Eye disease identification method, system, medium and equipment

Abstract

The invention discloses an eye disease recognition method and system, a medium and equipment, and relates to the field of image recognition, and the method comprises the steps: constructing a symptom recognition model based on improved ResNext50; in the improved ResNext50, a convolution layer of an original ResNext50 model is replaced by grouping depth separable convolution, a classification head of an original global average pooling and full connection layer is replaced by a feature fusion structure based on a multi-scale pyramid, and a space-channel double-attention mechanism is added at the tail end of a residual branch of an original Bottleneck structure; and carrying out positioning and focusing processing on a to-be-identified eye image, carrying out classification identification on the positioned and focused eye image through the trained symptom identification model, and obtaining an eye disease identification result.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and particularly to a method, system, medium, and device for identifying eye diseases. Background Art

[0002] The slit lamp microscope is an essential and important instrument for ophthalmic examinations. The slit lamp microscope consists of an illumination system and a binocular microscope. It can not only make superficial lesions clearly observable, but also adjust the focus and the width of the light source to create an "optical section" so that deep tissue lesions can also be clearly visualized. The slit lamp microscope is an active living tissue microscope. By controlling the diaphragm, a thin sheet of visible strong light beam is projected into the patient's eye to complete observation and diagnosis, and it is used for the observation and diagnosis of anterior segment diseases of the eye and anterior segment trauma affecting the structural characteristics of the eye. Clinically, it is used to diagnose lesions or tissue injuries in the anterior segment of the eye from the corneal epithelial tissue to the posterior capsule. In combination with a front lens, gonioscope, three-mirror lens, fundus contact lens, etc., it can examine the eyelids, conjunctiva, cornea, anterior chamber, iris, lens, vitreous body, and the entire retina from the optic disc to the ora serrata.

[0003] Conventional slit lamp microscopes require professional medical personnel to perform on-site operation and observation, which is difficult for patients in remote areas to accept for slit lamp microscope examinations. Even in areas with the equipment, without long-term guidance and training from professional medical personnel, effective detection still cannot be carried out due to limited self-level. Moreover, due to the lack of a lightweight deployment solution for end-edge-cloud collaboration in the prior art, it is difficult to achieve real-time analysis on low-computing-power devices (such as primary medical terminals), resulting in poor recognition efficiency of ophthalmic diseases. Summary of the Invention

[0004] The present invention provides a method, system, medium, and device for identifying eye diseases to solve the above problems existing in the prior art, that is, the problem of how to improve the recognition efficiency of eye diseases in the prior art. The present invention provides a method for identifying eye diseases, and the method includes:

[0005] Construct a symptom recognition model based on improved ResNext50; in the improved ResNext50, replace the convolutional layer of the original ResNext50 model with a grouped depthwise separable convolution, and replace the classification heads of the original global average pooling and fully connected layers with a feature fusion structure based on a multi-scale pyramid, and add a spatial-channel dual attention mechanism at the end of the residual branch of the original Bottleneck structure;

[0006] Perform positioning and focusing processing on the eye image to be recognized, and classify and recognize the positioned and focused eye image through the trained symptom recognition model to obtain the eye disease recognition result.

[0007] Optionally, the focal loss and dice loss are adopted to optimize the loss function of the improved ResNext50.

[0008] Optionally, the eye disease recognition result specifically includes:

[0009] Corneal epithelial defect, conjunctival congestion, and corneal edema.

[0010] The present invention provides an eye disease recognition system, including:

[0011] A construction module for constructing a symptom recognition model based on the improved ResNext50; in the improved ResNext50, the convolutional layer of the original ResNext50 model is replaced with a grouped depthwise separable convolution, and the classification head of the original global average pooling and fully connected layers is replaced with a feature fusion structure based on a multi-scale pyramid, and a spatial-channel dual attention mechanism is added at the end of the residual branch of the original Bottleneck structure;

[0012] An intelligent slit lamp microscope for positioning and focusing the eye image to be recognized, and classifying and recognizing the positioned and focused eye image through the trained symptom recognition model to obtain the eye disease recognition result.

[0013] The present invention provides a computer-readable storage medium, and the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned eye disease recognition method is realized.

[0014] The present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the above-mentioned eye disease recognition method is realized.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides an eye disease recognition method, which constructs a symptom recognition model based on the improved ResNext50 by introducing a multi-scale feature pyramid and a spatial-channel dual attention mechanism. By inputting the eye image information into the symptom recognition model, various eye diseases can be accurately and efficiently recognized, thereby assisting medical staff in the diagnosis and treatment of further ophthalmic diseases, being applicable to complex visual tasks, and improving the recognition efficiency of complex lesions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings here are incorporated into the specification and form a part of this specification, showing the embodiments in line with the present invention, and are used together with the specification to explain the principles of the present invention.

[0017] Figure 1Flowchart of an eye disease recognition method provided by an embodiment of the present invention;

[0018] Figure 2 Server software test results provided by an embodiment of the present invention;

[0019] Figure 3 Overall architecture diagram of the diagnostic platform provided by an embodiment of the present invention;

[0020] Figure 4 Structural diagram of the intelligent slit lamp hardware system provided by an embodiment of the present invention;

[0021] Figure 5 Schematic diagram of a computer device for the eye disease recognition method provided by an embodiment of the present invention. Detailed implementation manners

[0022] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] Figure 1 It is a flowchart of an eye disease recognition method provided by an embodiment of the present invention. As Figure 1 shown, an eye disease recognition method shown in this embodiment includes:

[0024] S1: Construct a symptom recognition model based on improved ResNext50; in the improved ResNext50, the convolutional layer of the original ResNext50 model is replaced with a grouped depthwise separable convolution, and the classification heads of the original global average pooling and fully connected layers are replaced with a feature fusion structure based on a multi-scale pyramid, and a spatial-channel dual attention mechanism is added at the end of the residual branch of the original Bottleneck structure.

[0025] Exemplarily, the present invention uses the ResNext50 model as the basic architecture, replaces the standard 3×3 convolution with a grouped depthwise separable convolution, reducing the number of parameters by 75%. Introduce a spatial-channel dual attention mechanism (SCDA), which improves the sensitivity to tiny lesions (such as corneal epithelial defects) through channel weight allocation and spatial region focusing. By introducing a multi-scale feature pyramid (MFPN), shallow texture information and deep semantic features are fused to solve the problem of large differences in lesion sizes.

[0026] Exemplarily, for the characteristics of eye images, random erasing of pathological regions and simulation generation of reflection spots can be adopted to improve the robustness of the model. Focal Loss and Dice Loss can be jointly optimized to optimize the loss function of the improved ResNext50 model and alleviate class imbalance (such as the too high proportion of normal samples).

[0027] S2: Locate and focus on the eye image to be recognized, and classify and recognize the located and focused eye image through the trained symptom recognition model to obtain the eye disease recognition result.

[0028] Optionally, the eye disease recognition result specifically includes corneal epithelial defect, conjunctival congestion, and corneal edema.

[0029] Exemplarily, when uploading data to the cloud, only the encrypted feature vectors (such as 256-dimensional) rather than the original images are uploaded to reduce bandwidth occupancy. Then, the feature vectors are decrypted and local model inference is performed to generate a diagnostic report.

[0030] Exemplarily, the present invention can adopt a federated learning framework to complete feature extraction at the local terminal and only upload the encrypted feature vectors to the cloud for analysis; by deploying a dynamic differential privacy algorithm, controllable noise is injected during model training to prevent the leakage of patient data.

[0031] Exemplarily, in this embodiment, by selecting 200 eye image data collected by an intelligent slit lamp microscope, the eye disease recognition method proposed by the present invention is used to recognize the eye disease conditions of patients, thereby assisting medical staff to obtain a diagnosis result, and comparing it with the diagnosis results independently diagnosed by 3 attending physicians. After the eye disease recognition method proposed by the present invention recognizes the eye image, the obtained diagnosis result has an accuracy rate of 95.2%.

[0032] Exemplarily, after opening the remote software, enter the username and password. After successful login, jump to the operation interface of the remote end. After finding the device name of the corresponding intelligent slit lamp microscope and successful binding, find the corresponding camera and obtain the list of patients waiting for diagnosis from the database. The patient call number is realized through the call number function of the detection end. When receiving the confirmation information from the detection end, the real-time video of the patient's eyes at the detection end is realized.

[0033] To test the connection of multiple detection end software and multiple remote end software, multiple remote end software and device end software are respectively opened on the computers of the detection end and the remote end, and different usernames and passwords are used to log in respectively as Figure 2As shown. The test results show that: the server software can enable doctors at four remote ends such as WZMU001PC and WZMU002PC to log in, and can also enable logins at four detection ends such as WZMU001VI and WZMU002VI, and can realize the binding and information interaction between the remote end and any idle detection end.

[0034] The test results of the intelligent slit lamp microscope are shown in Table 1. The stroke of the XYZ three-axis moving platform is 150*150*100mm, the precision is 0.01mm, the R axis can rotate 180 degrees, and the precision of the R axis is 0.5 degrees. The adjustment ranges of the slit length and width are both 0-18mm, the precision of each adjustment of the slit length or width is 0.5mm, the slit inclination angle can rotate 180 degrees, the minimum rotation angle each time is 30 degrees, and the slit has 6 brightness adjustments. The test results of the remote diagnosis platform are shown in Table 2. The video clarity transmitted by the diagnosis platform is 960*640 pixels, 20 frames can be transmitted per second, and the delay is 20ms.

[0035] Table 1 Test results of the movement of the intelligent slit lamp microscope

[0036]

[0037] Table 2 Test results of the remote diagnosis platform

[0038]

[0039] The above is the method for identifying eye diseases provided by one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding eye disease identification system, including:

[0040] A construction module for constructing a symptom recognition model based on the improved ResNext50; in the improved ResNext50, the convolutional layer of the original ResNext50 model is replaced with a grouped depthwise separable convolution, and the classification heads of the original global average pooling and fully connected layers are replaced with a feature fusion structure based on a multi-scale pyramid, and a spatial-channel dual attention mechanism is added to the end of the residual branch of the original Bottleneck structure.

[0041] An intelligent slit lamp microscope for performing positioning and focusing processing on the eye image to be recognized, and classifying and recognizing the positioned and focused eye image through the trained symptom recognition model to obtain the eye disease recognition result.

[0042] Exemplarily, the eye disease recognition system adopts a hierarchical design, and the specific architecture is as Figure 3 shown:

[0043] Terminal layer (intelligent hardware): An intelligent slit lamp microscope integrating multi-spectral imaging, motor control, and a local lightweight symptom recognition model, which completes eye image acquisition and preliminary analysis.

[0044] Edge layer (edge computing node): An edge server deployed inside a medical institution, responsible for receiving terminal data, running a high-precision symptom recognition model, and caching frequently accessed data (such as a common disease case library).

[0045] Cloud layer (central cloud platform): Provides distributed storage, federated learning model training, and remote expert consultation services, supporting real-time access from multiple terminals (PC, mobile).

[0046] Exemplarily, for example, the objective lens of the intelligent slit lamp microscope can be adjusted based on an image sharpness evaluation function (such as the Tenengrad gradient method) so that the modulation transfer function of the eye image is greater than or equal to 0.8. The eye image can be positioned through the X-axis or Y-axis of the three-axis motor inside the intelligent slit lamp microscope, focused through the Z-axis on the positioned eye image, the light source color can be switched and the slit width can be adjusted by using a PLD opto-mechanical module, and the CMOS sensor is used to obtain the eye image.

[0047] Such as Figure 4 As shown, the detection-end software can be deployed on a laptop computer. The USB interface of a digital camera is connected to the laptop computer's USB so that the laptop computer can read the video of the digital camera. The USB interface of the DLP opto-mechanical module is also connected to the laptop computer's USB interface so that the laptop computer can control the DLP opto-mechanical module through USB. The serial port of the intelligent slit lamp is connected to the laptop computer so that the laptop computer can control the PLC of the intelligent slit lamp, thereby controlling the three-dimensional movement of the stepping motor. Providing a mobile 5G network for the laptop computer greatly increases the mobility of the detection end.

[0048] For the specific limitations of the eye disease recognition system, reference can be made to the limitations of the eye disease recognition method in the above text, which will not be elaborated here. Each module in the above eye disease recognition system can be implemented in whole or in part through software, hardware, and their combinations. The above modules can be embedded in the processor of a computer device in hardware form or be independent of it, or be stored in the memory of the computer device in software form so that the processor can call and execute the operations corresponding to the above modules.

[0049] The present invention also provides a computer-readable storage medium, which stores a computer program that can be used to execute the above-provided eye disease recognition method.

[0050] The present invention also provides Figure 5 The structural schematic diagram of the computer device shown, such asFigure 5 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the eye disease recognition method provided in the above embodiments.

[0051] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope recorded by the present invention.

Claims

1. An eye disease recognition method, characterized in that, Comprising: Construct a symptom recognition model based on the improved ResNext50; in the improved ResNext50, replace the convolutional layer of the original ResNext50 model with a grouped depthwise separable convolution, and replace the classification head of the original global average pooling and fully connected layers with a feature fusion structure based on a multi-scale pyramid, and add a spatial-channel dual attention mechanism at the end of the residual branch of the original Bottleneck structure; Locate and focus the eye image to be recognized, and classify and recognize the located and focused eye image through the trained symptom recognition model to obtain the eye disease recognition result.

2. The eye disease recognition method according to claim 1, wherein, Adopt Focal Loss and Dice Loss to optimize the loss function of the improved ResNext50.

3. The eye disease recognition method according to claim 1, wherein The eye disease recognition result specifically includes: Corneal epithelial defect, conjunctival congestion and corneal edema.

4. An eye disease recognition system, characterized in that, Including: A construction module for constructing a symptom recognition model based on the improved ResNext50; in the improved ResNext50, replace the convolutional layer of the original ResNext50 model with a grouped depthwise separable convolution, and replace the classification head of the original global average pooling and fully connected layers with a feature fusion structure based on a multi-scale pyramid, and add a spatial-channel dual attention mechanism at the end of the residual branch of the original Bottleneck structure; An intelligent slit lamp microscope for locating and focusing the eye image to be recognized, and classifying and recognizing the located and focused eye image through the trained symptom recognition model to obtain the eye disease recognition result.

5. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and when the computer program is executed by a processor, the eye disease recognition method described in any one of claims 1 to 3 above is implemented.

6. A computer device, characterized in that, Including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the eye disease recognition method described in any one of claims 1 to 3 above is implemented.

Images