based on evaluation metrics involves, after calculating the metrics, statistically analyzing the target disparity and disparity severity of each output metric and the input dataset. This serves as the model's performance at the corresponding target disparity and disparity severity. Based on the changes in performance at different severity levels, the generalization performance of the model under evaluation is determined when facing target disparities. Specifically, a model with high generalization performance, in addition to exhibiting normal classification performance on standard datasets, should maintain high stability across all metrics when facing disparities different from standard datasets, reflecting its corresponding credibility in real-world applications.

Images