Improved recurrent laryngeal nerve image auxiliary segmentation method based on UNet segmentation model

Abstract

The invention discloses an improved recurrent laryngeal nerve image auxiliary segmentation method based on a UNet segmentation model, and the method comprises the following steps: S1, carrying out the data preprocessing, and obtaining a training set and a test set; s2, improving the network structure of the UNet segmentation model on the basis of the UNet segmentation model; s3, setting basic parameters of the improved UNet segmentation model; and S4, network structure model training is carried out by using the training set in the step S1, and model segmentation effect evaluation is carried out by using the test set image after each period of training is completed, so that the balance between the calculation efficiency and the performance is realized, and the calculation cost is remarkably reduced while the high precision is maintained; the method focuses on task key information, reduces redundant features, and reduces image noise interference. Efficient recognition of recurrent laryngeal nerves is achieved by improving the UNet segmentation model, and therefore operation errors in an operation are reduced, and accidental injuries in the operation are avoided.

Description

Technical Field

[0001] This invention relates to the fields of computer technology and medical image analysis technology, specifically to an image-assisted segmentation method for the recurrent laryngeal nerve based on an improved UNet segmentation model. Background Technology

[0002] With the continued rise in the incidence of thyroid nodules, this disease has become one of the most common endocrine system disorders worldwide. The recurrent laryngeal nerve is a crucial nerve that controls the basic functions of the larynx, including phonation, respiration, and swallowing, and is closely related to a patient's language and quality of life. In laparoscopic thyroid surgery, limited vision and the high precision required further increase the risk of recurrent laryngeal nerve injury.

[0003] Against this backdrop, research on AI-assisted automatic identification of the recurrent laryngeal nerve has become a research hotspot in the interdisciplinary development of medical image analysis and surgery. Deep learning models analyze and mine image data from thyroid surgery videos, and models based on convolutional neural networks (CNNs), such as the U-Net model, assist surgeons in accurately segmenting the location and contour of the recurrent laryngeal nerve, improving decision support for doctors and greatly reducing the risk of accidental nerve injury.

[0004] Current mainstream segmentation models, such as those used in research on recurrent laryngeal nerve (RLN) segmentation, selected five mainstream segmentation models (U-Net, FCN, SegNet, Deeplabv3+, and U-Net++). These mainstream models lack sufficient segmentation accuracy for small targets like the RLN; their global field of view is limited, making it difficult to capture cross-regional correlation information; and they struggle to clearly depict the contours and structural details of neural tissue. Furthermore, during model training, the networks often fail to converge, leading to training bottlenecks. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes an image-assisted segmentation method for the recurrent laryngeal nerve based on an improved UNet segmentation model. This method achieves efficient identification of the recurrent laryngeal nerve, thereby reducing intraoperative errors and avoiding accidental damage during surgery.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An image-assisted segmentation method for recurrent laryngeal nerve based on an improved UNet segmentation model includes the following steps: S1. Data preprocessing to obtain training and test sets; S2. Improve the network structure based on the UNet segmentation model; S3. Set the basic parameters of the improved UNet segmentation model; S4. Use the training set from step S1 to train the network structure model. After each training cycle is completed, use the test set images to evaluate the model segmentation effect.

[0007] As a preferred technical solution of the present invention: Step S1 is as follows: S11. Collect clinical surgical videos of laparoscopic thyroid surgery and acquire recurrent laryngeal nerve imaging data under the guidance of doctors in this department. S12. Preprocess the collected recurrent laryngeal nerve imaging data to obtain a dataset; S13. Perform data augmentation on the dataset from step S12; S14. Result classification, as follows: Use the latest SAM to depict neural labels on the dataset in step S13 and label them, where 80% is the training set and 20% is the test set.

[0008] As a preferred technical solution of the present invention: in step S12... Preprocessing of recurrent laryngeal nerve imaging data includes size uniformization, contrast enhancement, and noise reduction.

[0009] As a preferred technical solution of the present invention: In step S13, Data augmentation of the dataset includes random rotation, random translation, brightness adjustment, and the addition of Gaussian noise.

[0010] As a preferred technical solution of the present invention: in step S2... A combined architecture of "CBAM attention module + DO-Conv deep convolution module + CPAM fusion module" is constructed. The synergistic effect of each module in noise suppression and focusing, fine structure capture, and multi-scale feature integration enables accurate recognition and segmentation of RLN, as detailed below: CBAM Attention Module: A dual attention mechanism consisting of channel attention and spatial attention, enabling precise focusing on key information through two-dimensional feature extraction. Global average pooling and max pooling are used to generate channel-level feature details. The two pooling operations complement each other. Taking the intermediate feature map F as input, the one-dimensional channel attention map Mc and the two-dimensional spatial attention map Ms are derived step by step through the CBAM attention module. F′=Mc(F)⊗F; F′′=Ms(F′)⊗F′; DO-Conv module: Introduces a depthwise convolution expansion factor D into CNN convolutional layers. mul This forms a fused convolutional model, with convolution weights W'', which are formed by the CNN convolutional kernel matrix W' and the depthwise transpose matrix D.T Fusion generation; CPAM fusion module: consisting of channel attention branch M c Spatial attention branch M s Composition, channel attention branch M c Channel-level feature extraction is achieved by generating channel weights using global pooling, a lightweight MLP, and activation functions; spatial attention branch M s A two-dimensional spatial attention map is constructed using multi-scale spatial pooling, channel compression, and nonlinear activation to further extract key region features. The symbols are defined as follows: F1 is the first input feature map, F2 is the second input feature map; F′ is the feature map after channel attention processing, and F′′ is the feature map after spatial attention weighting. The final output feature map is denoted as F. out The corresponding calculation process is as follows: F′=M c (F1)⊗F1; F′′=M s (F′)⊗F′; F out =F′′+F2.

[0011] As a preferred technical solution of the present invention: Step S3 is as follows: The basic parameters to be set include the initial learning rate of the network, the learning rate decay method, the number of network iterations, and the loss function; The initial learning rate of the network was set to 0.0001, the learning rate decay method was cosine annealing, the number of network iterations was set to 200, and the loss function was cross-entropy loss function.

[0012] As a preferred technical solution of the present invention: Step S4 is as follows: The network model is trained using the training set. After each training cycle, the model segmentation effect is evaluated using test set images. The model is evaluated on the validation set every 5 rounds, based on the Iou and Dice metrics. The optimal weights are saved every 10 rounds based on the metrics of the validation set.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention structurally optimizes the UNet segmentation model to address some issues of current mainstream models, such as the limitation of single convolutional layers in fully extracting subtle features in complex scenes. The model achieves a balance between computational efficiency and performance, significantly reducing computational costs while maintaining high accuracy. It focuses on key task information, reducing redundant features and image noise interference. This invention effectively achieves efficient identification of the recurrent laryngeal nerve, thereby reducing intraoperative errors and avoiding accidental damage during surgery. Attached Figure Description

[0014] Figure 1 This is a model structure diagram of the present invention; Figure 2 It is a SAM annotation diagram; Figure 3 It is a data augmentation graph of the dataset; Figure 4 These are experimental results images; Figure 5 This is a table comparing the accuracy of this application with existing models. Detailed Implementation

[0015] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0016] like Figure 1 As shown, the present invention proposes an image-assisted segmentation method for recurrent laryngeal nerve based on an improved UNet segmentation model, comprising the following steps: S1. Data preprocessing to obtain the training and test sets, as follows: S11. Collect clinical surgical videos of laparoscopic thyroid surgery and acquire recurrent laryngeal nerve imaging data under the guidance of doctors in this department. S12. Preprocess the collected recurrent laryngeal nerve imaging data to obtain a dataset; S13. Perform data augmentation on the dataset from step S12; S14. Result classification, as detailed below: Using the latest SAM on the dataset in step S13, neural labels are depicted and annotated, accurately depicting the neural labels and improving the efficiency of annotation. Figure 2 As shown, (A) is SAM-assisted annotation and (B) is manual annotation. SAM-assisted annotation provides accurate and rich edge information, providing high-quality training data for subsequent model training. Among them, 80% is the training set and 20% is the test set.

[0017] In step S12, the preprocessing of the recurrent laryngeal nerve image data includes unifying the size, enhancing contrast, and reducing noise to improve the recurrent laryngeal nerve recognition.

[0018] In step S13, data augmentation of the dataset includes random rotation, random translation, brightness adjustment, and the addition of Gaussian noise to improve the recognition accuracy of the recurrent laryngeal nerve and create rich and diverse learning samples. Simulating shooting angle deviations enhances the model's ability to recognize targets from different perspectives; simulating target position shifts in the image avoids the model's over-reliance on the target's fixed position; simulating low-light scenes enhances the model's robustness to different lighting conditions; and simulating camera sensor noise or low-quality shooting scenes improves the model's robustness. Figure 3 As shown.

[0019] S2. Improve the network structure based on the UNet segmentation model: This study addresses the core challenges of blurred boundaries and background interference in the recognition and segmentation of recurrent laryngeal nerve (RLN). To address these challenges, a combined architecture of "CBAM attention module + DO-Conv deep convolution module + CPAM fusion module" is constructed. The synergistic effect of each module in noise suppression and focusing, fine structure capture, and multi-scale feature integration enables accurate recognition and segmentation of RLN.

[0020] in, The CBAM attention module extracts features in both spatial and channel dimensions, effectively suppressing background noise interference, and enabling the model to capture information more accurately in the target region.

[0021] The DO-Conv deep convolution module significantly improves the model's efficiency in capturing subtle structural information of RLN and directly optimizes the overall recognition performance. However, its computational complexity is high and it has certain requirements for hardware computing power.

[0022] The CPAM fusion module is responsible for capturing and integrating multi-scale information. The generated features contain rich semantic association information while maintaining high spatial accuracy.

[0023] Addressing the core challenges of blurred target boundaries and severe background interference in recurrent laryngeal nerve (RLN) recognition and segmentation tasks, this study proposes a combined architecture of a CBAM attention module, a DO-Conv deep convolution module, and a CPAM fusion module. Through the synergistic effect of these modules, the system accurately outputs RLN recognition and segmentation results.

[0024] Specifically as follows: A combined architecture of "CBAM attention module + DO-Conv deep convolution module + CPAM fusion module" is constructed. The synergistic effect of each module in noise suppression and focusing, fine structure capture and multi-scale feature integration enables accurate recognition and segmentation of RLN.

[0025] Specifically as follows: CBAM Attention Module: A dual attention mechanism consisting of channel attention and spatial attention. Dual-dimensional feature extraction enables precise focusing on key information. It generates channel-level feature details through global average pooling and max pooling, with the two pooling operations complementing each other.

[0026] Using the intermediate feature map F as input, the CBAM attention module progressively derives the one-dimensional channel attention map Mc and the two-dimensional spatial attention map Ms. This derivation process requires no manual setting of feature weights and relies entirely on the model's adaptive learning. ; ; Therefore, the CBAM attention module can effectively focus on key information while suppressing irrelevant noise. This design adds almost no computational burden, improving model performance without sacrificing efficiency.

[0027] DO-Conv module: Traditional Convolutional Neural Networks (CNNs) use small kernel convolutions to capture local spatial information in images, and their small receptive field allows for more accurate capture of local details. However, the small receptive field of small kernel convolutions also has some limitations: the model's ability to interact with information from distant features is insufficient, and it cannot engage in long-distance semantic communication. The DO-Conv deep convolution module designed in this invention integrates the advantages of traditional convolutions and deep convolutions, enhancing the model's feature representation capabilities and improving the efficiency of interaction with large-scale features, thus effectively compensating for the aforementioned shortcomings of small kernel convolutions.

[0028] Specifically: DO-Conv module: introduces a depthwise convolution expansion factor D into the CNN convolutional layers. mul To form a fused convolutional model, thereby enhancing the model's expressive power, the depthwise convolution expansion factor D... mul Based on the flexible configuration of channel expansion multiples, the model's ability to extract features in multiple dimensions is enhanced. The convolution weights of the fused convolutional model are W'', which are composed of the CNN convolutional kernel matrix W' and the depthwise transpose matrix D. T By fusing and generating features, the DO-Conv module combines the advantages of both traditional convolution and depthwise convolution, enhancing the model's feature representation capabilities and improving feature interaction efficiency. CPAM Fusion Module: CPAM (Channel and Spatial Attention Mechanism) is a highly efficient feature fusion module. Channel attention (Mc) and spatial attention (Ms) work synergistically, resulting in multi-scale feature extraction that balances global relevance with local detail. It dynamically calculates feature weights, enhancing feature representation in key regions while suppressing redundant or irrelevant information, thus highlighting important features at the neural edge. The symbols are defined as follows: F1 is the first input feature map, F2 is the second input feature map; F′ is the feature map after channel attention processing, F′′ is the feature map after spatial attention weighting, and the final output feature map is denoted as F. out The corresponding calculation process is as follows: ; ; .

[0029] Therefore, the CPAM fusion module can be applied to various complex feature fusion scenarios, enhancing the model's understanding of global semantic information and its ability to capture local details. It is very effective in tasks such as small object detection, semantic segmentation, and multi-scale feature processing.

[0030] S3. Set the basic parameters of the improved UNet segmentation model as follows: The basic parameters to be set include the initial learning rate of the network, the learning rate decay method, the number of network iterations, and the loss function; The initial learning rate of the network was set to 0.0001, the learning rate decay method was cosine annealing, the number of network iterations was set to 200, and the loss function was cross-entropy loss function.

[0031] S4. Train the network structure model using the training set from step S1. After each training cycle is completed, evaluate the model segmentation performance using the test set images, as follows: The network model is trained using the training set. After each training cycle, the model segmentation effect is evaluated using test set images. The model is evaluated on the validation set every 5 rounds, based on the Iou and Dice metrics. The optimal weights are saved every 10 rounds based on the metrics of the validation set.

[0032] In the experiment, the segmentation results generated by the model (represented by green outlines) were visually compared with the neural outlines (blue) annotated by specialists. In images with good visual conditions and clear neural boundaries, the model was able to accurately identify and segment neural structures, demonstrating high consistency and good robustness and stability.

[0033] In scenes with poor image quality, interference factors such as uneven lighting, image blurring, tissue occlusion, or complex edge structures significantly affect the segmentation performance of the model, resulting in a certain degree of decline in its recognition ability.

[0034] The above phenomena indicate that image quality plays a crucial role in neural segmentation tasks, especially in surgical applications. The clarity and stability of image acquisition directly affect the practicality and accuracy of the model, as shown in the following specific effects. Figure 4 As shown. The accuracy of this application is compared with that of five mainstream models, for example. Figure 5 As shown.

[0035] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for image-assisted segmentation of the recurrent laryngeal nerve based on an improved UNet segmentation model, characterized in that, Includes the following steps: S1. Data preprocessing to obtain training and test sets; S2. Improve the network structure based on the UNet segmentation model; S3. Set the basic parameters of the improved UNet segmentation model; S4. Use the training set from step S1 to train the network structure model. After each training cycle is completed, use the test set images to evaluate the model segmentation effect.

2. The image-assisted segmentation method for recurrent laryngeal nerve based on an improved UNet segmentation model according to claim 1, characterized in that, Step S1 is as follows: S11. Collect clinical surgical videos of laparoscopic thyroid surgery and acquire recurrent laryngeal nerve imaging data under the guidance of doctors in this department. S12. Preprocess the collected recurrent laryngeal nerve imaging data to obtain a dataset; S13. Perform data augmentation on the dataset from step S12; S14. Result classification, as follows: Use the latest SAM to depict neural labels on the dataset in step S13 and label them, where 80% is the training set and 20% is the test set.

3. The image-assisted segmentation method for the recurrent laryngeal nerve based on the improved UNet segmentation model according to claim 2, characterized in that, In step S12, Preprocessing of recurrent laryngeal nerve imaging data includes size uniformization, contrast enhancement, and noise reduction.

4. The image-assisted segmentation method for the recurrent laryngeal nerve based on the improved UNet segmentation model according to claim 2, characterized in that, In step S13, Data augmentation of the dataset includes random rotation, random translation, brightness adjustment, and the addition of Gaussian noise.

5. The image-assisted segmentation method for the recurrent laryngeal nerve based on the improved UNet segmentation model according to claim 1, characterized in that, In step S2, A combined architecture of "CBAM attention module + DO-Conv deep convolution module + CPAM fusion module" is constructed. The synergistic effect of each module in noise suppression and focusing, fine structure capture, and multi-scale feature integration enables accurate recognition and segmentation of RLN, as detailed below: CBAM Attention Module: A dual attention mechanism consisting of channel attention and spatial attention, enabling precise focusing on key information through two-dimensional feature extraction. Global average pooling and max pooling are used to generate channel-level feature details. The two pooling operations complement each other. Taking the intermediate feature map F as input, the one-dimensional channel attention map Mc and the two-dimensional spatial attention map Ms are derived step by step through the CBAM attention module. F′=Mc(F)⊗F; F′′=Ms(F′)⊗F′; DO-Conv module: Introduces a depthwise convolution expansion factor D into CNN convolutional layers. mul This forms a fused convolutional model, with convolution weights W'', which are formed by the CNN convolutional kernel matrix W' and the depthwise transpose matrix D. T Fusion generation; CPAM fusion module: consisting of channel attention branch M c Spatial attention branch M s Composition, channel attention branch M c Channel-level feature extraction is achieved by generating channel weights based on global pooling, lightweight MLP, and activation functions; space Attention branch M s A two-dimensional spatial attention map is constructed using multi-scale spatial pooling, channel compression, and nonlinear activation to further extract key region features. The symbols are defined as follows: F1 is the first input feature map, F2 is the second input feature map; F′ is the feature map after channel attention processing, F′′ is the feature map after spatial attention weighting, and the final output feature map is denoted as F. out The corresponding calculation process is as follows: F′=M c (F1)⊗F1; F′′=M s (F′)⊗F′; F out =F′′+F2。 6. The image-assisted segmentation method for recurrent laryngeal nerve based on the improved UNet segmentation model according to claim 1, characterized in that, Step S3 is as follows: The basic parameters to be set include the initial learning rate of the network, the learning rate decay method, the number of network iterations, and the loss function; The initial learning rate of the network was set to 0.0001, the learning rate decay method was cosine annealing, the number of network iterations was set to 200, and the loss function was cross-entropy loss function.

7. The image-assisted segmentation method for recurrent laryngeal nerve based on an improved UNet segmentation model according to claim 1, characterized in that, Step S4 is as follows: The network model is trained using the training set. After each training cycle, the model segmentation effect is evaluated using test set images. The model is evaluated on the validation set every 5 rounds, based on the Iou and Dice metrics. The optimal weights are saved every 10 rounds based on the metrics of the validation set.

Images