System for monitoring intraocular perfusion pressure in real time to assist vitreoretinal surgery based on deep learning

Abstract

The invention provides a system for monitoring intraocular perfusion pressure in real time to assist vitreoretinal surgery based on deep learning, and belongs to the field of artificial intelligence assisted surgery. According to the system disclosed by the invention, the intraocular perfusion pressure is monitored in real time and feedback and early warning are carried out by monitoring the gray value change of the optic disc area. The system can assist a surgeon in monitoring intraocular perfusion pressure in real time, interference of subjective factors is avoided, higher accuracy and objectivity can be provided, safety and reliability of an operation are enhanced, and lagging judgment caused by fatigue or distraction in the operation process is also avoided. The system can continuously and stably monitor the whole operation process in real time and feed back optic disc state changes in the operation process, it is ensured that a doctor always masters accurate data of intraocular perfusion pressure and adjusts operation in time, the risk of eye tissue damage caused by the fact that abnormity is not perceived in time is effectively reduced, potential damage to optic nerves of a patient is reduced, and the patient experience is improved. Good application prospects are realized.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence-assisted surgical operations, and particularly relates to a system for assisting vitreoretinal surgery by real-time monitoring of intraocular perfusion pressure based on deep learning. Background Art

[0002] Vitreoretinal surgery is one of the complex and delicate surgical types in the field of ophthalmology, and its purpose is to treat various vitreous and retinal diseases, such as retinal detachment, vitreous hemorrhage, macular lesions, etc. During vitreoretinal surgery, the intraocular perfusion pressure is often actively increased to stop bleeding. In addition, operations such as peripheral vitrectomy, heavy water injection, and silicone oil injection by pressing on the sclera often cause a sudden increase in intraocular pressure. Research shows that high intraocular perfusion pressure may cause adverse consequences such as reduced intraocular blood supply, retinal damage, reduced choroidal blood flow, and visual impairment. Therefore, it is very necessary to monitor the intraocular perfusion pressure during vitreoretinal surgery.

[0003] However, during vitreoretinal surgery, the surgeon needs to concentrate on performing vitrectomy and cannot take into account the changes in intraocular perfusion pressure. Therefore, it is particularly important to assist in monitoring the changes in intraocular perfusion pressure and provide real-time feedback. Most current methods for monitoring intraocular perfusion pressure rely on doctors' experience and traditional instruments and equipment. For example, using a Zeiss microscope for observation, this method highly depends on the doctor's personal experience. Another example is to detect the water flow pressure at the distal end of the perfusion tube of the cutting machine, but this pressure is inaccurate and has a delay. Therefore, these methods have problems such as poor real-time performance, insufficient accuracy, and reliance on manual active monitoring, which are not only time-consuming and laborious, but also easily interfered by subjective factors.

[0004] With the rapid development of deep learning technology, it has been widely used in the field of medical image processing and analysis. Deep learning algorithms can automatically learn and extract features in images, perform operations such as image classification, segmentation, and recognition, providing new ideas and methods for medical diagnosis. In the field of ophthalmology, deep learning has been used for the analysis of fundus images, such as the diagnosis of retinal diseases and the segmentation of retinal blood vessels, and has achieved good results. However, there is currently a lack of a deep learning system for real-time monitoring of intraocular perfusion pressure during vitreoretinal surgery.

[0005] Therefore, developing a system for assisting vitreoretinal surgery by real-time monitoring of intraocular perfusion pressure based on deep learning has important clinical significance and application value. Summary of the Invention

[0006] The purpose of the present invention is to provide a system for assisting vitreoretinal surgery by real-time monitoring of intraocular perfusion pressure based on deep learning.

[0007] The present invention provides a system for real-time monitoring of intraocular perfusion pressure to assist vitreoretinal surgery based on deep learning, which monitors the intraocular perfusion pressure by using the gray value of the optic disc area.

[0008] Further, the aforementioned system includes:

[0009] Input module: configured to input the video during the vitreoretinal surgery;

[0010] Image extraction module: configured to extract the video into images;

[0011] Optic disc area positioning module: uses the Yolo network to position the optic disc area in the image to obtain the optic disc area image;

[0012] Image gray value conversion module: configured to convert the pixels of the optic disc area image into gray values;

[0013] Output module: compares the obtained gray value with the set gray value threshold, and issues an alarm when it is higher than the threshold.

[0014] Further, in the input module, the real-time video during the vitreoretinal surgery is input.

[0015] Further, in the image extraction module, the video is extracted into images at intervals of 1 to 15 frames; preferably, the resolution of the images is 1920×1080 pixels.

[0016] Preferably, in the image extraction module, the Open CV module is used to extract the images.

[0017] Further, in the optic disc area positioning module, the Yolo network is the Yolov5 network.

[0018] Further, in the optic disc area positioning module, the Yolov5 network is an optimized Yolov5 network, and the optimization method is: adding a detection layer in the Yolov5 network and / or adjusting the anchor box size of the Yolov5 network;

[0019] Among them, the detection layer is added at the 13th layer, C3 position of the Yolov5 network;

[0020] Adjust the anchor box size of the Yolov5 network to adjust the original anchor box size of the Yolov5 network to the first layer [55, 61, 66, 71, 79, 83], the second layer [94, 98, 107, 110, 123, 129], the third layer [127, 153, 158, 164, 186, 176], and the fourth layer [202, 189, 243, 256, 332, 342] of the four detection feature layers.

[0021] Further, in the optic disc area positioning module, when using the Yolo network for positioning, images with a confidence level less than 90% are excluded.

[0022] Further, in the image grayscale value conversion module, the pixels of the optic disc area image are converted into grayscale values according to the RGB weighting coefficients. The RGB weighting coefficients are as follows: the weight of red is 0.5, the weight of green is 0.3, and the weight of blue is 0.2.

[0023] Preferably, when converting, based on the center of the optic disc area image, the pixels of the area image with 2 / 3 of the middle area of the optic disc area image are selected for conversion.

[0024] Further, in the output module, the threshold of the grayscale value is 17 - 25.

[0025] The present invention also provides a computer-readable storage medium, on which there is stored a computer program for implementing the aforementioned system for real-time monitoring of intraocular perfusion pressure to assist vitreoretinal surgery based on deep learning.

[0026] In the present invention, the Yolo network is a real-time object detection algorithm based on deep learning. Its main purpose is to directly predict the category and bounding box of the object from the input image through a single neural network. The Yolov5 network is the 5th version of the Yolo series.

[0027] The present invention is the first to use the optic disc state to monitor the intraocular perfusion pressure. Under high intraocular perfusion pressure, the optic disc will turn pale, the blood vessels will become thinner, and the grayscale value of the optic disc area will increase. Therefore, the present invention uses the grayscale value of the optic disc area to monitor the intraocular perfusion pressure in real time during vitreoretinal surgery, so as to assist the surgeon to adjust the intraocular perfusion pressure in time and minimize the potential damage to the patient's optic disc nerve tissue.

[0028] The system of the present invention mainly consists of two key steps: (1) using the optimized Yolov5 network to continuously and stably track the optic disc area during the operation in real time, providing reliable data input for subsequent analysis; (2) converting the pixels of the optic disc area detected by the Yolov5 network into grayscale values according to specific RGB weighting coefficients to quantify the change in the optic disc state. Under normal physiological conditions, the brightness of the optic disc area is maintained within the reference range. When high perfusion occurs, due to the vasoconstriction of the optic disc blood vessels and ischemic pallor of the tissue, the grayscale value of the target area shows a significant increase. The system monitors the increase rate of the grayscale value of the optic disc area in real time and dynamically compares it with the preset threshold. When it detects that the increase rate of the grayscale value exceeds the threshold, it immediately triggers a type I alarm to indicate the risk of acute high perfusion, and at the same time starts a timer to record the duration of the abnormal state. If the cumulative duration exceeds the set critical threshold, it triggers a type II alarm to indicate persistent high perfusion damage. The dual warning mechanism feedback ensures the timeliness of clinical intervention.

[0029] The present invention has achieved the following beneficial effects:

[0030] The present invention provides a system for real-time monitoring of intraocular perfusion pressure based on deep learning to assist vitreoretinal surgery. The system monitors the change of gray value in the optic disc area to monitor the intraocular perfusion pressure in real time and gives feedback and warning. The system of the present invention can assist surgeons in real-time monitoring of intraocular perfusion pressure. Compared with relying on doctors' subjective experience judgment and traditional monitoring methods, it avoids the interference of subjective factors, can provide higher accuracy and objectivity, enhances the safety and reliability of the surgery, and also avoids the lag judgment caused by fatigue or distraction during the surgery. The system can continuously and stably monitor the whole surgical process in real time and feedback the change of the optic disc state during the surgery, ensuring that doctors always master the state of intraocular perfusion pressure, timely adjust the operation, effectively reduce the risk of ocular tissue damage caused by failure to detect abnormalities in time, reduce the potential damage to the patient's optic nerve, and has good application prospects.

[0031] Obviously, based on the above content of the present invention, according to the common general knowledge and conventional means in the art, without departing from the above basic technical idea of the present invention, various other forms of modification, substitution or change can be made.

[0032] The following is a further detailed description of the above content of the present invention through specific embodiments in the form of examples. However, this should not be construed as limiting the scope of the above subject matter of the present invention to the following examples. All technologies implemented based on the above content of the present invention belong to the scope of the present invention. Specific Embodiments

[0033] It should be particularly noted that the algorithms for data acquisition, transmission, storage and processing steps not specifically described in the embodiments, as well as the hardware structures and circuit connections not specifically described, can all be implemented through the content publicly disclosed in the prior art.

[0034] Embodiment 1. The system of the present invention for real-time monitoring of intraocular perfusion pressure based on deep learning to assist vitreoretinal surgery

[0035] The system of the present invention for real-time monitoring of intraocular perfusion pressure based on deep learning to assist vitreoretinal surgery includes:

[0036] Input module: configured to input the real-time video during vitreoretinal surgery.

[0037] Image extraction module: configured to extract images from the video at intervals of 10 frames based on the OpenCV module, with the image resolution of 1920×1080 pixels, and the obtained images are passed to the next module.

[0038] Optic disc area positioning module: Configured to use the optimized Yolov5 network to perform real-time tracking and positioning on the optic disc area in the image, and obtain the optic disc area image. When using the Yolov5 network for positioning, images with a confidence level less than 90% will be automatically excluded.

[0039] Since the optic disc area in ophthalmic surgery is small, in order to detect the target more precisely, the present invention optimizes the Yolov5 network structure. The optimization method is as follows: (1) Add a detection layer at the C3 position of the 13th layer of the Yolov5 network structure; (2) Adjust the anchor box size. Specifically, the original anchor box sizes in the original Yolov5 network structure, the first layer [10, 13, 16, 30, 33, 23], the second layer [30, 61, 62, 45, 59, 119], the third layer [116, 90, 156, 198, 373, 326] are adjusted to the first layer [55, 61, 66, 71, 79, 83], the second layer [94, 98, 107, 110, 123, 129], the third layer [127, 153, 158, 164, 186, 176], and the fourth layer [202, 189, 243, 256, 332, 342] for four detection feature layers.

[0040] Image grayscale value conversion module: Configured to convert the pixels of the optic disc area image into grayscale values according to the RGB weighting coefficients. When converting, the center of the optic disc area is used as a reference, and an area with 2 / 3 of the area in the middle part of the optic disc area is selected for conversion.

[0041] In order to enhance the detection sensitivity, the present invention adjusts the standard weighting ratio [0.2989, 0.5870, 0.1140] to [0.5, 0.3, 0.2] to increase the weight of the red channel.

[0042] Output module: Set the threshold of the grayscale value to 17 and the cumulative duration threshold to 10s. When the image grayscale value exceeds the threshold, a type I alarm is issued to prompt the risk of acute hyperperfusion, and at the same time, a timer is started to record the duration of the abnormal state. When the cumulative duration of the abnormal state exceeds the threshold, a type II alarm is issued to prompt persistent hyperperfusion injury.

[0043] The beneficial effects of the present invention are demonstrated by the following specific test examples.

[0044] Test Example 1. Verification of the effect of the system of the present invention

[0045] The images used in this test example are from the Vitreous and Retina Disease Group of Sichuan Provincial People's Hospital, covering 80 vitreoretinal surgery videos during the period from July 2023 to December 2024, including cases such as silicone oil injection, heavy water injection, and indentation of the peripheral retina. Images were extracted every 10 frames, with a resolution of 1920×1080 pixels, and blurred or overexposed images caused by environmental or equipment factors were excluded. Three ophthalmologists with clinical experience were invited to label the optic disc area in the images using Labelimg, and one mid-level seniority physician was invited to review the labeled areas.

[0046] To improve data quality, the images were grayscale processed to reduce noise and computational complexity. At the same time, the dataset was augmented through operations such as horizontal flipping, vertical flipping, and random rotation to simulate different viewing angles and enhance the robustness of the model. Finally, a dataset containing 56,708 optic disc images was constructed. 80% of this dataset was used for training, 10% for validation, and 10% for testing.

[0047] In this invention, the original Yolov5 network and the optimized Yolov5 network in Example 1 were verified for their effectiveness on 7569 external validation set images.

[0048] The results show that:

[0049] After external validation (the image data did not participate in model training), the detection effects of the optimized Yolov5 network and the original Yolov5 network of this invention are shown in Table 1.

[0050] Table 1. Detection effects of the optimized Yolov5 network and the original Yolov5 network

[0051]

[0052]

[0053] As can be seen from Table 1; when using the optimized Yolov5 network for intraocular perfusion pressure monitoring, compared with the original Yolov5 network, the precision, recall rate, mAP50, and F1 score were improved by 5.0%, 1.6%, 2.4%, and 3.5% respectively, indicating that the optimized model has higher accuracy and reliability in detecting targets and can better meet the need for real-time monitoring of intraocular perfusion pressure during vitreoretinal surgery.

[0054] In summary, the present invention provides a system for real-time monitoring of intraocular perfusion pressure to assist vitreoretinal surgery based on deep learning. The system monitors the change in gray value of the optic disc region to monitor the intraocular perfusion pressure in real time and gives feedback and warning. The system of the present invention can assist surgeons in real-time monitoring of intraocular perfusion pressure. Compared with relying on doctors' subjective experience judgment and traditional monitoring methods, it avoids the interference of subjective factors, can provide higher accuracy and objectivity, enhances the safety and reliability of the surgery, and also avoids the lag judgment caused by fatigue or distraction during the surgery. The system can continuously and stably monitor the entire surgical process in real time and feedback the changes in the optic disc state during the surgery, ensuring that doctors always master the accurate data of intraocular perfusion pressure, timely adjust the operation, effectively reduce the risk of ocular tissue damage caused by failure to detect abnormalities in time, reduce the potential damage to the patient's optic nerve, and has good application prospects.

Claims

1. A system for real-time monitoring of intraocular perfusion pressure to assist vitreoretinal surgery based on deep learning, characterized in that: It uses the grayscale value of the optic disc area to monitor the intraocular perfusion pressure.

2. The system according to claim 1, characterized in that: include: Input module: configured to input a video during vitreoretinal surgery; Image extraction module: configured to extract the video into images; Optic disc area positioning module: Use the Yolo network to locate the optic disc area in the image and obtain the optic disc area image; An image grayscale value conversion module: configured to convert pixels of the optic disc area image into grayscale values; Output module: Compare the obtained grayscale value with the set grayscale value threshold, and issue an alarm when it is higher than the threshold.

3. The system according to claim 2, characterized in that: In the input module, real-time video during vitreoretinal surgery is input.

4. The system according to claim 2, characterized in that: In the image extraction module, the video is extracted into images at intervals of 1 to 15 frames; Preferably, in the image extraction module, an Open CV module is used to extract the image.

5. The system according to claim 2, characterized in that: In the optic disc area positioning module, the Yolo network is a Yolov5 network.

6. The system according to claim 5, characterized in that: In the optic disc area positioning module, the Yolov5 network is an optimized Yolov5 network, and the optimization method is: adding a detection layer to the Yolov5 network and / or adjusting the anchor frame size of the Yolov5 network; Among them, the detection layer is added to the 13th layer of the Yolov5 network, at position C3; Adjust the anchor box size of the Yolov5 network to adjust the original anchor box size of the Yolov5 network to four detection feature layers: the first layer [55, 61, 66, 71, 79, 83], the second layer [94, 98, 107, 110, 123, 129], the third layer [127, 153, 158, 164, 186, 176], and the fourth layer [202, 189, 243, 256, 332, 342].

7. The system according to claim 2, characterized in that: In the optic disc area positioning module, images with a confidence level less than 90% are excluded when positioning is performed using the Yolo network.

8. The system according to claim 2, characterized in that: In the image grayscale value conversion module, the pixels of the optic disc area image are converted into grayscale values ​​according to the RGB weighting coefficients, wherein the RGB weighting coefficients are: the weight of red is 0.5, the weight of green is 0.3, and the weight of blue is 0.2; Preferably, during the conversion, the center of the optic disc area image is taken as a reference, and pixels of the area image of the middle 2 / 3 of the optic disc area image are selected for conversion.

9. The system according to claim 2, characterized in that: In the output module, the threshold of the grayscale value is 17-25.

10. A computer-readable storage medium, characterized in that: Stored thereon is: a computer program for implementing a system for real-time monitoring of intraocular perfusion pressure based on deep learning to assist vitreoretinal surgery as described in any one of claims 1 to 9.

Images