Dry eye diagnosis and treatment integrated system based on deep neural network

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a dry eye diagnosis and treatment integrated system based on a deep neural network, which comprises an input unit module, a diagnosis unit module, a data matching module and an output unit module, and is characterized in that dry eye symptoms and OSDI scale result characteristics of a patient to be diagnosed are input through the input unit module; the diagnosis unit module obtains the dry eye symptom, the OSDI scale and the corresponding relation between the lower tear river height and the mild dryness, moderate dryness and severe dryness of a patient, the data matching module matches the result features of the dry eye symptom, the OSDI scale and the psychological survey scale of the patient to be diagnosed into the corresponding relation, and a corresponding matching diagnosis report is generated. And finally, outputting grading diagnosis of light, moderate and heavy dry eyes of the patient to be diagnosed through an output unit module. The system can help clinicians to improve the diagnosis efficiency of dry eye grading and reduce missed diagnosis and misdiagnosis, and is high in operability.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a comprehensive dry eye diagnosis and treatment system based on deep neural network. Background Art

[0002] Dry eye is one of the most common eye diseases in the world. The symptoms of dry eye mainly include eye irritation, photophobia and visual fluctuations, which can lead to eye pain and may cause permanent damage to the cornea. The prevalence rate is 5%-50%, and the prevalence rate is affected by race and environmental factors. Epidemiological studies have shown that dry eye is more common in women, and its prevalence increases with age. At present, people's work and lifestyle have changed, and the dependence on electronic products, the widespread use of corneal contact lenses and the increase in the prevalence of sleep disorders have led to an increasing trend in the incidence of dry eye in people of different ages. Other related risk factors include diabetes and air pollution. Dry eye may also affect the quality of life of patients and reduce work efficiency, resulting in significant direct or indirect public health costs and increased personal economic burden. Only by diagnosing dry eye quickly and accurately and conducting precise treatment can patients be provided with a higher quality of life and work efficiency.

[0003] According to the 2020 "Chinese Dry Eye Expert Consensus: Definition and Classification (2020)", dry eye refers to a type of disease caused by tear film instability and (or) damage to the ocular surface due to abnormal tear quantity or quality, which leads to ocular discomfort symptoms. Diagnostic criteria for dry eye: 1. The patient complains of one of the subjective symptoms such as dry eyes, foreign body sensation, burning sensation, fatigue, discomfort, red eyes, and visual fluctuations. The Chinese Dry Eye Questionnaire Scale is ≥7 points or the Ocular Surface Disease Index (OSDI) is ≥13 points; at the same time, the patient's tear film stability, tear secretion, ocular surface cell staining, ophthalmic imaging examination, laboratory auxiliary examination, eyelid margin and meibomian gland examination and systemic examination need to be tested. In summary, the current dry eye detection items are complicated and too professional. In addition, professional equipment is required to check dry eyes, which is expensive and has not been popularized. For some specialists and grassroots doctors, it lacks certain operability, which is easy to cause missed diagnosis and misdiagnosis. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a comprehensive dry eye diagnosis and treatment system based on deep neural network. The present invention aims to solve the problem that the existing dry eye detection process has complicated detection items, poor operability, and is prone to missed diagnosis and misdiagnosis.

[0005] To achieve the above-mentioned object, the present invention provides a comprehensive system for dry eye diagnosis and treatment based on deep neural network, comprising an input unit module, a diagnosis unit module, a data matching module and an output unit module; The input unit module is used to input the OSDI scale and symptom result characteristics of the patient to be diagnosed; the diagnosis unit module is used to train the initial dry eye grading diagnosis neural network model based on a large number of clinical history data of dry eye patients, so as to obtain the corresponding relationship between the patient's dry eye OSDI scale, symptoms and lower tear meniscus height examination result characteristics and the clinical diagnosis grading of dry eye; The diagnosis unit module includes a model training subunit module; The data matching module is used to match the dry eye diagnosis result characteristics of the patient to be diagnosed to the above corresponding relationship and generate a corresponding matching diagnosis report; The output unit module is used to output the clinical grade of dry eye of the patient to be diagnosed.

[0006] Furthermore, the model training subunit module includes a data input module, an initial diagnosis module and a diagnosis typing module; The data input module is used to obtain a large amount of historical clinical data of dry eye patients, set data preprocessing rules, and process the data information according to the data preprocessing rules to obtain a diagnostic data set; The initial diagnosis module is used to randomly divide the diagnosis data set processed by the data input module and the image data set collected by the mobile phone into a training set, a validation set and a test set, and then input the data of the training set and the validation set into the initial dry eye grading diagnosis neural network model for model training, and finally obtain the dry eye grading diagnosis neural network model; The diagnostic classification module is used to obtain the corresponding relationship between the OSDI scale, symptoms and lower tear meniscus height result characteristics of dry eye patients and mild dry eyes, moderate dry eyes and severe dry eyes.

[0007] Further, the historical clinical data of dry eye patients acquired by the data input module include OSDI scale, symptom result characteristics and lower tear meniscus height result characteristic data; Among them, the OSDI scale refers to the dry eye questionnaire; the symptom result characteristics include the degree or frequency of dry eyes, foreign body sensation, burning sensation, fatigue, discomfort, red eyes, and vision fluctuations.

[0008] Furthermore, the data preprocessing rules include data type conversion, data filling and data deletion; Data type conversion includes binary data conversion and multi-value data conversion; Data filling is to fill the empty value fields of each item with mean value; Data deletion is to delete images where the eyelashes block the tear canal over a large area or where the pupil is not imaged.

[0009] Furthermore, the diagnosis process of the diagnosis unit module is: S1. Obtain ocular surface images of dry eye patients as historical data sets, and preprocess the historical data of dry eye patients; S2. Positioning and labeling the pupil center and tear meniscus region of the ocular surface image of the dataset preprocessed in step S1; S3. Calculate the lower tear meniscus height according to the annotation results of step S2, and grade the severity of dry eye symptoms; S4. Construct a neural network model for dry eye grading diagnosis; S5. Input the ocular surface image data of the patient to be diagnosed into the model of step S4 to obtain the dry eye grading diagnosis result.

[0010] Furthermore, in step S2, the ocular surface image is located based on the Fast R-CNN model, the region of interest, i.e., the tear meniscus and the corresponding position of the pupil, is obtained, and the predicted candidate box is output; the pupil center of the ocular surface image is located and marked according to two easily identifiable biological features of the pupil, as follows: The pupil is roughly located on the ocular surface image. The difference in grayscale values ​​between the pupil area and other areas of the eye is used in combination with binarization technology to determine the area with high grayscale values ​​including the pupil. The pupil color is the position with the lowest grayscale value of the entire eye. The image is processed using morphological operators including opening and closing operations, dilation and erosion to remove unnecessary parts. The geometric features of the pupil are used to determine the size and center position of the pupil area using the Hough circle transform. The center of the circle is the pupil center.

[0011] Furthermore, in step S2, the tear meniscus area is segmented by building a semantic segmentation network model of the tear meniscus area, specifically as follows: the tear meniscus area is manually marked on the ocular surface image in the historical data set of step S1, and divided into a training set and a verification set, the training set is input into the semantic segmentation network for training, and the test set is input into the trained segmentation network for verification testing, and finally a semantic segmentation network model of the tear meniscus area is obtained.

[0012] Further, in step S3, the height of the tear mane is calculated by subtracting the distance from the center of the pupil to the upper edge of the tear mane from the distance from the center of the pupil to the lower edge of the tear mane in the six-point direction of the center of the pupil, and the difference is the height of the tear mane; The calculation expression is: In the formula, It is the height of the River of Tears; N Indicates that the tear river is taken near the six o'clock position of the pupil center. N The pixels are averaged; It represents the distance from the pupil center to the lower edge of the tear meniscus; Indicates the distance from the center of the pupil to the upper edge of the tear meniscus; Dry eye symptom severity grading: The tear meniscus height is used as a grading indicator, specifically divided into normal, mild dry eyes, moderate dry eyes, and severe dry eyes; The grading range is as follows: tear meniscus height > 0.2 mm is the normal range; tear meniscus height ≤ 0.2 mm, the patient has mild dry eyes; tear meniscus height ≤ 0.1 mm, the patient has moderate dry eyes; when the tear meniscus height cannot be measured, the patient has severe dry eyes.

[0013] Furthermore, in step S4, the dry eye grading diagnosis neural network model is as follows: In the formula, R represents the marked area; S A text description representing the marked area; I Represents the entire image; θ Represents the parameters of the convolutional neural network model and the recursive neural network model. Further, the data input by the input unit module also includes the patient's psychological survey scale. Beneficial Effects

[0014] 1. The present invention provides a comprehensive dry eye diagnosis and treatment system based on a deep neural network, including an input unit module, a diagnosis unit module, a data matching module and an output unit module. The dry eye symptoms of the patient to be diagnosed and the OSDI scale result characteristics are input through the input unit module, and the corresponding relationship between the patient's dry eye symptoms, OSDI scale and lower tear meniscus height and mild dry, moderate and severe dryness is obtained through the diagnosis unit module. The dry eye symptoms of the patient to be diagnosed and the OSDI scale result characteristics are matched to the above-mentioned corresponding relationship through the data matching module, and a corresponding matching diagnosis report is generated. Finally, the output unit module outputs the graded diagnosis of mild, moderate and severe dry eyes of the patient to be diagnosed. The use of the invented system can help clinicians improve the diagnostic efficiency of dry eye grading, reduce the occurrence of missed diagnosis and misdiagnosis, and has strong operability. The present invention also has a psychological survey scale, which pays attention to dry eye diseases caused by psychological factors. Screening of the eyes and psychology at the same time, and giving corresponding guidance and treatment.

[0015] 2. The comprehensive dry eye diagnosis and treatment system based on deep neural network of the present invention can improve the screening efficiency of dry eye, save diagnosis and treatment time and cost, enable the subjects to actively and timely cooperate with doctors to formulate and adjust the diagnosis and treatment plan without time and geographical restrictions, improve bad work and life habits, and reduce the ocular surface damage caused by dry eyes, which leads to reduced learning and work efficiency, insomnia, depression and other adverse psychological problems, so as to prevent social harm. In the clinical aspect of ophthalmology: The system of the present invention greatly reduces the workload of ophthalmologists, reduces the expenditure of medical insurance funds, and helps to alleviate the difficulties caused by the relative shortage, uneven distribution and different levels of medical resources, as well as the irreversible vision loss caused by insufficient knowledge of dry eyes and low economic income.

[0016] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a module diagram of a comprehensive system for dry eye diagnosis and treatment based on a deep neural network according to the present invention; Figure 2 It is the workflow diagram of the diagnosis unit module; Figure 3 This is the overall flow chart of semantic segmentation and classification of dry eye syndrome; Figure 4 This is the specific structure of the tear river segmentation network. DETAILED DESCRIPTION

[0018] In order to make the technical solutions, advantages and purposes of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of this application.

[0019] like Figure 1~Figure 3 As shown, the present invention provides a comprehensive dry eye diagnosis and treatment system based on a deep neural network, including an input unit module, a diagnosis unit module, a data matching module and an output unit module; The input unit module is used to input the OSDI scale and symptom result characteristics of the patient to be diagnosed; This embodiment uses the CSO Antares ocular surface comprehensive analyzer produced by Shenzhen Keyikang Electronic Instrument Co., Ltd. as the data acquisition device. During the ocular surface analyzer acquisition process, the examinee needs to sit in front of the operating table, fix the mandible and forehead, and select the anterior segment long lens single-line scanning mode. The patient needs to look at the fixation light in the device and scan at the 6 o'clock position below the cornea. The tear meniscus is placed in the center of the scanning window by adjusting the lens position, and the OCT lens is fine-tuned to make the tear meniscus cross-section image clearest.

[0020] The data input by the input unit module also includes the patient's psychological survey scale, which is used to judge anxiety, mania, and depression.

[0021] Psychological survey scales can preliminarily determine the presence and severity of the above-mentioned mental illnesses.

[0022] The diagnostic unit module is used to train the initial dry eye grading diagnosis neural network model based on a large amount of clinical history data of dry eye patients, so as to obtain the correspondence between the patient's dry eye OSDI scale, symptoms and lower tear meniscus height examination result characteristics and the clinical diagnostic grading of dry eye; The data matching module is used to match the dry eye diagnosis result characteristics of the patient to be diagnosed to the above corresponding relationship and generate a corresponding matching diagnosis report; The output unit module is used to output the clinical grade of dry eye of the patient to be diagnosed.

[0023] The diagnosis unit module includes a model training sub-unit module, and the model training sub-unit module includes a data input module, an initial diagnosis module and a diagnosis typing module; The data input module is used to obtain a large amount of historical clinical data of dry eye patients, set data preprocessing rules, and process the data information through the data preprocessing rules to obtain a diagnostic data set; The initial diagnosis module is used to randomly divide the diagnosis data set obtained by the data input module and the image data set collected by the mobile phone into a training set, a validation set and a test set, and then input the data of the corresponding training set and validation set into the initial dry eye grading diagnosis neural network model for model training. When the training conditions are met, the training is stopped to obtain the dry eye grading diagnosis neural network model; The diagnostic classification module is used to further analyze the symptoms of dry eye patients, the OSDI scale, and the lower tear meniscus height examination results, and use the cluster analysis method to derive the correspondence between the symptoms of dry eye patients, the OSDI scale, and the lower tear meniscus height results and mild dry eye, moderate dry eye, and severe dry eye.

[0024] The historical clinical data of dry eye patients obtained by the data input module include OSDI scale, psychological survey scale, symptom result characteristics and lower tear meniscus height result characteristic data. Among them, OSDI scale refers to dry eye questionnaire; symptom result characteristics include the degree or frequency of eye dryness, foreign body sensation, burning sensation, fatigue, discomfort, red eyes, and visual fluctuations.

[0025] Data preprocessing rules include data type transformation, data filling, and data deletion. Data type transformation includes binary data conversion (for example, the gender field "male" or "female" can represent "0" or "1" respectively) and multi-value data conversion (for example, symptoms "dry eyes", "foreign body sensation", "burning sensation", "fatigue", "discomfort", "red eyes", "visual fluctuations" can be converted to "1", "2", "3", "4", "5", "6", "7"); data filling is to fill the empty value fields of each item (such as patient ID, ID number, age, gender, etc.) with the mean value to improve the accuracy of model training; data deletion is to delete images where eyelashes block the tear river or pupil imaging is not captured.

[0026] The data input module obtains the anterior segment photographic image data of the dry eye comprehensive analyzer, locates the target based on the Fast R-CNN model, obtains the region of interest, that is, the corresponding position of the tear meniscus and the pupil, outputs the predicted candidate frame, identifies the upper and lower edges of the tear meniscus, and calculates the height of the tear meniscus.

[0027] According to the tear meniscus height, the severity of dry eyes can be judged. Finally, the historical dry eye basic data information and the corresponding dry eye classification standards are obtained based on the dry eye basic data information. The dry eye data information is classified and labeled according to the historical dry eye basic data information and the corresponding dry eye classification annotation categories, and at the same time associated with the dry eye data set. Using the built deep network structure, the pictures collected by the mobile phone are processed to obtain the tear meniscus height of the person taking the photo, so as to judge whether the person in the photo suffers from dry eyes and the severity of dry eyes.

[0028] The diagnostic process of the diagnostic unit module is: S1. Obtain ocular surface images of dry eye patients as historical data sets, and preprocess the historical data of dry eye patients; S2. Positioning and labeling the pupil center and tear meniscus region of the ocular surface image of the dataset preprocessed in step S1; In step S2, the ocular surface image is located based on the Fast R-CNN model to obtain the region of interest, i.e., the position corresponding to the tear meniscus and the pupil, and the predicted candidate box is output; To ensure that the extracted area includes the complete tear meniscus area, the size of the extracted area is finally set to 512×288 to remove redundant information. To ensure the accuracy of the subsequent neural network segmentation results, the ocular surface image content needs to be annotated at the pixel level. The labels are made using image J software and are jointly annotated by multiple physicians with sufficient experience and level. At the same time, the consistency of the annotations is regularly analyzed to ensure that the data labels are clearly annotated. The specific annotation locations are the pupil and tear meniscus areas of the ocular surface image set.

[0029] The pupil center of the ocular surface image is located and annotated based on two easily identifiable biological features of the pupil, as follows: The pupil is roughly located on the ocular surface image. The difference in grayscale values ​​between the pupil area and other areas of the eye is used in combination with binarization technology to determine the area with high grayscale values ​​including the pupil. The pupil color is the position with the lowest grayscale value of the entire eye. The image is processed by morphological operators including opening and closing operations, dilation and erosion to remove unnecessary parts. The geometric characteristics of the pupil are circular or elliptical, and the Hough circle transform is used to determine the size and center position of the pupil area. The center of the circle is the pupil center. In step S2, the tear meniscus area is segmented by building a semantic segmentation network model of the tear meniscus area, as follows: the tear meniscus area is manually marked on the ocular surface image in the historical data set of step S1, and divided into a training set and a verification set, the training set is input into the semantic segmentation network for training, and the test set is input into the trained segmentation network for verification testing, and finally a semantic segmentation network model of the tear meniscus area is obtained.

[0030] In the actual clinical diagnosis and detection of dry eyes, doctors are most concerned about the height of the tear river in the six directions of the tear river pupil center, rather than the specific target type. Therefore, the tear river area in the original image is used as the foreground, and the remaining pixels are used as the background to extract the tear river area in the ocular surface image. Figure 4 As shown in the figure, the tear river segmentation network architecture consists of 9 dilated convolution modules, 1×1 convolution dimension reduction layer, 3×3 convolution layer and pixel classification sublayer. The specific composition structure of the network model is analyzed through the following three parts: (1) Dilated convolution module In the specific network structure of Tear River segmentation, each dilated convolution module consists of two 3×3 dilated convolution units with the same dilation coefficient. Each 3×3 dilated convolution unit contains a 3×3 dilated convolution layer, a BN batch normalization layer, and a ReLU activation function. In order to ensure that the network receptive field is large enough and meets the accuracy requirements to achieve better segmentation results, two 3×3 convolution kernels are used instead of one 5×5 convolution kernel. The reasons for this are: Model deepening: The receptive field size of two 3×3 convolutional layers is equivalent to that of a 5×5 convolution. Using two layers of 3×3 convolution can deepen the model and give it more nonlinear expression capabilities.

[0031] The number of parameters is reduced: for two 3×3 convolution kernels, the total number of parameters used is 2×(3x3)×channels, and for one 5×5 convolution kernel, 5×5×channels. The amount of computation is reduced by 28%. The receptive field size does not change.

[0032] (2) Cross-layer connection method As the number of convolutional layers increases, the details of the original image may be lost. The present invention uses a cross-layer connection method to combine the local information of the first four dilated convolutional modules with the global information output by the last module to comprehensively utilize the fine local information of the output feature maps of the first four modules. This is conducive to improving the semantic segmentation performance and thus improving the accuracy of tear river edge segmentation.

[0033] (3) Feature fusion method First, assume that a 3-channel color image is input, and a 3×3×3 convolution kernel is used to perform sliding convolution on the original image to obtain a feature map. From this, we can conclude that the number of channels of the feature map is determined by the number of convolution kernels.

[0034] When 64 convolution kernels of size 3×3×3 are used to perform sliding convolution on the original image, 64 different feature maps are generated. After two layers of convolution layers are stacked in the channel dimension, the feature dimension of the image increases, while the amount of information in each dimension remains unchanged.

[0035] The Concatenation feature fusion method is used in the network model to integrate the information of feature maps. It uses the semantic information of feature maps of different scales, increases the width of the network by increasing the number of channels, and improves the segmentation accuracy.

[0036] Finally, after feature fusion, 1×1 convolution is used for dimensionality reduction, and a continuous convolution layer is used to extract features, which are combined to obtain a more accurate output. Then, the pixel classification sublayer is used to distinguish foreground and background pixels, and the segmentation result is output.

[0037] S3. Calculate the lower tear meniscus height according to the annotation results of step S2, and grade the severity of dry eye symptoms; In step S3, the height of the tear mane is calculated by subtracting the distance from the center of the pupil to the upper edge of the tear mane from the distance from the center of the pupil to the lower edge of the tear mane in the six-point direction of the center of the pupil, and the difference is the height of the tear mane; The calculation expression is: In the formula, It is the height of the River of Tears; N Indicates that the tear river is taken near the six o'clock position of the pupil center. N The pixels are averaged; It represents the distance from the pupil center to the lower edge of the tear meniscus; It represents the distance from the pupil center to the upper edge of the tear meniscus; about N The value of . NIf the value of is too large, it will include more uncertain factors. For example, some tear menisci are short, resulting in a sudden change in height. Or the tear meniscus is usually thicker in the middle and gradually decreases towards the sides of the eye corners. This inherent height change may cause the value after taking the average value to change. N If the value of is too small, the credibility of the result will be reduced. If the height calculation of a single pixel at six points is used, there will be certain randomness and errors. Calculating the average value of the tear river of multiple pixels can reduce the measurement error and prevent generalization. After many experiments, the final N The value is 9 pixel values.

[0038] Dry eye symptom severity grading: The tear meniscus height is used as a grading indicator, which is specifically divided into normal, mild dry eyes, moderate dry eyes, and severe dry eyes; The grading range is as follows: tear meniscus height > 0.2 mm is the normal range; tear meniscus height ≤ 0.2 mm, the patient has mild dry eyes; tear meniscus height ≤ 0.1 mm, the patient has moderate dry eyes; when the tear meniscus height cannot be measured, the patient has severe dry eyes.

[0039] S4. Construct a neural network model for dry eye grading diagnosis; In step S4, the dry eye grading diagnosis neural network model is as follows: In the formula, R represents the marked area; S A text description representing the marked area; I Represents the entire image; θ Represents the parameters of the convolutional neural network model and the recurrent neural network model.

[0040] Input during training I , θ , S ,get R , and adjusted through training θ , input during prediction I ,get R and S During diagnosis, all ocular surface images are first identified and classified into two categories: dry eye and non-dry eye. Then, the semantic attribute prediction layer and semantic attribute prediction vector of dry eye severity are obtained based on the mapping between images and semantic attributes. Then, the LSTM model is used to generate a mapping from the semantic attribute prediction vector to the word sequence, and finally the dry eye severity described by the marker is output.

[0041] S5. Input the ocular surface image data of the patient to be diagnosed into the model of step S4 to obtain the dry eye grading diagnosis result.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the protection scope of the present invention.

Claims

1. A comprehensive system for dry eye diagnosis and treatment based on deep neural network, characterized by: It includes an input unit module, a diagnosis unit module, a data matching module and an output unit module; The input unit module is used to input the OSDI scale and symptom result characteristics of the patient to be diagnosed; The diagnostic unit module is used to train an initial dry eye grading diagnosis neural network model based on a large amount of clinical history data of dry eye patients, so as to obtain the correspondence between the patient's dry eye OSDI scale, symptoms and lower tear meniscus height examination result characteristics and the dry eye clinical diagnosis grading; The diagnosis unit module includes a model training subunit module; The data matching module is used to match the dry eye diagnosis result characteristics of the patient to be diagnosed to the above corresponding relationship and generate a corresponding matching diagnosis report; The output unit module is used to output the clinical grade of dry eye of the patient to be diagnosed.

2. According to claim 1, a comprehensive system for dry eye diagnosis and treatment based on deep neural network, characterized in that: The model training subunit module includes a data input module, an initial diagnosis module and a diagnosis typing module; The data input module is used to obtain a large amount of historical clinical data of dry eye patients, set data preprocessing rules, and process the data information according to the data preprocessing rules to obtain a diagnostic data set; The initial diagnosis module is used to randomly divide the diagnosis data set processed by the data input module and the image data set collected by the mobile phone into a training set, a validation set and a test set, and then input the data of the training set and the validation set into the initial dry eye grading diagnosis neural network model for model training, and finally obtain the dry eye grading diagnosis neural network model; The diagnostic classification module is used to obtain the correspondence between the OSDI scale, symptoms and lower tear meniscus height result characteristics of dry eye patients and mild dry eyes, moderate dry eyes and severe dry eyes.

3. The dry eye diagnosis and treatment comprehensive system based on deep neural network according to claim 2, characterized in that: The historical clinical data of dry eye patients acquired by the data input module include OSDI scale, symptom result characteristics and lower tear meniscus height result characteristic data; Among them, the OSDI scale refers to the dry eye questionnaire; the symptom result characteristics include the degree or frequency of eye dryness, foreign body sensation, burning sensation, fatigue, discomfort, red eyes, and vision fluctuations.

4. The comprehensive dry eye diagnosis and treatment system based on deep neural network according to claim 3, characterized in that: The data preprocessing rules include data type conversion, data filling and data deletion; Data type conversion includes binary data conversion and multi-value data conversion; Data filling is to fill the empty value fields of each item with mean value; Data deletion is to delete images where the eyelashes block the tear canal over a large area or where the pupil is not imaged.

5. A comprehensive dry eye diagnosis and treatment system based on deep neural network according to claim 4, characterized in that: The diagnostic process of the diagnostic unit module is as follows: S1. Obtain ocular surface images of dry eye patients as historical data sets, and preprocess the historical data of dry eye patients; S2. Positioning and labeling the pupil center and tear meniscus region of the ocular surface image of the dataset preprocessed in step S1; S3. Calculate the lower tear meniscus height according to the annotation results of step S2, and grade the severity of dry eye symptoms; S4. Construct a neural network model for dry eye grading diagnosis; S5. Input the ocular surface image data of the patient to be diagnosed into the model of step S4 to obtain the dry eye grading diagnosis result.

6. A comprehensive dry eye diagnosis and treatment system based on deep neural network according to claim 5, characterized in that: In step S2, the ocular surface image is located based on the Fast R-CNN model, the region of interest, i.e., the position corresponding to the tear meniscus and the pupil, is obtained, and the predicted candidate box is output; the pupil center of the ocular surface image is located and marked according to two easily identifiable biological features of the pupil, as follows: The pupil is roughly located on the ocular surface image. The difference in grayscale values ​​between the pupil area and other areas of the eye is used in combination with binarization technology to determine the area with high grayscale values ​​including the pupil. The pupil color is the position with the lowest grayscale value of the entire eye. The image is processed by morphological operators including opening and closing operations, dilation and erosion to remove unnecessary parts. The geometric characteristics of the pupil are used to determine the size and center position of the pupil area using the Hough circle transform. The center of the circle is the pupil center.

7. A comprehensive dry eye diagnosis and treatment system based on deep neural network according to claim 6, characterized in that: In the step S2, the tear meniscus area is segmented by building a semantic segmentation network model of the tear meniscus area, specifically as follows: the tear meniscus area is manually marked on the ocular surface image in the historical data set of step S1, and divided into a training set and a verification set, the training set is input into the semantic segmentation network for training, and the test set is input into the trained segmentation network for verification testing, and finally a semantic segmentation network model of the tear meniscus area is obtained.

8. The dry eye diagnosis and treatment comprehensive system based on deep neural network according to claim 7, characterized in that: In step S3, the height of the tear mane is calculated by subtracting the distance from the center of the pupil to the upper edge of the tear mane from the distance from the center of the pupil to the lower edge of the tear mane in the six-point direction of the center of the pupil, and the difference is the height of the tear mane; The calculation expression is: In the formula, It is the height of the River of Tears; N Indicates that the tear river is taken near the six o'clock position of the pupil center. N The pixels are averaged; It represents the distance from the pupil center to the lower edge of the tear meniscus; It represents the distance from the pupil center to the upper edge of the tear meniscus; Dry eye symptom severity grading: The tear meniscus height is used as a grading indicator, which is specifically divided into normal, mild dry eyes, moderate dry eyes, and severe dry eyes; The grading range is as follows: tear meniscus height > 0.2 mm is the normal range; tear meniscus height ≤ 0.2 mm, the patient has mild dry eyes; tear meniscus height ≤ 0.1 mm, the patient has moderate dry eyes; when the tear meniscus height cannot be measured, the patient has severe dry eyes.

9. A comprehensive dry eye diagnosis and treatment system based on deep neural network according to claim 8, characterized in that: In step S4, the dry eye grading diagnosis neural network model is as follows: In the formula, R represents the marked area; S A text description representing the marked area; I Represents the entire image; θ Represents the parameters of the convolutional neural network model and the recurrent neural network model.

10. A comprehensive dry eye diagnosis and treatment system based on deep neural network according to claim 9, characterized in that: The data input by the input unit module also includes the patient's psychological survey scale.

Images