Lightweight head and neck lymph node segmentation method and system

Abstract

The application discloses a lightweight head and neck lymph node image segmentation method and system, and relates to the technical field of image processing. The application comprises the following steps: receiving a lymph node image for preprocessing, and dividing the image data into a training set and a test set; constructing a segmentation network structure based on a U-Net network, wherein the segmentation network structure comprises an encoder, a volume partial convolution layer, a convolution block and a decoder; adding a lightweight boundary refinement output module at the output end of the segmentation network structure for further refining the edge; training the segmentation network structure by using the training set, taking the minimization of the overall loss function as the network optimization objective function, and optimizing the segmentation network structure. The application adds the volume partial convolution, the convolution block and the lightweight boundary refinement output module between the encoder and the decoder, which can effectively reduce the calculation amount and the parameter amount while enhancing the accuracy of the segmentation result, and improve the calculation efficiency and the model performance.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a lightweight head and neck lymph node image segmentation method and system. Background Art

[0002] Head and neck cancer is the sixth most common malignancy worldwide, and involvement of cervical lymph nodes in these patients increases the risk of distant metastasis and death. Therefore, accurate lymph node segmentation is crucial for clinical radiotherapy planning to maximize the chance of cure and minimize toxicity for head and neck cancer patients. However, manual labeling of lymph nodes is a highly labor-intensive and time-consuming task, taking experts approximately 50 minutes to process each case. This increases patient waiting time and limits the frequency of radiotherapy plan adjustments. Therefore, there is a significant demand for automatic lymph node segmentation. Computed tomography (CT) has become one of the most important imaging technologies in the diagnosis of head and neck cancer because it contains rich information and has low risk to patients. However, this task is challenging because lymph nodes can resemble surrounding tissues on CT and have variable morphology.

[0003] The development of deep learning methods has achieved remarkable success in the automatic segmentation of organs at risk, which is crucial for computer-aided diagnosis and image-guided surgery systems. In particular, the U-shaped structure network using convolutional neural networks, with its unique encoder-decoder structure and jump connection, can effectively utilize a limited number of supervised data sets to extract detailed features, thereby achieving excellent segmentation performance and has achieved success in the field of medical image segmentation. At the same time, the three-dimensional variant of the U-shaped structure network is widely used in lymph node segmentation tasks. The morphology and volume of lymph nodes are crucial for diagnosis, which requires the use of three-dimensional convolution for feature extraction. However, the three-dimensional convolution operation will incur a lot of computational cost and memory consumption. To this end, we propose a lightweight head and neck lymph node image segmentation method and system. Summary of the Invention

[0004] The purpose of the present invention is to provide a lightweight head and neck lymph node image segmentation method and system, which can significantly reduce the amount of calculation and parameter while maintaining high segmentation accuracy.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a lightweight head and neck lymph node image segmentation method, comprising the following steps:

[0006] Receive lymph node images for preprocessing and divide the image data into training and test sets;

[0007] Construct a segmentation network structure based on the U-Net network, which includes an encoder, a volumetric convolutional layer, a convolutional block, and a decoder;

[0008] Add a lightweight boundary refinement output module at the output of the segmentation network structure to further refine the edges;

[0009] The segmentation network structure is trained using the training set, and the minimization of the overall loss function is used as the network optimization objective function to optimize the segmentation network structure;

[0010] The optimized segmentation network structure is used to identify and segment the lymph nodes in the lymph image to be identified.

[0011] Furthermore, the lymph node images are received and preprocessed, and the image data is divided into a training set and a test set, as follows:

[0012] (1) Receive lymph node image data and manually annotate the lymph node regions in the patient's head and neck CT image;

[0013] (2) Randomly select some of the annotated images as training set samples and the remaining images as test set samples, perform data augmentation on the training set samples, and expand the training samples.

[0014] Furthermore, data augmentation processing is performed on the training set samples, including horizontal flipping, vertical flipping and random rotation.

[0015] Furthermore, the encoder uses the encoding part of UNet as the feature extractor, which includes three stages, wherein the first stage includes 1 convolution block, the second stage includes 2 convolution blocks, and the third stage includes 2 convolution blocks. Convolution downsampling is used between each stage to reduce the feature map.

[0016] Furthermore, each convolution block includes a 3×3×3 volume partial convolution layer, a point-by-point convolution layer, a normalization layer, and an activation function layer.

[0017] Furthermore, for a 3D input image, the number of floating-point operations of a single volumetric partial convolution layer is:

[0018]

[0019] The memory access amount of the volumetric convolution layer is:

[0020]

[0021] Where h, w, d represent the length, width and depth of the feature map respectively, k is the size of the convolution kernel, c p is the number of channels affected by the convolution.

[0022] Furthermore, the convolution block includes a volumetric partial convolution layer and two point-by-point convolution layers.

[0023] Furthermore, the decoder uses the decoding part of UNet to reconstruct the high-resolution feature map, which includes 3 stages, each stage consists of a convolution block, each convolution block contains a 3×3×3 volume partial convolution layer, a point-by-point convolution layer, a normalization layer and an activation function layer, and a transposed convolution layer is used between each stage to enlarge the feature map.

[0024] Furthermore, the overall loss function L total The details are as follows:

[0025] L total =L focal +λ·L dice 111

[0026] Where, L focal is the focal loss function, L dice is the dice loss function, and λ is the weight of the two loss functions set to 1;

[0027] The focal loss function is specifically:

[0028]

[0029] Where N represents the total number of pixels, C is the number of categories, represents the predicted value of pixel n of category c, represents the manually annotated value of pixel n of category c, and γ is the adjustment factor set to 2;

[0030] The dice loss function is specifically:

[0031]

[0032] In the formula, the value of ε is set to 1e-6.

[0033] According to one aspect of the present invention, the present invention provides a lightweight head and neck lymph node image segmentation system, comprising:

[0034] A preprocessing unit, configured to receive lymph node images for preprocessing and divide the image data into a training set and a test set;

[0035] The network structure construction unit is used to construct a segmentation network structure based on the U-Net network. The segmentation network structure includes an encoder, a volumetric convolution layer, a convolution block, and a decoder.

[0036] Boundary refinement unit, which is used to add a lightweight boundary refinement output module at the output end of the segmentation network structure to further refine the edges;

[0037] A training unit is used to train the segmentation network structure using a training set, taking minimization of the overall loss function as the network optimization objective function, and optimizing the segmentation network structure;

[0038] The recognition and segmentation unit is used to recognize and segment the lymph nodes in the lymph image to be identified using the optimized segmentation network structure.

[0039] The present invention has at least the following beneficial effects:

[0040] 1. This invention adds volumetric partial convolution and convolution blocks between the encoder and decoder. Volumetric partial convolution is used to selectively extract important features. The convolution blocks expand the receptive field through point-by-point convolution, improving the effectiveness of feature extraction. The set volumetric partial convolution and convolution blocks can effectively reduce the amount of calculation and parameters, improving computational efficiency and model performance.

[0041] 2. This paper adopts a lightweight boundary refinement output module, which further refines the edges through large-kernel deep convolution, thereby enhancing the accuracy of the segmentation results;

[0042] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic diagram of the structure of the segmented network structure of the present invention;

[0044] Figure 2 Schematic diagram of the structure of the volume partial convolution of the present invention

[0045] Figure 3 Schematic comparison of the traditional 3D U-Net module (a) and the convolutional block in the present invention (b);

[0046] Figure 4 Schematic diagram of the structure of the lightweight boundary refinement output module of the present invention;

[0047] Figure 5 Schematic diagram comparing the segmentation results of the method of the present invention and the existing method. DETAILED DESCRIPTION

[0048] The following will be combined with the accompanying drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the embodiments described are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0049] Example 1:

[0050] See also Figure 1-5 The present invention provides a technical solution: a lightweight head and neck lymph node image segmentation method, comprising the following steps:

[0051] S1. Receive lymph node images, perform preprocessing, and divide the image data into training and test sets as follows:

[0052] (1) Receive lymph node image data and manually annotate the lymph node regions in the patient's head and neck CT image;

[0053] (2) Randomly select some of the annotated images as training set samples and the remaining images as test set samples, perform data augmentation on the training set samples, and expand the training samples;

[0054] For the technical solution of this embodiment, the data comes from head and neck CT images of 123 patients, containing a total of 678 lymph nodes. Doctors manually marked the lymph nodes as target areas;

[0055] Furthermore, data augmentation processing is performed on the training set samples, including horizontal flipping, vertical flipping and random rotation, to alleviate the overfitting phenomenon of deep neural networks;

[0056] S2. Construct a segmentation network structure based on the U-Net network, which includes an encoder, a volumetric convolutional layer, a convolutional block, and a decoder;

[0057] like Figure 1 As shown, the input of the network is a three-dimensional volume image. For the technical solution of this embodiment, the encoder uses the encoding part of UNet as a feature extractor, which includes three stages, each of which consists of a different number of convolution blocks. The first stage contains one convolution block, the second stage contains two convolution blocks, and the third stage contains two convolution blocks. Each convolution block contains a 3×3×3 volume part convolution layer, a point convolution layer, a normalization layer, and an activation function layer. Convolution downsampling is used between each stage to reduce the feature map, thereby increasing the receptive field of the convolution.

[0058] In this embodiment, volumetric partial convolution (VPConv) extracts spatial features from selected input channels to capture key patterns and structures in the data. The remaining channels remain unchanged to preserve the information integrity of the original input data. This channel selection mechanism enables VPConv to effectively reduce computational burden and memory requirements while preserving key information.

[0059] For a 3D input image, the number of floating-point operations (FLOPs) of a single VPConv is only:

[0060]

[0061] Among them, h, w, d represent the length, width and depth of the feature map respectively, k is the size of the convolution kernel, c p is the number of channels affected by convolution. In practical implementations, there is usually a ratio r = c p / c=1 / 4, which makes the FLOPs of VPConv only 1 / 16 of regular convolution;

[0062] At the same time, the memory access amount of VPConv is:

[0063]

[0064] The memory access of VPConv is only 1 / 4 of that of regular convolution, because the remaining (cc p ) channels do not participate in the calculation, thus eliminating the need for memory access to these channels. The subsequent point-by-point convolution layer (PWConv) fully utilizes these channels, allowing feature information to flow in all channels;

[0065] The convolution block consists of a volumetric partial convolution layer (VPConv) and two point-wise convolution layers (PWConv), which together form a reverse residual block. The number of channels in the intermediate layer is quadrupled, and a shortcut is added to reuse the input features. This structure allows independent expansion and compression of each channel to enrich the feature representation.

[0066] It should be noted that if Figure 3 As shown, the combination of point-by-point convolution layer (PWConv) and volumetric partial convolution layer (VPConv) on the input feature map is similar to T-shaped convolution (d) in terms of its effective receptive field. Compared with the ordinary convolution of 3D U-Net (e), this combination allocates more computing resources to the central position, which has been proven to be effective in the lymph node segmentation task.

[0067] Activation layers and normalization layers are crucial in convolutional neural networks, but excessive use will limit feature diversity and thus slow down calculation speed. Therefore, the convolution block of this embodiment uses fewer activation functions and normalization layers, only uses activation functions between PWConv layers, and only uses normalization layers between VPConv and the first PWConv layer. Compared with the common U-Net, this embodiment replaces ReLU with LeakyReLU to solve the problem of neuron deactivation caused by hard truncation of negative gradients, thereby improving performance on small data sets. Instance normalization considers all elements within a single sample and a single channel when calculating normalization statistics. It is a normalization algorithm that is more suitable for scenarios with higher requirements on single pixels, such as lymph node segmentation tasks.

[0068] For the technical solution of this embodiment, the decoder uses the decoding part of UNet to reconstruct the high-resolution feature map. It consists of three stages, each of which consists of a convolution block. Each convolution block contains a 3×3×3 volume partial convolution layer, a point convolution layer, a normalization layer, and an activation function layer. A transposed convolution layer is used between each stage to increase the feature map and gradually restore the image resolution.

[0069] S3. Add a lightweight boundary refinement output module at the output of the segmentation network structure to further refine the edges;

[0070] Specifically, if Figure 4 As shown in the figure, the lightweight boundary refinement output module (LBROM) further refines the edge through a combination of large kernel depth convolution and point-by-point convolution. It first uses transposed convolution to upsample to the required number of channels, then enhances the edge features through large kernel depth convolution, and finally outputs the segmentation result through point-by-point convolution.

[0071] S4. Using the training set to train the segmentation network structure, minimizing the overall loss function as the network optimization objective function, and optimizing the segmentation network structure;

[0072] For the technical solution of this embodiment, the overall loss function is the sum of Focal Loss and Dice loss;

[0073] Focal Loss is an improved loss function based on cross entropy loss, which aims to solve the problem of class imbalance, as shown in formula (3):

[0074]

[0075] Where N represents the total number of pixels, C is the number of categories, represents the predicted value of pixel n of category c, represents the manually annotated value of pixel n of category c, and γ is the adjustment factor set to 2;

[0076] Since the sizes of lymph nodes vary greatly, using only the focal loss function will result in missing lymph nodes with smaller areas. Therefore, the dice loss function is used to make the network pay more attention to and learn the features of small targets, as shown in formula (4):

[0077]

[0078] The value of ε is set to 1e-6 to prevent the denominator from being 0;

[0079] The focal loss function and the dice loss function are added together as the final overall loss function of the network, as follows:

[0080] Ltotal =L focal +λ·L dice #(5)

[0081] Where λ is the weight of the two loss functions, which is set to 1;

[0082] It should be further explained that the model of this embodiment is based on the PyTorch deep learning framework, the GPU configuration is NVIDIA RTX4090 with 24G memory, a total of 200 cycles of training, using a multivariate learning rate strategy, the learning rate lr = 0.0004, the batch_size value during training is set to 8, and it is set to 1 during testing and verification. This embodiment adopts the AdamW optimization algorithm. Considering that small lymph nodes are more difficult to distinguish than large lymph nodes, in order to test the performance of the model on smaller lymph nodes, we divided all test lymph nodes into three groups, including small lymph nodes (volume ≤ 404mm 3 ), large lymph nodes (volume > 404mm 3 ) and all lymph nodes, of which 404 mm 3 represents the median volume of all lymph nodes;

[0083] S5. Use the optimized segmentation network structure to identify and segment the lymph nodes in the lymph image to be identified.

[0084] In order to quantitatively evaluate the performance of the proposed method, the Dice coefficient (DSC) and intersection over union (IoU), two commonly used evaluation indicators in medical image segmentation, are used as the measurement criteria for the experimental results.

[0085] DSC is a function that evaluates the similarity between the prediction results and the manual annotations, which is defined as:

[0086]

[0087] Where TP (True Positive) is the number of true positive pixels, FP (False Positive) is the number of false positive pixels, and FN (False Negative) is the number of false negative pixels;

[0088] The intersection-over-union ratio is the overlap of two regions divided by the combined region of the two regions, and is defined as:

[0089]

[0090] Where A is the predicted area and B is the true area; |A∩B| represents the intersection area of ​​the predicted segmentation area and the true segmentation area (that is, the number of overlapping pixels), and |A∪B| represents the union area of ​​the predicted segmentation area and the true segmentation area (that is, the total number of pixels, including all pixels in the predicted and true areas).

[0091] To demonstrate the effectiveness of the module, the 3D Partial U-Net proposed in this example is selected as the basic network. The corresponding ablation experiment results are shown in Table 1. The results show that the VPConv and LBROM modules proposed in this paper are effective.

[0092] As shown in Table 1, we first start with the original model and then remove the LBROM, which results in a slight decrease in segmentation accuracy. It is worth noting that using the LBROM is more effective in segmenting small lymph nodes. Then, the convolution block is replaced with a 3D U-Net block, which results in a significant increase in the number of model parameters and FLOPs. However, the accuracy only decreases slightly, which shows the effectiveness of the convolution block in this embodiment.

[0093] The number of channels is a major hyperparameter that affects the number of parameters, complexity, and performance of the model. In Table 1, the experiments in this embodiment show two different configurations of the model; the ordinary model follows C1=32, C2=64, C3=128, C5=256, the model 3D Partial U-Net-L follows C1=64, C2=128, C3=256, and C4=512, while the 3D PartialU-Net-M follows C1=16, C2=32, C3=128, and C4=160. It can be observed that increasing the number of channels further improves the performance but also increases the computational overhead. Although reducing the number of channels will reduce the performance, a very lightweight model is obtained, and the performance degradation is not significant.

[0094] Table 1 Comparison of segmentation results of ablation experiments

[0095]

[0096]

[0097] Furthermore, this embodiment compares the proposed method with 3D U-Net, SegResNet, AttentionU-Net, Dense V-Net, UNETR, MedNeXt, 3D UX-Net and PMSFNet. The segmentation results of different methods are shown in the figure. Figure 5 As shown;

[0098] The method of this embodiment achieves more accurate results than other methods while also reducing the number of parameters and computational complexity. Table 2 shows the quantitative segmentation results of different methods. The results show that the method of this embodiment has the highest Dice coefficient and IoU coefficient for all lymph node segmentations.

[0099] Table 2 Performance comparison with other 3D medical image segmentation models

[0100]

[0101]

[0102] In summary, the technical solution proposed in this embodiment can improve the accuracy and efficiency of segmentation while reducing the amount of calculation and the number of parameters.

[0103] Example 2:

[0104] The present invention provides a lightweight head and neck lymph node image segmentation system, comprising:

[0105] A preprocessing unit, configured to receive lymph node images for preprocessing and divide the image data into a training set and a test set;

[0106] The network structure construction unit is used to construct a segmentation network structure based on the U-Net network. The segmentation network structure includes an encoder, a volumetric convolution layer, a convolution block, and a decoder.

[0107] Boundary refinement unit, which is used to add a lightweight boundary refinement output module at the output end of the segmentation network structure to further refine the edges;

[0108] A training unit is used to train the segmentation network structure using a training set, taking minimization of the overall loss function as the network optimization objective function, and optimizing the segmentation network structure;

[0109] The recognition and segmentation unit is used to recognize and segment the lymph nodes in the lymph image to be identified using the optimized segmentation network structure.

[0110] Specifically, the above-mentioned preprocessing unit, network structure construction unit, boundary refinement unit, training unit and recognition and segmentation unit can be embedded in a computer processing system. The computer calls the above-mentioned units to complete the task of equivalent modeling based on the lightweight head and neck lymph node image segmentation method provided above; the above-mentioned preprocessing unit, network structure construction unit, boundary refinement unit, training unit and recognition and segmentation unit can perform operations according to the specific steps given in the lightweight head and neck lymph node image segmentation method.

[0111] It should be understood that the division of the units of the above system is only a logical function division, and all or part of the units can be integrated into one physical entity or physically separated in actual implementation, and the units can all be implemented in the form of software called by a processing element; or all can be implemented in the form of hardware; or part of the units can be implemented in the form of software called by a processing element, and part of the units can be implemented in the form of hardware. For example, the receiving unit can be a separately established processing element, or can be integrated in a chip of the above device, in addition, the receiving unit can also be stored in the form of program code in the memory of the above device, and the function of the above signal processing unit is called and executed by a processing element of the above device, and the implementation of other units is similar. In addition, all or part of the units can be integrated together or independently implemented, and the processing element described herein can be an integrated circuit having a signal processing capability, and in the implementation process, each step of the above method or each unit can be completed by an integrated logic circuit of hardware in the processing element or an instruction in the form of software.

[0112] For example, the above units can be one or more integrated circuits configured to implement the above method, such as one or more application specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs), etc. For another example, when a certain unit above is implemented in the form of a processing element scheduling program code, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, the units can be integrated together to implement in the form of a system on a chip (SOC).

[0113] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0114] For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on", "installed on", "fixed on" or "set on" another element, it can be directly on the other element or there can be a central element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there can be a central element at the same time. The terms "vertical", "horizontal", "up", "down", "left", "right" and similar expressions used herein are for illustrative purposes only and are not intended to be the only embodiment.

[0115] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

[0116] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Claims

1. A lightweight head and neck lymph node image segmentation method, characterized by: The following steps are involved: Receive lymph node images for preprocessing and divide the image data into training and test sets; Construct a segmentation network structure based on the U-Net network, which includes an encoder and a decoder; Add a lightweight boundary refinement output module at the output of the segmentation network structure to further refine the edges; The segmentation network structure is trained using the training set, and the minimization of the overall loss function is used as the network optimization objective function to optimize the segmentation network structure; The optimized segmentation network structure is used to identify and segment the lymph nodes in the lymph image to be identified; The encoder uses the encoding part of UNet as the feature extractor, which consists of three stages. The first stage contains one convolution block, the second stage contains two convolution blocks, and the third stage contains two convolution blocks. Convolution downsampling is used between each stage to reduce the feature map. Each convolution block contains a 3×3×3 volumetric partial convolution layer, a point-by-point convolution layer, a normalization layer, and an activation function layer; Volumetric partial convolution extracts spatial features from selected input channels to capture key patterns and structures in the data, while remaining unchanged for the remaining channels to preserve the information integrity of the original input data; The lightweight boundary refinement output module further refines the edges through a combination of large kernel depthwise convolution and point-by-point convolution. It first uses transposed convolution to upsample to the required number of channels, then enhances edge features through large kernel depthwise convolution, and finally outputs the segmentation result through point-by-point convolution.

2. The lightweight head and neck lymph node image segmentation method according to claim 1, characterized in that: Receive lymph node images for preprocessing and divide the image data into training and test sets as follows: (1) Receive lymph node image data and manually annotate the lymph node areas in the patient's head and neck CT images; (2) Randomly select some of the annotated images as training set samples and the remaining images as test set samples. Perform data augmentation on the training set samples to expand the training samples.

3. The lightweight head and neck lymph node image segmentation method according to claim 2, characterized in that: Data augmentation processing of training set samples includes horizontal flipping, vertical flipping and random rotation.

4. The lightweight head and neck lymph node image segmentation method according to claim 3, characterized in that: For a 3D input image, the number of floating-point operations for a single volumetric partial convolution layer is: The memory access amount of the volumetric convolution layer is: Where, Represent the length, width and depth of the feature map respectively, is the convolution kernel size, is the number of channels affected by the convolution.

5. The lightweight head and neck lymph node image segmentation method according to claim 4, characterized in that: The convolution block includes a volumetric partial convolution layer and two point-by-point convolution layers.

6. The lightweight head and neck lymph node image segmentation method according to claim 5, characterized in that: The decoder uses the decoding part of UNet to reconstruct high-resolution feature maps. It consists of three stages, each of which consists of a convolution block. Each convolution block contains a 3×3×3 volume partial convolution layer, a point-by-point convolution layer, a normalization layer, and an activation function layer. A transposed convolution layer is used between each stage to enlarge the feature map.

7. The lightweight head and neck lymph node image segmentation method according to claim 6, characterized in that: The overall loss function The details are as follows: 111 Where, is the focal loss function, is the dice loss function, The weights of the two loss functions are set to 1; The focal loss function is specifically: Where N represents the total number of pixels, C is the number of categories, represents the predicted value of pixel n of category c, represents the manually annotated value of pixel n of category c, and γ is the adjustment factor set to 2; The dice loss function is specifically: In the formula, the value of ε is set to 1e-6.

8. A lightweight head and neck lymph node image segmentation system, used to implement the lightweight head and neck lymph node image segmentation method according to any one of claims 1 to 7, characterized in that: include: A preprocessing unit, configured to receive lymph node images for preprocessing and divide the image data into a training set and a test set; The network structure construction unit is used to construct a segmentation network structure based on the U-Net network. The segmentation network structure includes an encoder, a volumetric convolution layer, a convolution block, and a decoder. Boundary refinement unit, which is used to add a lightweight boundary refinement output module at the output end of the segmentation network structure to further refine the edges; A training unit is used to train the segmentation network structure using a training set, taking minimization of the overall loss function as the network optimization objective function, and optimizing the segmentation network structure; The recognition and segmentation unit is used to recognize and segment the lymph nodes in the lymph image to be identified using the optimized segmentation network structure.

Images