Full-automatic method for optimal frame detection and optic nerve sheath width measurement of fundus ultrasonic video

Abstract

The invention provides a full-automatic method for optimal frame detection and optic nerve sheath width measurement of a fundus ultrasonic video. The method comprises the following steps: firstly, training a deep network model for automatically segmenting eyeballs and optic nerve sheaths in a fundus ultrasound image; and then, segmenting each frame of image in each fundus ultrasonic video, designing a fundus ultrasonic image scoring model to score each frame of ultrasonic image, and outputting five images with highest scores in each fundus ultrasonic video as an optimal frame set. And finally, designing an algorithm based on a center line according to a segmentation result of the eyeball and the optic nerve sheath to automatically calculate the optic nerve sheath width at a position 3 mm below the eyeball, and taking a trimming average value of the optic nerve sheath width in the optimal frame image as a width measurement value of the fundus optic nerve sheath. According to the method, under the condition that only the fundus ultrasonic video is input, automatic segmentation of the eyeball and the optic nerve sheath, automatic selection of the optimal frame and automatic measurement of the optic nerve sheath width can be achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of medical image processing and analysis, and in particular relates to a fully automatic method for optimal frame detection of fundus ultrasound video and measurement of optic nerve sheath width. Background Art

[0002] Increased intracranial pressure (ICP) has serious consequences for brain function, and may limit blood flow to the brain, leading to hypoxia and malnutrition, which can cause irreversible brain damage. It may be caused by a variety of factors, including intracranial lesions (such as tumors, cerebral hemorrhage, and cerebral edema), cerebrospinal fluid circulation disorders (such as hydrocephalus), and head trauma (such as traffic accidents or falls). Studying and monitoring increased ICP is essential for timely intervention and improving patient outcomes.

[0003] In clinical practice, lumbar puncture is a commonly used invasive method to measure cerebrospinal fluid pressure to accurately assess ICP. However, during this procedure, patients may experience significant discomfort or pain, and need to be alert to potential risks such as infection, bleeding, and post-puncture headache. Fundus ultrasound imaging provides a safer and more convenient alternative as a non-invasive ICP measurement technique. Fundus ultrasound can be used to assess multiple indicators such as optic nerve diameter (OND) and optic nerve sheath diameter (ONSD). Since the optic nerve sheath is the continuation of the intracranial dura mater and the subarachnoid space, increased intracranial pressure will directly lead to an increase in the optic nerve sheath diameter. Therefore, the optic nerve sheath diameter (ONSD) is considered to be an important parameter for predicting increased intracranial pressure (ICP). In trauma patients, ICP can be continuously monitored by measuring ONSD. Under normal circumstances, the ONSD value is less than 5 mm, and if it is higher than 6 mm, it indicates a significant increase in intracranial pressure. Therefore, it is urgent and very meaningful to develop a fully automatic system that can accurately calculate the width of the optic nerve sheath from fundus ultrasound videos, which relies on automatic segmentation of fundus ultrasound images and automatic detection of optimal frames.

[0004] Deep learning models have been widely used in medical image analysis tasks. Since the U-Net network architecture was proposed in 2015, its design featuring a symmetrical encoding-decoding structure has significantly improved the accuracy of image segmentation, especially in medical image analysis tasks. Subsequently, various variants and innovative methods such as Mask R-CNN and DeepLab series have continued to achieve excellent performance in segmentation accuracy and efficiency. Among them, nnU-Net, as an enhanced and automated U-Net model, can automatically adapt to the characteristics of different data sets without manual adjustment of hyperparameters, greatly simplifying the application process of the model. With its outstanding performance, nnU-Net has demonstrated excellent results in multiple medical image segmentation open challenges, further promoting the development of deep learning in the field of medical image segmentation.

[0005] Existing deep network models for video key frame detection can be roughly divided into two categories. The first category regards key frame detection as a binary classification problem. The detection model is trained through the cross entropy loss function to classify each frame in the video as a key frame or a non-key frame. This method takes advantage of the classification algorithm to extract the features of each frame and make classification predictions. The second method regards key frame detection as a regression problem, estimating a score for each frame based on relevance or importance, indicating its possibility of becoming a key frame. This type of model is trained using mean square error loss and can delicately evaluate the importance of each frame. Regression methods provide greater flexibility in showing the relative importance of frames. However, both types of deep learning models need to label all frames in the video during training, especially the second type of regression-based deep network model, which needs to accurately score each frame, which is very difficult and time-consuming for doctors.

[0006] In addition, accurate measurement of the optic nerve sheath diameter (ONSD) is crucial. This process requires high experience and operational skills of clinicians. However, the quality of ocular ultrasound images is poor, usually manifested as low signal-to-noise ratio, insufficient contrast, and blurred boundaries of the optic nerve sheath, all of which affect the accuracy of clinical measurements. In addition, when measuring the width of the optic nerve sheath 3 mm below the eyeball, the doctor needs to manually measure 3 mm along the optic nerve sheath toward the eyeball and draw a vertical line to calculate the distance between the intersection of the vertical line and the boundary of the optic nerve sheath. Due to the many interferences and limitations of manual measurement of the optic nerve sheath, the consistency and stability of the measurement results of different doctors and the same doctor at different time points are poor. Therefore, it is urgent to develop an algorithm that can automatically, stably and accurately measure ONSD.

[0007] In order to cope with the challenges in these key frame detection and optic nerve sheath measurement, the present invention introduces a fully automatic method for optimal frame detection and optic nerve sheath width measurement of fundus ultrasound video. This method aims to use the international expert consensus on optic nerve sheath diameter imaging and measurement to diagnose the optimal frame. Unlike the existing ultrasound video optimal frame detection algorithm, the method of the present application strictly follows the international expert consensus for feature selection and overall scoring, has strong interpretability, and does not require doctors to annotate and score each frame. The present invention is a fully automatic model, including three key modules: an automatic segmentation module of the eyeball and optic nerve sheath in the ultrasound image, an automatic detection module of the video key frame, and an automatic measurement module of ONSD 3 mm below the eyeball. During model training, the ultrasound video is split into image frames, and only the eyeball and optic nerve sheath of a small number of image frames need to be annotated. After the eyeball and optic nerve sheath are automatically segmented based on the trained segmentation model, it is only necessary to use the image's own information to automatically detect the optimal frame and automatically measure ONSD. The model of the present application is the first fully automatic method that strictly follows the international expert consensus to detect the optimal frame of the fundus ultrasound video and measure the width of the optic nerve sheath. Summary of the invention

[0008] Purpose of the invention: In view of the challenges faced by ultrasound video segmentation, optimal frame detection and optic nerve sheath width measurement, including a large amount of image noise and artifacts, blurred optic nerve sheath boundary contours and difficulty in quantifying the optimal frame standard, the present invention redesigns the optic nerve sheath width measurement task of fundus ultrasound video and proposes a new method for fully automatic fundus ultrasound video segmentation, optimal frame detection and optic nerve sheath width measurement. The model uses the automatic segmentation results of fundus images and the consensus of international experts to perform optimal frame detection and calculate the optic nerve sheath width 3 mm below the eyeball in the extracted optimal frame image.

[0009] Technical solution: A fully automatic method for optimal frame detection of fundus ultrasound video and measurement of optic nerve sheath width, comprising the following steps:

[0010] a) Literature research: Collect public literature and patents in related fields, including the current status of fundus ultrasound video processing, the current status of optimal frame detection in medical videos, and the current status of optic nerve sheath width measurement.

[0011] b) Data collection: 70 fundus ultrasound video data were collected from Nanjing Drum Tower Hospital, and 762 frames of images were randomly selected for pixel annotation of the eyeball and optic nerve sheath area; the collected data set was randomly divided into a training set and a test set for the segmentation model in a ratio of 8:2, and corresponding preprocessing was performed. In addition, 40 fundus ultrasound videos were labeled with 160 optimal frames and 160 unqualified frames selected by doctors, according to 3:

[0012] The data were randomly divided into training and test sets for the scoring model in a ratio of 2.

[0013] c) Establish an eyeball and optic nerve sheath segmentation model for fundus ultrasound images: Based on the nnU-Net segmentation network, an end-to-end 2D fundus ultrasound image segmentation model was constructed. The model was trained on a training set consisting of 609 images with eyeball and optic nerve sheath region annotations to achieve automatic segmentation of the eyeball and optic nerve sheath. After the training, the segmentation model was tested on an independent test set consisting of 153 frames of images.

[0014] d) Design an heuristic algorithm for optimal frame extraction of fundus ultrasound video: Based on the four qualitative standard rules in the international expert consensus on the measurement of optic nerve sheath width in fundus ultrasound video, extract the corresponding features based on the segmentation results and quantify the four rules, and design a linear regression scoring model. Based on the regression model algorithm, automatically determine the corresponding weight coefficient for each rule, obtain the score of each frame in the fundus ultrasound video, and output the five frames with the highest scores as the optimal frame set of the video.

[0015] e) Design an automatic measurement algorithm for the width of the optic nerve sheath: For the five best frame images extracted from each video, based on the segmentation results of the eyeball and the optic nerve sheath, calculate the center line of the optic nerve sheath, and locate the intersection of the center line and the eyeball boundary to 3 mm below the eyeball to automatically measure the width of the optic nerve sheath. After removing the maximum and minimum values ​​of the five width values, the average of the remaining three width values ​​is used as the final optic nerve sheath width.

[0016] f) Training the segmentation model: using an end-to-end training method to train and optimize the eyeball and optic nerve sheath segmentation model constructed in c) on the training set;

[0017] g) After training, the segmentation model obtained in f) is applied to the test set to segment all images in each patient's video. The original images and segmentation results are combined, and the algorithm constructed in d) is used to obtain the optimal frame set of the left and right eye ultrasound video of each patient. The algorithm in e) is used to obtain the final optic nerve sheath width of the video.

[0018] Specifically, the number of ultrasound video frames in step b) ranges from 22 to 501. The video frame images have two formats: the first format has an image size of 1024×768 and a pixel resolution of [0.0834 mm, 0.0834 mm]; the second format has an image size of 800×600 and a pixel resolution of [0.0652 mm, 0.0652 mm].

[0019] Specifically, two senior doctors from Gulou Hospital annotated the eyeball and optic nerve sheath at the pixel level for 762 2D ultrasound images of the 70 fundus ultrasound video data collected in step b), and then randomly divided them into training set and test set in a ratio of 609:153. The preprocessing of the data set is specifically implemented as follows: first, the pixel values ​​of all images are normalized to [0,1] by the Z-score method; second, the diversity of training data is increased by random horizontal flipping, random vertical flipping, adding random noise and other operations, and finally, the images in the training set are randomly cropped to 512×512 to improve training efficiency.

[0020] Specifically, the 40 fundus ultrasound videos with optimal frames and unqualified frames in step b) contain frame numbers of 4 optimal frames and 4 unqualified frames. The image data of these optimal frames and unqualified frames are used to train the weight coefficients in the linear regression model of the optimal frame score.

[0021] Specifically, the eyeball and optic nerve sheath segmentation model in step c) is divided into feature extraction, multi-resolution feature fusion, and adaptive mechanism training. The feature extraction adopts the nnU-Net framework, which uses a variant of the U-Net architecture for feature extraction, extracts low-order features of the image through the convolution layer and pooling layer in the encoder; uses a multi-level feature fusion strategy to combine shallow and deep features to achieve a balance between resolution and semantics, thereby improving segmentation accuracy; dynamic learning rate adjustment and data enhancement technology (random rotation, random flipping and random scaling) are used during training to enhance the generalization ability of the model. Among them, the loss function used for training the nnU-Net network is a combination of the average value of Dice loss and cross entropy (CE) loss. The final output is the segmentation result of the image, that is, the entire image is divided into three parts: background area, eyeball area and optic nerve sheath area. The trained nnU-Net segmentation model has a segmentation accuracy of Dice = 97.36% for the eyeball area and a segmentation accuracy of Dice = 90.05% for the optic nerve sheath on the test set.

[0022] Specifically, the optimal frame extraction algorithm for fundus ultrasound video in step d) needs to quantify the four rules in the international expert consensus on the measurement of optic nerve sheath width in fundus ultrasound images and then establish a scoring model. This application obtains the segmentation results of the eyeball and optic nerve sheath area for each frame image in the fundus ultrasound video based on the trained nnU-Net segmentation model. For the optimal frame selection rule 1: whether there is a lens and anterior chamber in the ultrasound image, which is a prerequisite for determining whether the frame image is the optimal frame, the quantification rules are as follows:

[0023]

[0024] Where I(x,y) represents the pixel value of the original image at position (x,y), M(x,y) represents the segmentation mask of the eyeball, which is 1 in the eyeball area and 0 in other areas. Represents the process of performing image erosion operation within the structural unit K.

[0025] For the optimal frame selection rule 2: whether the edge of the optic nerve sheath area is clear, the quantification rule is as follows:

[0026]

[0027] Where I(x,y) represents the pixel value of the original image at position (x,y), and N(x,y) represents the segmentation mask of the optic nerve sheath. Represents the structural unit K e The process of image erosion operation is performed inside. Represents the structural unit K dIn the process of image dilation operation, the mean grayscale values ​​of the image obtained by dilation and corrosion are subtracted. This difference reflects the brightness difference of the edge area of ​​the optic nerve sheath in the ultrasound image.

[0028] For optimal frame selection rule 3: whether there are two bright and three dark areas in the optic nerve sheath area in the image, where the "two bright" part is the arachnoid area in medicine, and the "three dark" part is the dura mater and optic nerve area. This application calculates the absolute difference of the pixel average value between the alternating areas of the "two bright" part and the "three dark" part. There are four such absolute differences, and they are summed to get var 3 The larger the summation result, the more significant the change in brightness.

[0029] For optimal frame selection rule 4: whether the optic nerve sheath is displayed in the vertical direction, this application considers the angle between the center line of the optic nerve sheath and the vertical direction of the image. The smaller the angle, the greater the absolute value of the center line slope, and the closer the optic nerve sheath is to vertical; the larger the angle, the smaller the absolute value of the center line slope, and the more inclined the optic nerve sheath, so var 4 It is the angle between the center line of the nerve sheath and the vertical direction of the image.

[0030] After quantifying the four rules, we can get the corresponding four variable values. This application de-dimensionalizes these four variables, and the final scoring model is as follows:

[0031] Var=w 1 ×var 1 +w 2 ×var 2 +w 3 ×var 3 +w 4 ×var 4

[0032] Based on the classification and annotation of the best frames and unqualified frames in 40 fundus ultrasound videos by doctors, this application randomly selected 24 cases as the training set of the scoring model, and determined the four weight coefficients in the above scoring model through the LDA (Linear Discriminant Analysis) method, where w 1 =-0.2873, w 2 =0.3585,w 3 =0.1115,w 4 =-0.1179.

[0033] Finally, this application sorts all image frames in the fundus ultrasound video from high to low according to the score of the optimal frame scoring model, and selects the top five frames with the highest scores as the optimal frame set of the video. In the test set of all 16 fundus ultrasound videos with optimal frame annotations, this application compares the top five frames selected by the above scoring model with the doctor's optimal frame annotations, and finds that the number of cases in which the frame number with the highest score predicted by the model differs by no more than 2 from the optimal frame number annotated by the doctor accounts for 52.17% of the total test set (i.e., the indicator top1). The number of cases in which at least one of the three frames with the highest scores selected by the above scoring model differs by no more than 2 from the optimal frame number annotated by the doctor accounts for 86.96% of the total test set (i.e., the indicator top3). The number of cases in which at least one of the top five frames with the highest scores selected by the above scoring model differs by no more than 2 from the optimal frame number annotated by the doctor accounts for 100% of the total test set (i.e., the indicator top5).

[0034] Specifically, the calculation process of the automatic measurement algorithm of the optic nerve sheath width in the fundus ultrasound video in step e) is as follows: first, the optic nerve sheath centerline l is extracted according to the segmentation results of the five best frame images with the highest scores selected by the previous scoring model. 1 , and calculate the intersection point x of the center line and the eyeball 1 Then, according to the pixel resolution of the image, calculate the intersection point x 1 Point x 3 mm below the center line of the optic nerve sheath 2 At point x 2 Center line l 1 The perpendicular line l 2 , then calculate l 2 x of the intersection with the edge of the optic nerve sheath 3 and x 4 , and finally calculate x 3 and x 4 For the five frames in the optimal frame set of the video, the optic nerve sheath width values ​​are calculated according to the above process, and the maximum value and the minimum value are removed, and the remaining three values ​​are averaged as the final optic nerve sheath width value.

[0035] Beneficial effects: The present invention uses the automatic segmentation results of the eyeball and optic nerve sheath in the fundus ultrasound image and the consensus of international experts to detect the optimal frame of the ultrasound video, and automatically measures the width of the optic nerve sheath 3 mm below the eyeball from the extracted optimal frame image. It is worth noting that the optimal frame extraction algorithm of the present invention is a mathematical modeling and quantification of medical diagnostic standards. Different from other existing scoring methods based on deep network models, it is an interpretable quantitative scoring method, and does not require doctors to mark all the optimal frames of the images in the fundus ultrasound video, but only needs to provide a small number of typical optimal frames and unqualified frames.

[0036] Figure content

[0037] Figure 1 It is the overall framework diagram of the method proposed in the present invention;

[0038] Figure 2 A comparison diagram of the segmentation results and the doctor's annotations in the framework of the method proposed in the present invention, wherein the first column is the original image frame, the second column is the original image frame and the doctor's annotations, and the third column is the original image frame and the segmentation results;

[0039] Figure 3 The visualization results of the quantitative indicators in the framework of the fundus ultrasound optimal frame detection method proposed in the present invention, (a) is the visualization of standard plane selection rule 1: "Whether there is a lens and anterior chamber in the ultrasound image", (b) is the visualization of standard plane selection rule 2: "Whether the edge of the optic nerve sheath area is clear", (c) is the visualization of standard plane selection rule 3: "Whether there are two bright and three dark areas alternating in the optic nerve sheath area", (d) is the visualization of standard plane selection rule 4: "Whether the optic nerve sheath is displayed in the vertical direction";

[0040] Figure 4 It is a visualization diagram of the contribution of four quantitative indicators in the fundus ultrasound image scoring model proposed by the present invention to the model;

[0041] Figure 5 This is a visualization diagram of the method for automatically measuring the width of the optic nerve sheath in fundus ultrasound images proposed in the present invention. DETAILED DESCRIPTION

[0042] The objectives, technical solutions and advantages of the present invention are described in detail below in conjunction with the accompanying drawings and specific embodiments of the present invention.

[0043] The present invention proposes a fully automatic method for optimal frame detection and optic nerve sheath width measurement of fundus ultrasound video, which can automatically segment the eyeball and optic nerve sheath area in the fundus ultrasound image, and then automatically score each frame image in the fundus ultrasound video based on the segmentation result and four rules in the quantitative international expert consensus, and then select the 5 frames with the highest scores as the optimal frame set, and finally automatically measure the optic nerve sheath width 3 mm below the eyeball in the selected 5 optimal frame images, remove the maximum and minimum width values, and average the remaining three width values ​​as the final optic nerve sheath width value. The overall flow chart of the present invention is as follows Figure 1 As shown, the specific steps include the following steps.

[0044] a) Data collection: 70 fundus ultrasound video data were collected from Nanjing Drum Tower Hospital. For each video, it was first divided into image frames in chronological order, and all image frames were converted into grayscale images. Then doctors were asked to select and annotate the fundus area and optic nerve sheath area in 762 frames of images as the data set for the deep network segmentation model. These annotated images were randomly divided into training and test sets in a ratio of 8:2.

[0045] b) Establishing the eyeball and optic nerve sheath segmentation model: Based on the nnU-Net segmentation network, an end-to-end 2D fundus ultrasound image segmentation model was constructed. The parameters used for model training are shown in Table 1. After training, the eyeball and optic nerve sheath were segmented for all image frames of all patients to extract the optimal frame later.

[0046] Table 1 Parameters used in model training

[0047] Batch size 13 Initial epochs 1000 Patch size (448,576) Initial lr 0.01 Weightdecay 0.0003

[0048] c) Design a heuristic video optimal frame extraction algorithm: Based on the segmentation results of all image frames and the international consensus statement standards, the following four rule quantification indicators and the final optic nerve sheath detection indicators are designed. Figure 3 Visualization of the information quantified for each metric:

[0049] var 1 To quantify rule 1: "Whether there is a lens and anterior chamber in the ultrasound image",

[0050]

[0051] var 2 To quantify rule 2: "Is the edge of the optic nerve sheath area clear?"

[0052]

[0053] var 3 To quantify rule 3: "Whether there are two bright and three dark areas alternating in the optic nerve sheath area", var 4 To quantify rule 4: "Whether the optic nerve sheath is displayed in the vertical direction". Then, through linear discriminant analysis

[0054] (LDA) method, based on a small number of optimal frames and unqualified frames of the video annotated by doctors, four weight coefficients were determined. Finally, the scoring rules of this application are as follows:

[0055] Var=-0.2873×var 1 +0.3585×var 2 +0.11.15×var 3 -0.1179×var4

[0056] All image frames of the fundus ultrasound video are scored according to the final indicator Var, and the five frames with the highest scores are selected as the optimal frame set of the video.

[0057] d) Design of optic nerve sheath width measurement algorithm: For each optimal frame set selected in fundus ultrasound video, the present application automatically calculates the optic nerve sheath width in five frames based on the segmentation results. Subsequently, the maximum and minimum values ​​of the five width values ​​are removed, and the remaining three values ​​are averaged as the final optic nerve sheath width. The visualization result is as follows: Figure 3 shown.

[0058] The present invention first obtains a 2D fundus ultrasound dataset by processing fundus ultrasound videos and adding the doctor's annotation results of the eyeball and optic nerve sheath, and trains an automatic segmentation model of the eyeball and optic nerve sheath in fundus ultrasound images based on the nnU-Net deep network. The segmentation result of the segmentation model for the eyeball area on the test dataset is Dice = 97.36%, and the segmentation result for the optic nerve sheath part is Dice = 90.05%. Then, the present application uses the segmentation model to segment all image frames of all ultrasound videos, and designs a quantifiable and interpretable scoring model based on the international expert consensus on the measurement of the width of the optic nerve sheath in fundus ultrasound images. This scoring model is used to extract the five frames of images with the highest scores in each video as the optimal frame set. Compared with the optimal frame set selected by the doctor, the five frames of images selected by the optimal frame extraction algorithm of the present application have a top1 detection accuracy of 52.17%, a top3 detection accuracy of 86.96%, and a top5 detection accuracy of 100%. Finally, the application automatically calculates the width of the optic nerve sheath in the best five frames of images in each video, and then removes the highest and lowest scores to calculate the trimmed average value as the optic nerve sheath width measured in the fundus ultrasound video.

Claims

1. A fully automatic method for optimal frame detection of fundus ultrasound video and measurement of optic nerve sheath width, characterized in that: The steps include: a) Data collection: Collect fundus ultrasound video data; randomly divide the collected data into training set and test set according to a certain ratio, and perform corresponding preprocessing; b) Establish an eyeball and optic nerve sheath segmentation model: Based on the nnU-Net segmentation network, an end-to-end 2D ultrasound image segmentation model is constructed to automatically segment the eyeball and optic nerve sheath; c) Design a heuristic algorithm for extracting optimal video frames: Based on the standard plane selection rules for ultrasound video, quantify the rules and score each frame of each video, and output the five frames with the highest scores as the optimal frame set; d) Design an optic nerve sheath width measurement algorithm: Based on the original image information and segmentation results, the optic nerve sheath width at 3 mm below the eyeball of each image in the optimal frame set is calculated using the optic nerve sheath centerline, and then the trimmed average of the optic nerve sheath width values ​​of all images in the optimal frame set is calculated as the optic nerve sheath width measurement value in the video; e) Training the segmentation model: using an end-to-end training method to train and optimize the eyeball and optic nerve sheath segmentation model constructed in b) on the training set; f) After training, the segmentation model obtained in e) is applied to the test set to segment all images in each patient's video. The original images and segmentation results are combined, and the algorithm constructed in c) is used to obtain the optimal frame set of the left and right eye ultrasound video of each patient. The algorithm in d) is used to obtain the final optic nerve sheath width of the video.

2. The fully automatic method for optimal frame detection of fundus ultrasound video and measurement of optic nerve sheath width according to claim 1, characterized in that: In step a), the collected data is randomly divided into a training set and a test set of the segmentation model in a ratio of 8:2; the preprocessing is specifically as follows: first, the pixel values ​​of all fundus ultrasound video frame images are normalized to [0, 1] by the Z-score method; secondly, random horizontal flipping, random vertical flipping and random noise are performed, and finally, the fundus ultrasound video frame images in the training set are randomly cropped to 512×512.

3. The fully automatic method for optimal frame detection of fundus ultrasound video and measurement of optic nerve sheath width according to claim 1, characterized in that: In step b), the loss function used to train the nnU-Net network is a combination of the average value of the Dice loss and the cross entropy CE loss; the final output is the image segmentation result, that is, the entire image is divided into three parts: background area, eyeball area and optic nerve sheath area.

4. The fully automatic method for optimal frame detection of fundus ultrasound video and measurement of optic nerve sheath width according to claim 1, characterized in that: In step c), the heuristic video optimal frame extraction algorithm includes quantifying four ultrasound video standard plane selection rules in the international expert consensus and scoring each frame of the video; Among them, the four ultrasound video standard plane selection rules and corresponding quantization rules are as follows: Optimal frame selection rule 1: whether there is a lens and anterior chamber in the fundus ultrasound video frame image; Where I(x,y) represents the pixel value of the original image at position (x,y), M(x,y) represents the segmentation mask of the eyeball, which is 1 in the eyeball area and 0 in other areas. Represents the process of image erosion operation in the structural unit K; Optimal frame selection rule 2: whether the edge of the optic nerve sheath area is clear; Where I(x,y) represents the pixel value of the original image at position (x,y), N(x,y) represents the segmentation mask of the optic nerve sheath, Represents the structural unit K e The process of image erosion operation is performed inside. Represents the structural unit K d The process of image dilation operation is performed in the image, and the mean grayscale values ​​of the image obtained by dilation and corrosion are subtracted. This difference reflects the brightness difference of the edge area of ​​the optic nerve sheath in the ultrasound image. Optimal frame selection rule 3: whether there are two bright and three dark areas alternating in the optic nerve sheath area in the image; The "two bright" parts are the arachnoid area in medicine, and the "three dark" parts are the dura mater and optic nerve area. The absolute difference between the average values ​​of pixels in the alternating areas of the "two bright" parts and the "three dark" parts is calculated. There are four absolute differences, and they are summed to get var3. The larger the sum result, the more significant the change in brightness and darkness. Optimal frame selection rule 4: whether the optic nerve sheath is displayed in the vertical direction; Consider the angle between the center line of the optic nerve sheath and the vertical direction of the image. The smaller the angle, the larger the absolute value of the center line slope, and the closer the optic nerve sheath is to vertical; the larger the angle, the smaller the absolute value of the center line slope, and the more inclined the optic nerve sheath. Var4 is the angle between the center line of the nerve sheath and the vertical direction of the image. The scoring formula is as follows: Var=w1×var1+w2×var2+w3×var3+w4×var4 Wherein, w1, w2, w3 and w4 are the weight coefficients of the four quantization rules: "whether there is a lens and anterior chamber in the ultrasound image", "whether the edge of the optic nerve sheath area is clear", "whether there are two bright and three dark areas alternating in the optic nerve sheath area", and "whether the optic nerve sheath is displayed in the vertical direction".

5. The fully automatic method for optimal frame detection of fundus ultrasound video and measurement of optic nerve sheath width according to claim 1, characterized in that: In step d), the optic nerve sheath width measurement algorithm calculates the center line of the optic nerve sheath based on the optimal frame of the patient input, locates the intersection of the center line and the boundary of the eyeball to 3 mm below the eyeball, and automatically measures the width of the optic nerve sheath; after removing the maximum and minimum values ​​of the five width values, the average of the remaining three width values ​​is used as the final optic nerve sheath width.

Images