Optic nerve sheath ultrasound image segmentation method, system and medical device

Abstract

The application discloses an optic nerve sheath ultrasonic image segmentation method and system and medical equipment. The application is based on a YOLOV5s model, a brand-new feature fusion is designed for the optic nerve sheath ultrasonic image, a C3STR module containing a WTA structure, a BOT3 module containing an MHSA structure and a CRTN module containing a branch structure and a reversible block operation are combined in the YOLOv5s model, the window attention mechanism of the C3STR module and the multi-head self-attention mechanism of the BOT3 module are used to capture the global features of the optic nerve sheath image, in addition, the multi-layer nonlinear transformation and the residual structure in the CRTN module are used to better capture the fuzzy boundary of the optic nerve sheath, so that the network is more accurate for the optic nerve sheath segmentation under a complex image. The fusion of the local features and the global features of the input image is realized, and the segmentation of the optic nerve sheath ultrasonic image is realized.

Description

Technical Field

[0001] This invention relates to the field of medical segmentation, specifically to a method, system, and medical device for segmenting ultrasound images of the optic nerve sheath. Background Technology

[0002] Elevated intracranial pressure (ICP) is a common but challenging and potentially fatal complication in neurosurgery, neurology, pediatrics, and ophthalmology. It is typically caused by acute traumatic brain injury, intracranial tumors, cerebral hemorrhage, hydrocephalus, and intracranial inflammation, requiring timely intervention to prevent adverse consequences. Therefore, physicians need to accurately and quickly differentiate between patients with symptoms of elevated ICP and those without. In recent years, technological advancements have led to various methods for measuring ICP, including non-invasive and invasive approaches. Invasive methods generally involve directly measuring intracranial pressure by implanting sensors, such as strain gauges, pneumatic sensors, and fiber optic sensors. While highly accurate and reliable, invasive methods carry risks of local wound infection and bleeding. In contrast, non-invasive methods, although less intuitive, are sufficiently safe and are more appropriate for patients with milder symptoms. Non-invasive methods include transcranial Doppler ultrasound (TCD) and measurement of the optic nerve sheath diameter (ONSD). The optic nerve sheath is a thin membrane that surrounds the optic nerve, and its main function is to provide protection and support. Increased intracranial pressure can lead to optic nerve sheath edema, which is the accumulation and expansion of fluid within the optic nerve sheath. Therefore, observing and assessing changes in the optic nerve sheath is crucial for evaluating the state of intracranial pressure in intracranial pressure monitoring and the diagnosis of related diseases.

[0003] Due to the inherent subjectivity and operator dependence of ultrasound technology, acquiring standard ONSD images and performing accurate measurements is a challenging and time-consuming task. Artificial intelligence-based image segmentation can improve the accuracy and reliability of ultrasound diagnosis and reduce doctors' image interpretation time. With the rapid development of artificial intelligence technology in recent years, many new deep learning-based methods have emerged in medical segmentation. Medical segmentation techniques based on Convolutional Neural Networks (CNNs) are becoming increasingly mature in model design. CNNs can extract multi-level features from the original image and gradually abstract detailed information from the image through layer-by-layer propagation. In recent years, some improved CNN models, such as U-Net, VGG, and ResNet, have achieved excellent results in medical image segmentation. In addition, graph cut-based methods, Gaussian mixture models, and level set methods have also been widely used in medical image segmentation. However, since the design of these CNN networks focuses on local features, especially in high-resolution image segmentation tasks, these segmentation algorithms struggle to accurately identify and segment target regions when faced with the complexity of medical images. Summary of the Invention

[0004] This addresses the challenge of segmenting the optic nerve sheath in patient ocular ultrasound images due to the diversity of acquisition angles, acquisition instruments, and individual patient differences.

[0005] This invention proposes a method, system, and medical device for optic nerve sheath segmentation based on CBC-YOLOv5s. Based on the YOLOv5s model, this invention designs a novel feature fusion method for optic nerve sheath ultrasound images. The YOLOv5s model incorporates a C3STR module with a WTA structure, a BOT3 module with an MHSA structure, and a CRTN module with branching structures and reversible block operations, achieving the fusion of local and global features of the input image. This addresses the difficulty of optic nerve sheath segmentation in the lower quadrant.

[0006] The technical solution adopted in this invention includes the following steps:

[0007] S1. Collect eye data:

[0008] Ultrasound videos of the eyes of different patients were acquired using an ultrasound instrument. The ultrasound videos were then processed by frame extraction to form an ultrasound image dataset. The dataset was manually labeled, with the labeling content being the lower half of the orbit and the optic nerve sheath. The labeled dataset was then divided into a training set and a test set.

[0009] S2. Data processing of the optic nerve sheath portion under complex ocular ultrasound images based on the CBC-YOLOV5s network:

[0010] S2-1. In the backbone network part for feature extraction, a five-layer convolutional module is used to progressively obtain feature maps containing local features of the input ocular ultrasound image.

[0011] S2-2. Between the second and third convolutional modules, and between the third and fourth convolutional modules, C3STR modules are introduced respectively. Each C3STR module consists of a single-layer C3, a double-layer LN, a single-layer WTA, a single-layer MLP, and two residual structures. The characteristics of the C3, WTA, and MLP modules are used to mine the local and global features of the optic nerve sheath. The third convolutional module performs a convolution operation on the feature map containing local and global features output by the first C3STR module. The convolved feature map is then passed through the second C3STR module to extract more local and global features of the optic nerve sheath. Finally, the fourth convolutional module performs a convolution on the feature map output by the second C3STR module, outputting the first feature map after fusing the two local and global features.

[0012] S2-3. Between the fourth and fifth convolutional layers, and between the fifth and SPPF layers, BOT3 modules are introduced. Each BOT3 module contains an MHSA structure. The first BOT3 module uses the MHSA structure to project the feature map output from the fourth convolutional layer onto multiple attention heads, each of which performs self-attention computation. The outputs of each attention head are concatenated and weighted through an additional linear transformation. Finally, a second feature map with the same scale as the output feature map is obtained through matrix transformation. The second feature map then passes through the fifth convolutional layer and the second BOT3 layer before entering the SPPF module. Since each attention head can focus on different ranges of contextual information, the MHSA structure can better capture long-range dependencies in the feature map.

[0013] S2-4. At the tail of the Backbone feature extraction backbone network, the SPPF module is used to further fuse the local and global features in the feature map output by the second layer BOT3 module to obtain the third feature map.

[0014] S2-5. In the upsampling feature fusion part of the feature pyramid, the third feature map output by the SPPF module is first convolved and upsampled by convolution module I to obtain the fourth feature map. The scale of the fourth feature map is matched with the scale of the second feature map output by the first layer BOT3 module in the feature extraction backbone network. Then, the CONCAT module is used to concatenate the two scale-matched feature maps to obtain the concatenated feature map. Figure I ;

[0015] S2-6. Splicing Features Figure IAfter processing by a C3STR module, the fifth feature map is obtained through a convolution module II, which performs a convolution and upsampling operation. The scale of the fifth feature map is then matched with the scale of the first feature map output by the second layer of the C3STR module in the feature extraction backbone network. Finally, the CONCAT module concatenates the two scale-matched feature maps to obtain the concatenated feature map. Figure II ;

[0016] S2-7. Feature pyramid downsampling feature fusion part, which includes three CRTN modules and two convolutional modules, first concatenates the features. Figure II In the first CRTN module, the CRTN module utilizes its branching structure and reversible block operations to concatenate the input features. Figure II The global features are further fused to obtain the fused features. Figure I ;

[0017] S2-8. Fusion features after processing by a CRTN module Figure I Perform convolution operations and make the convolutioned features Figure VI The scale is consistent with the feature map VII output by convolutional module II in the upsampling feature fusion part, and then the features are... Figure VI Feature map VII is fed into the CONCAT module for feature concatenation, and the concatenated feature map is then fed into the CRTN module for further feature fusion to obtain the fused features. Figure II ;

[0018] S2-9. Fusion features after processing by two CRTN modules Figure II A convolution operation is performed, ensuring that the scale of the convolutional feature map VIII is consistent with that of feature map IX output from convolutional module I in the upsampling feature fusion part. Then, feature maps VIII and IX are fed into the CONCAT module for feature concatenation. The concatenated feature map is then input into the CRTN module for further feature fusion to obtain the fused feature map. Figure III ;

[0019] S2-10. Using three segment dividers to fuse features at different scales from the outputs of three CRTN modules. Figure I Fusion characteristics Figure II and fusion features Figure III Perform semantic segmentation.

[0020] S3. Construct a loss function using the training set, and train the CBC-YOLOV5s network using the loss function. After N training rounds, obtain the best-performing CBC-YOLOV5s network.

[0021] S4. Input the data from the test set described in step S1 into the trained CBC-YOLOV5s network, extract the optic nerve sheath image features and perform semantic segmentation.

[0022] A lower optic nerve sheath ultrasound image segmentation system includes a data acquisition module, a CBC-YOLOV5s network module, a training module, and a testing module;

[0023] Data acquisition module: Collects eye data and divides it into training and test sets;

[0024] CBC-YOLOV5s network module: Based on the YOLOV5s model, it designs a new feature fusion for optic nerve sheath ultrasound images. It combines the C3STR module containing the WTA structure, the BOT3 module containing the MHSA structure, and the CRTN module containing the branch structure and reversible block operation in the YOLOv5s model to achieve the fusion of local and global features of the input image and perform data processing on the optic nerve sheath part under complex ocular ultrasound images.

[0025] Training module: Construct a loss function using the training set, and train the CBC-YOLOV5s network using the loss function. After N training rounds, obtain the best-performing CBC-YOLOV5s network.

[0026] Test module: Input the test set into the trained CBC-YOLOV5s network, extract optic nerve sheath image features and perform semantic segmentation.

[0027] A medical device incorporating the aforementioned ultrasound image segmentation method for the lower optic nerve sheath.

[0028] The beneficial effects of this invention are as follows:

[0029] This invention proposes a CBC-YOLOV5s-based method for optic nerve sheath ultrasound image segmentation. When using Convolutional Neural Networks (CNNs) to extract optic nerve sheath features, global contextual information is often ignored, resulting in the extraction of only local features of the optic nerve sheath image. While YOLOV5s networks also focus on local features, they offer greater flexibility and scalability, allowing for the overemphasis on local features to be compensated for by modifying the network structure. This invention introduces C3STR, BOT3, and CRTN modules. The window attention mechanism of the C3STR module and the multi-head self-attention mechanism of the BOT3 module are used to capture global features of the optic nerve sheath image. Furthermore, the multi-layer nonlinear transformation and residual structure in the CRTN module are used to better capture the blurred boundaries of the optic nerve sheath, resulting in more accurate optic nerve sheath segmentation.

[0030] This invention proposes a method for optic nerve sheath ultrasound image segmentation based on CBC-YOLOV5s, which can effectively solve the problem of optic nerve sheath segmentation and achieve excellent segmentation performance in semantic segmentation. Attached Figure Description

[0031] Figure 1 The structure of the CBC-YOLOV5s optic nerve sheath segmentation algorithm;

[0032] Figure 2 Here is a structural diagram of the C3STR module;

[0033] Figure 3 This is a diagram of the BOT3 module structure.

[0034] Figure 4 Here is a structural diagram of the CRTN module;

[0035] Figure 5 Annotate before-and-after comparison images of ocular ultrasound images;

[0036] Figure 6 Heatmaps of different models under normal and special images. Detailed Implementation

[0037] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0038] This embodiment uses a CBC-YOLOV5s-based ultrasound image segmentation method for the optic nerve sheath, including the following steps:

[0039] S1. Constructing the dataset

[0040] S1-1: Obtain ocular ultrasound images from ocular ultrasound videos of different patients to form the dataset Z2HOSPITAL-5000 required for this invention. The dataset contains 5000 ocular ultrasound images, of which 3834 are used as the training set, 426 as the validation set, and 740 as the test set, including 342 images of poor quality.

[0041] S1-2: Manually annotate the above dataset, focusing on the lower half of the orbit and the optic nerve sheath. Z2HOSPITAL-5000 partial ocular ultrasound images and annotated images are shown below. Figure 5 As shown.

[0042] S2. Data processing of the optic nerve sheath portion under complex ocular ultrasound images based on the CBC-YOLOV5s network.

[0043] The basic framework of the method proposed in this invention is as follows: Figure 1 As shown.

[0044] Figure 1The diagram shows the basic framework of the method of the present invention. The input image first passes through a five-layer convolutional module, a two-layer C3STR module, a two-layer BOT3 module, and a single-layer SPPF module in the feature extraction backbone network. C3STR modules are introduced between the second and third layers and between the third and fourth convolutional modules to initially mine the global features of the optic nerve sheath.

[0045] The first feature map output by the second C3STR module is fed into the first BOT3 module after passing through another convolutional module. The multi-head self-attention mechanism of the BOT3 module is used to capture and output a second feature map containing global features of the optic nerve sheath. The second feature map is then passed to another convolutional module.

[0046] The second feature map is input into the second layer BOT3 module after passing through the convolution module. The BOT3 module continues to mine feature maps containing global features of the optic nerve sheath, and the output feature map is passed to the SPPF module. The SPPF module is used to achieve deep fusion of global and local features to obtain the fused third feature map.

[0047] The third feature map, fused by the SPPF module, is input into the upsampling feature fusion part of the feature pyramid. After passing through convolution module I and upsampling operations, the scale of the resulting fourth feature map is matched with the scale of the second feature map in the backbone network, and then concatenated using the CONCAT module. Figure I ;

[0048] splicing features Figure I After passing through a C3STR module and a convolutional module II, the fifth feature map is obtained through an upsampling module, while maintaining the same scale as the first feature map. This image is then input into the CONCAT module for concatenating the two feature maps, resulting in the concatenated feature map. Figure II ;

[0049] In the feature pyramid downsampling and feature fusion section, features are concatenated. Figure II The first-layer CRTN module is input, and the CRTN module utilizes its branching structure and reversible block operations to concatenate the input features. Figure II The global features are further fused to obtain the fused features. Figure I ;

[0050] Fusion features Figure I Features obtained after passing through a convolutional module Figure VI The feature map VII output from convolutional module I is concatenated with the feature map VII, and the concatenated feature map is then fed into the second-layer CRTN module to output the fused feature map. Figure II ;

[0051] The fusion features output by the second-layer CRTN module Figure IIPerform another convolution operation and concatenate the output feature map VIII with the output feature map IX of convolution module I. The concatenated feature map is then fed into the third-layer CRTN module to obtain the fused features. Figure III ;

[0052] Finally, the fused features at different scales output by the three CRTN modules are processed using three segmentation dividers. Figure I Fusion characteristics Figure II and fusion features Figure III Perform semantic segmentation.

[0053] C3STR module

[0054] like Figure 2 As shown, the C3STR module consists of a single-layer C3, a two-layer LN, a single-layer WTA, a single-layer MLP, and two residual structures. First, the C3 module uses convolution to fuse optic nerve sheath features and performs LayerNorm normalization on the feature maps. Second, the WTA module combines window attention (W-MSA) and shifted window attention (SW-MSA). The W-MSA mechanism divides the input into local windows and performs self-attention calculations on these windows to facilitate the interaction between local and global information. However, using W-MSA alone can sometimes be limited by the window size and fail to capture the global features of the optic nerve sheath in ocular ultrasound images. Therefore, SW-MSA is introduced, which shifts the window by a fixed size within each time frame, helping the model better understand the entire ocular ultrasound image and better capture the global features of the optic nerve sheath. Simultaneously, the fully connected layers of the MLP module adjust and reorganize the features after interaction with the WTA module, helping the model better utilize the optic nerve sheath features acquired through the attention mechanism. Finally, the optic nerve sheath features processed by the first residual connection are combined with the optic nerve sheath features extracted by the MLP module to perform a second residual connection, further increasing the network depth.

[0055] BOT3 module

[0056] like Figure 3As shown, the BOT3 module consists of three layers: CBS, BT, and CONCAT. The feature map of the input module is divided into two branches. One branch extracts features through CBS and then passes through the BT module, utilizing the MHSA (Multi-Head Self-Attention) module included in the BT module to capture the global relationships between features. The other branch only extracts features through the CBS module and then concatenates them with the features processed by the previous branch. The BOT3 module achieves deep processing of optic nerve sheath features and captures the global feature relationships of the optic nerve sheath through the attention mechanism implemented by the MHSA module, which helps to obtain optic nerve sheath features more comprehensively and accurately.

[0057] CRTN module

[0058] like Figure 4 As shown, the CRTN module consists of two CBS modules incorporating the Swish activation function and a three-layer BBR module. After feature maps are extracted through a CBS structure, they split into two branches. One branch enters the three-layer BBR structure, while the other branch concatenates the feature maps that have undergone one to three BBR module operations. Finally, the concatenated feature map is fed back into the CBS structure for a second feature extraction. Through its branching structure and reversible block operations, the CRTN module achieves cross-stage information transfer and efficient feature learning, which helps to better acquire the global features of the optic nerve sheath.

[0059] S3. Train the entire model in 100 rounds, and finally retain the model with the best performance on the validation set.

[0060] S4. Input the test set images into the trained model to extract optic nerve sheath features and perform semantic segmentation.

[0061] The experimental server used in this invention employed an NVIDIA TITAN RTX 4090 GPU, and the deep learning framework used was PyTorch. The input images were 640*640 resolution 3-channel ocular ultrasound images. The training batch size was 16, and the training run consisted of 100 epochs. The loss function was optimized using a Stochastic Gradient Descent (SGD) optimizer with a learning rate of 10. -2 The weight decay factor is 10. -5 .

[0062] Example

[0063] Using the CBC-YOLOV5s optic nerve sheath segmentation algorithm, YOLOV5s algorithm, and U-Net model proposed in this invention, visualization heatmaps were plotted on the features extracted by the visualization model from 740 total test images and 342 specific image test images in the Z2HOSPITAL-5000 dataset, as shown below. Figure 6 As shown in the figure, compared with the other two algorithms, the optic nerve sheath segmentation algorithm proposed in this invention can correctly focus on the optic nerve sheath and the lower half of the orbit in both normal test sets and special test sets, and suppress background interference.

[0064] The following is an experimental data analysis of the algorithm proposed in this invention based on a partitioned test set database. The performance of this invention is compared with that of the YOLOv5s algorithm and the U-Net algorithm on the total test set and specific test sets of the Z2HOSPITAL-5000 dataset.

[0065] To validate the superiority of our model, we evaluated it using four metrics: precision, recall, intersection-over-union (IoU), and model memory usage. Precision refers to the proportion of samples predicted as positive by the model that are actually positive. It measures the model's accuracy in predicting positive examples, and the formula is as follows:

[0066]

[0067] TP stands for True Positives, and FP stands for False Positives.

[0068] Recall is the proportion of true positive samples that are correctly predicted as positive by the model. It measures the model's coverage of positive samples, and is calculated using the following formula:

[0069]

[0070] FN stands for False Negative.

[0071] The intersection-union ratio (IUGR) is a metric used to measure the degree of overlap between the predicted results and the ground truth annotations of an object detection or semantic segmentation model. It evaluates model performance by calculating the ratio of the intersection to the union of the predicted and ground truth regions. The formula is as follows:

[0072]

[0073] Wherein, Intersection represents the area of ​​the intersection between the predicted region and the ground truth labeled region, and Union represents the area of ​​the union of the two.

[0074] The segmentation performance of the three algorithms on the total test set and specific test sets is shown in Table 1 and Table 2.

[0075] Table 1. Top-1 performance comparison of different algorithms on the total dataset.

[0076] U-Net 85.1 78.3 68.8 YOLOV5s-seg 90.5 92.6 84.4 Algorithm of this invention 92.3 93 86.3

[0077] Analysis of the segmentation results of the Z2HOSPITAL-5000 test set: Compared with the traditional YOLOV5s-seg network, the precision, recall, and crossover ratio of the algorithm of this invention are 1.8%, 0.4%, and 2.1% higher, respectively; compared with the U-Net network, the precision, recall, and crossover ratio of the algorithm of this invention are 6.8%, 14.7%, and 17.5% higher, respectively.

[0078] Table 2 Comparison of segmentation performance of different algorithms on a specific test set.

[0079]

[0080]

[0081] Analysis of the segmentation results on the Z2HOSPITAL-5000 special test set: Compared with the traditional YOLOV5s-seg network, the precision, recall, and crossover ratio of the algorithm of this invention are 4.1%, 2.1%, and 4.5% higher, respectively; compared with the U-Net network, the precision, recall, and crossover ratio of the algorithm of this invention are 9.2%, 21%, and 19.7% higher, respectively.

[0082] This invention addresses the challenge of optic nerve sheath segmentation in patient ocular ultrasound images due to the diversity of acquisition angles, acquisition instruments, and individual patient differences. It proposes an optic nerve sheath segmentation algorithm based on CBC-YOLOV5s, which fully integrates information from both local and global features of the optic nerve sheath image, effectively improving segmentation performance compared to current mainstream models. This demonstrates the effectiveness of the proposed method.

[0083] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A method for segmenting ultrasound images of the optic nerve sheath, characterized in that, Based on the YOLOv5s model, a novel feature fusion method is designed for ultrasound images of the optic nerve sheath. This method combines the C3STR module (containing WTA structure), the BOT3 module (containing MHSA structure), and the CRTN module (containing branching structure and reversible block operations) within the YOLOv5s model to achieve the fusion of local and global features of the input image, thus enabling optic nerve sheath segmentation. The specific steps include: S1. Collect eye data and divide it into training and test sets; S2. Data processing of the optic nerve sheath in complex ocular ultrasound images based on the CBC-YOLOV5s network; S3. Construct a loss function using the training set, and train the CBC-YOLOV5s network using the loss function. After N training rounds, obtain the best-performing CBC-YOLOV5s network. S4. Input the test set into the trained CBC-YOLOV5s network, extract the optic nerve sheath image features and perform semantic segmentation; The CBC-YOLOV5s network described in step S2 includes a backbone feature extraction network, an upsampled feature fusion part of the feature pyramid, and a downsampled feature fusion part of the feature pyramid. The specific implementation of the backbone network part for feature extraction is as follows: S2-1. In the backbone network part of the feature extraction, a five-layer convolutional module is used to gradually obtain feature maps containing local features of the input ocular ultrasound image; S2-2. Between the second and third convolutional modules, and between the third and fourth convolutional modules, C3STR modules are introduced respectively. Each C3STR module consists of a single-layer C3, a double-layer LN, a single-layer WTA, a single-layer MLP, and two residual structures. The characteristics of the C3, WTA, and MLP modules are used to mine the local and global features of the optic nerve sheath. The third convolutional module performs a convolution operation on the feature map containing local and global features output by the first C3STR module. The convolved feature map is then passed through the second C3STR module to extract more local and global features of the optic nerve sheath. Finally, the fourth convolutional module performs a convolution on the feature map output by the second C3STR module, outputting the first feature map after fusing the two local and global features. S2-3. Between the fourth and fifth convolutional modules, and between the fifth convolutional module and the SPPF module, BOT3 modules are introduced respectively. Each BOT3 module contains an MHSA structure. The first BOT3 module uses the MHSA structure to project the feature map output by the fourth convolutional module onto multiple attention heads. Each attention head performs self-attention calculation. The outputs of each attention head are concatenated and weighted by an additional linear transformation. Finally, a second feature map with the same scale as the output feature map is obtained through matrix transformation. The second feature map passes through the fifth convolutional module and the second BOT3 module again before entering the SPPF module. S2-4. At the tail of the Backbone feature extraction backbone network, the SPPF module is used to fuse the local and global features in the feature map output by the second layer BOT3 module to obtain the third feature map. The specific implementation of the upsampling feature fusion part of the feature pyramid is as follows: S2-5. The third feature map output by the SPPF module is convolved and upsampled once by the convolution module I to obtain the fourth feature map. The scale of the fourth feature map is matched with the scale of the second feature map output by the first layer BOT3 module in the feature extraction backbone network. The two feature maps with the matched scales are then concatenated using the CONCAT module to obtain the concatenated feature map I. S2-6. After the concatenated feature map I is processed by a C3STR module, it is then convolved and upsampled by the convolution module II to obtain the fifth feature map. The scale of the fifth feature map is matched with the scale of the first feature map output by the second layer C3STR module in the feature extraction backbone network. The two scale-matched feature maps are then concatenated by the CONCAT module to obtain the concatenated feature map II. The feature pyramid downsampling feature fusion part consists of three CRTN modules and two convolutional modules, as implemented below: S2-7. First, the spliced ​​feature map II is input into the first CRTN module. The CRTN module uses its branch structure and reversible block operation to fuse the global features of the input spliced ​​feature map II to obtain the fused feature map I. S2-8. Perform a convolution operation on the fused feature map I after processing by a CRTN module, and make the scale of the convolutional feature map VI consistent with the feature map VII output by the convolution module II of the upsampled feature fusion part. Then, input feature map VI and feature map VII into the CONCAT module for feature concatenation, and input the concatenated feature map into the CRTN module again for feature fusion to obtain the fused feature map II. S2-9. Perform a convolution operation on the fused feature map II after processing by the two CRTN modules, and make the scale of the convolutional feature map VIII consistent with the feature map IX output by the convolution module I of the upsampled feature fusion part. Then, input feature map VIII and feature map IX into the CONCAT module for feature concatenation, and input the concatenated feature map into the CRTN module again for feature fusion to obtain the fused feature map III.

2. The method for segmenting ultrasound images of the optic nerve sheath according to claim 1, characterized in that... Step S1 is implemented as follows: Ultrasound videos of the eyes of different patients were acquired using an ultrasound instrument. The ultrasound videos were then processed by frame extraction to form an ultrasound image dataset of the eyes. The dataset was manually labeled, with the labeling content being the lower half of the orbit and the optic nerve sheath. The labeled dataset was then divided into a training set and a test set.

3. The method for segmenting ultrasound images of the optic nerve sheath according to claim 1, characterized in that, The three segmenters in the CBC-YOLOV5s network perform semantic segmentation on the fused feature maps I, II, and III at different scales output by the three CRTN modules.

4. The method for segmenting ultrasound images of the optic nerve sheath according to claim 1, characterized in that, The BOT3 module consists of three layers: CBS, BT, and concat. The input feature map is divided into two branches. The first branch extracts features through CBS and then passes through the BT module. The MHSA module contained in the BT module is used to capture the global relationship between features, and finally outputs the feature map. The second branch only passes through the CBS module. After the CBS module extracts the feature map, it is concatenated with the feature map processed by the first branch.

5. An ultrasound image segmentation system for the optic nerve sheath, characterized in that, The system is used to implement the method as described in claim 1, and includes a data acquisition module, a CBC-YOLOV5s network module, a training module, and a testing module; Data acquisition module: Collects eye data and divides it into training and test sets; CBC-YOLOV5s network module: Based on the YOLOV5s model, it designs a new feature fusion for optic nerve sheath ultrasound images. It combines the C3STR module containing the WTA structure, the BOT3 module containing the MHSA structure, and the CRTN module containing the branch structure and reversible block operation in the YOLOv5s model to achieve the fusion of local and global features of the input image and perform data processing on the optic nerve sheath part under complex ocular ultrasound images. Training module: Construct a loss function using the training set, and train the CBC-YOLOV5s network using the loss function. After N training rounds, obtain the best-performing CBC-YOLOV5s network. Test module: Input the test set into the trained CBC-YOLOV5s network, extract optic nerve sheath image features and perform semantic segmentation.

6. A medical device, characterized in that... The device incorporates the ultrasound image segmentation method for the optic nerve sheath as described in any one of claims 1-4.

Images