A method, system, device and storage medium for bronchoscopic image segmentation

Abstract

This invention discloses a method, system, device, and storage medium for bronchoscopic image segmentation, belonging to the field of image processing technology. The method includes acquiring a bronchoscopic image; inputting the bronchoscopic image into a pre-trained image segmentation model to obtain a labeled image of the bronchial aperture region; wherein the image segmentation model includes an encoder module, a decoder module, and a skip connection layer; the encoder module includes a multi-layer encoding module, each layer including an ODConv module and an EMA module; the skip connection layer is used to input the second output features of each layer into the decoder module; the decoder module includes a multi-layer decoding module, each layer including an ODConv module and an EMA module. This invention solves the problem of bronchoscope collision or puncture during bronchoscopic robotic surgery due to unclear bronchial aperture regions.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to a method, system, device and storage medium for bronchoscopic image segmentation. Background Technology

[0002] In the current field of bronchoscopic lung surgery robots in China, surgeons rely heavily on preoperative training and past experience. During the procedure, surgeons can only rely on the real-time images captured by the bronchoscope controller and the bronchoscope's perspective. However, these real-time images lack the functionality of bronchial channel marking and central pathway visualization. Therefore, in current lung surgery robot operations, surgeons are prone to incidents such as friction between the bronchoscope and bronchial mucosa, or collisions or even punctures between the bronchoscope tip and the bronchial wall during forward and backward movements of the bronchoscope.

[0003] In the current field of medical image segmentation, several algorithmic structures have been proposed. Traditional medical image segmentation methods are typically based on mathematical representations, requiring no large amounts of labeled data for training. They are easily integrated with the anatomical knowledge and experience of clinicians, such as using lesion information provided by doctors to reasonably determine segmentation thresholds and initial contours. However, these methods are highly dependent on human intervention. For example, thresholding methods are efficient, but threshold settings are often too empirical. While segmentation methods such as region growing and clustering incorporate pixel neighborhood information and have relatively low computational cost, they face significant challenges in selecting initial seed point locations, distinguishing between the segmentation target and background, and integrating the doctor's prior knowledge. Furthermore, they are prone to undersegmentation and oversegmentation. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, system, device and storage medium for bronchoscopic image segmentation, which solves the problem of bronchoscope hitting the wall or being punctured during bronchoscopic robotic surgery due to unclear bronchial passage areas.

[0005] This invention provides the following technical solution:

[0006] In a first aspect, a bronchoscopic image segmentation method is provided, comprising: acquiring a bronchoscopic image; inputting the bronchoscopic image into a pre-trained image segmentation model to obtain a labeled image of the bronchoscopic aperture region; wherein the image segmentation model includes an encoder module, a decoder module, and a skip connection layer; the encoder module includes a multi-layer encoding module, each layer of which includes an ODConv module and an EMA module; the bronchoscopic image is input into the multi-layer encoding module, and convolved by the first layer of the ODConv module to obtain a first output feature; the first output feature of each layer is input into each layer of the EMA module and the next layer of the encoding module respectively; each layer of the EMA module extracts features from the first output feature to obtain a second output feature; and the next layer of the encoding module extracts features from the first output feature of the previous layer. After downsampling, the output feature is input into the ODConv module of the next layer; the skip connection layer is used to input the second output feature of each layer into the decoder module respectively; the decoder module includes a multi-layer decoding module, and each layer decoding module includes an ODConv module and an EMA module; the first layer decoding module upsamples the second output feature of the last layer encoding module and splices the second output features of the last two layers encoding modules to obtain the spliced ​​feature. The spliced ​​feature is then processed by the ODConv module and the EMA module to obtain the third output feature. The remaining layer decoding modules upsample the third output feature of the previous layer and splice the third output feature of the previous layer with the second output feature of the corresponding encoding module until the last layer decoding module outputs the marked image of the bronchoscope channel region.

[0007] As a preferred embodiment of the present invention, the bronchoscopic image is input into a multi-layer coding module, and convolved by a first-layer ODConv module to obtain a first output feature, including:

[0008] The first-layer ODConv module performs a convolution operation on the bronchoscopic images, represented as:

[0009] ;

[0010] in, This represents the first output feature of the first-layer ODConv module. This represents the nth convolutional kernel of the first-layer ODConv module. Represents the convolution kernel Attention scalar Indicates the convolution kernel Attention scalar in the spatial dimension Indicates the convolution kernel Attention scalar at the input channel dimension. Indicates the convolution kernel The attention scalar of the output channel dimension, where X represents the bronchoscopic image.

[0011] As a preferred embodiment of the present invention, the first output feature of each layer is input to each layer's EMA module and the next layer's encoding module, respectively. The next layer's encoding module downsamples the first output feature of the previous layer and inputs it to the next layer's ODConv module, including:

[0012] The first output feature of the ODConv module in the i-th layer coding module is represented as follows: The first output feature of the i-th layer coding module Input the EMA module in the i-th layer encoding module and the next layer encoding module respectively;

[0013] Except for the first layer encoding module, all other layers of encoding modules include a max pooling layer, which is used to reduce the resolution of the first output feature of the previous layer to half of its original value, and represent it as the first input feature. The (i+1)th layer ODConv module processes the first input features Performing a convolution operation is represented as:

[0014] ;

[0015] in, This represents the first output feature of the (i+1)th layer ODConv module. Represents the first input feature of the i-th layer ODConv module .

[0016] As a preferred embodiment of the present invention, each layer of the EMA module performs feature extraction on the first output feature to obtain the second output feature, including:

[0017] The EMA module includes three branches, two of which are... Branch and one Branches; The branch performs a one-dimensional global average pooling operation on the first output feature. The branch performs the first output feature Convolution concatenates the outputs of the three branches to obtain the second output feature;

[0018] The second output feature of the EMA module in the i-th layer coding module is represented as follows: .

[0019] As a preferred embodiment of the present invention, the first layer decoding module upsamples the second output feature of the last layer encoding module and concatenates the second output features of the last two layers encoding modules to obtain the concatenated feature, including:

[0020] The first-layer decoding module upsamples the second output feature of the last-layer encoding module to double the resolution of the second output feature of the last-layer encoding module, thus obtaining the second input feature.

[0021] Except for the last coding module, each coding module corresponds to a skip connection layer, which is used to input the second output feature of the coding module to the corresponding decoding module;

[0022] The second output feature of the last layer encoding module is input into the skip connection layer above, concatenated with the second output feature of the previous layer, and then input into the first layer decoding module to concatenate the second input feature, thus obtaining the concatenated feature.

[0023] Secondly, a bronchoscopic image segmentation system is provided, comprising: an image acquisition module for acquiring bronchoscopic images;

[0024] The segmentation module is used to input the bronchoscopic image into a pre-trained image segmentation model to obtain a labeled image of the bronchoscopic aperture region;

[0025] The image segmentation model includes an encoder module, a decoder module, and a skip connection layer.

[0026] The encoder module includes a multi-layer encoding module, and each layer encoding module includes an ODConv module and an EMA module. The bronchoscopic image is input into the multi-layer encoding module, and is convolved by the first layer ODConv module to obtain the first output feature. The first output feature of each layer is input into the EMA module of each layer and the next layer encoding module respectively. Each layer EMA module extracts features from the first output feature to obtain the second output feature. The next layer encoding module downsamples the first output feature of the previous layer and inputs it into the ODConv module of the next layer.

[0027] The skip connection layer is used to input the second output feature of each layer into the decoder module respectively;

[0028] The decoder module includes a multi-layer decoding module, each of which includes an ODConv module and an EMA module. The first-layer decoding module upsamples the second output feature of the last-layer encoding module and splices the second output features of the last two encoding modules to obtain a spliced ​​feature. The spliced ​​feature is then processed by the ODConv module and the EMA module to obtain a third output feature. The remaining layer decoding modules upsample the third output feature of the previous layer and splice the third output feature of the previous layer with the second output feature of the corresponding encoding module until the last layer decoding module outputs a marked image of the bronchoscope channel region.

[0029] Thirdly, a bronchoscopic image segmentation apparatus is provided, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method described in the first aspect.

[0030] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the steps of the method described in the first aspect.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] This invention provides a bronchoscopic image segmentation method. By setting an ODConv module, the shape and size of the convolution kernel are dynamically adjusted according to the input bronchoscopic image to extract features. By setting an EMA module, cross-channel information can be encoded to adjust the importance of different channels and to retain precise spatial structure information in the channels. By obtaining a labeled image of the bronchoscopic orifice region through the image segmentation model, the problem of bronchoscope collision or puncture during bronchoscopic robotic surgery due to unclear bronchial orifice regions is solved. Attached Figure Description

[0033] Figure 1 This is a flowchart of the bronchoscopic image segmentation method in an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the image segmentation model method in an embodiment of the present invention. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0036] Example 1

[0037] This embodiment provides a method for bronchoscopic image segmentation. For example... Figure 1 As shown, the specific steps include the following:

[0038] Step 1: Obtain bronchoscopic images.

[0039] Step 2: Input the bronchoscope image into a pre-trained image segmentation model to obtain a labeled image of the bronchoscope channel region.

[0040] The image segmentation model includes an encoder module, a decoder module, and a skip connection layer.

[0041] The encoder module includes a multi-layer encoding module, and each layer encoding module includes an ODConv module and an EMA module. The bronchoscopic image is input into the multi-layer encoding module, and is convolved by the first layer ODConv module to obtain the first output feature. The first output feature of each layer is input into the EMA module of each layer and the next layer encoding module respectively. Each layer EMA module extracts features from the first output feature to obtain the second output feature. The next layer encoding module downsamples the first output feature of the previous layer and inputs it into the ODConv module of the next layer.

[0042] Specifically, the ODConv (Omni-Dimensional Dynamic Convolution) module dynamically adjusts the shape and size of the convolution kernel based on the features of the input data to adapt to different input data. The first layer of the ODConv module performs a convolution operation on the bronchoscopic images, as shown below:

[0043] ;

[0044] in, This represents the first output feature of the first-layer ODConv module. This represents the nth convolutional kernel of the first-layer ODConv module. Represents the convolution kernel Attention scalar Indicates the convolution kernel Attention scalar in the spatial dimension Indicates the convolution kernel Attention scalar at the input channel dimension. Indicates the convolution kernel The attention scalar of the output channel dimension, where X represents the bronchoscopic image.

[0045] The first output feature of the ODConv module in the i-th layer coding module is represented as follows: The first output feature of the i-th layer coding module Input the EMA module in the i-th layer encoding module and the next layer encoding module respectively;

[0046] Except for the first layer encoding module, all other layers of encoding modules include a max pooling layer, which is used to reduce the resolution of the first output feature of the previous layer to half of its original value, and represent it as the first input feature. The (i+1)th layer ODConv module processes the first input features Performing a convolution operation is represented as:

[0047] ;

[0048] in, This represents the first output feature of the (i+1)th layer ODConv module. Represents the first input feature of the i-th layer ODConv module .

[0049] The EMA (Efficient Multi-Scale Attention) module comprises three branches, namely two... Branch and one Branches; The branch performs a one-dimensional global average pooling operation on the first output feature. The branch performs the first output feature Convolution concatenates the outputs of the three branches to obtain the second output feature. The second output feature of the EMA module in the i-th layer encoding module is represented as... .

[0050] Furthermore, One-dimensional global average pooling is used in the branch to encode channel information in two spatial directions respectively. Branch through Convolution captures multi-scale feature representations. The EMA module not only encodes information across channels to adjust the importance of different channels, but also preserves precise spatial structure information within the channels. The output of the branch encodes global spatial information using two-dimensional global average pooling, while The outputs of each branch are directly converted to their corresponding dimensional shapes. These outputs are then aggregated using a matrix dot product operation to generate a second output feature.

[0051] like Figure 2 As shown, in this embodiment, the encoder module includes a 5-layer encoding module. The bronchoscopic images are captured using a three-channel array. A pixel-sized image, after passing through the ODConv module in the first layer encoding module, has 64 channels, resulting in the first output feature. First output feature The input to the second-layer encoding module is transformed into the first input feature after passing through the max-pooling layer. First input feature The first input feature is obtained after inputting into the ODConv module of the second layer. Repeat the above steps until the first input feature of the fourth-layer ODConv module is reached. Input the fifth layer encoding module.

[0052] The skip connection layer is used to input the second output feature of each layer into the decoder module. For example... Figure 2As shown, the skip connection layer includes an attention gate. In this embodiment, the encoding modules from the first to the fourth layer are all skip-connected to the attention gate, and the second output feature output by the EMA module in the encoding modules from the first to the fourth layer... The input is fed into the corresponding attention gate. The second output feature is the output of the EMA module in the fifth-layer encoding module. Input into the attention gate of the fourth layer.

[0053] The decoder module includes a multi-layer decoding module, each of which includes an ODConv module and an EMA module. The first-layer decoding module upsamples the second output feature of the last-layer encoding module and splices the second output features of the last two encoding modules to obtain a spliced ​​feature. The spliced ​​feature is then processed by the ODConv module and the EMA module to obtain a third output feature. The remaining layer decoding modules upsample the third output feature of the previous layer and splice the third output feature of the previous layer with the second output feature of the corresponding encoding module until the last layer decoding module outputs a marked image of the bronchoscope channel region.

[0054] Furthermore, the first-layer decoding module upsamples the second output feature of the last-layer encoding module to double its resolution, thus obtaining the second input feature. Except for the last-layer encoding module, each encoding module corresponds to a skip connection layer, which is used to input the second output feature of that layer into the corresponding decoding module. The second output feature of the last-layer encoding module is input into the skip connection layer of the previous layer, concatenated with the second output feature of the previous layer, and then input into the first-layer decoding module to concatenate the second input feature, thus obtaining the concatenated feature.

[0055] like Figure 2 As shown, in this embodiment, the decoder module includes a 4-layer decoding module. Second output feature The inputs are respectively the skip connection layer corresponding to the fourth layer encoding module and the first layer decoding module. In the attention gate of the skip connection layer, the second output feature... splicing the second output feature Output , It contains attention information from both spatial and channel perspectives, as well as interaction information from both spatial and channel perspectives, and also possesses the localization capability of the skip connection layer itself. In the first-layer decoding module, the second output feature... Upsampling is performed to obtain the second input feature. .Will and After concatenation, the number of channels becomes 1024. Then, the signals are input into the ODConv and EMA modules in the first-layer decoding module to obtain the third output feature. Third output feature Input the skip connection layer corresponding to the third layer encoding module and the second layer decoding module respectively, and repeat the above concatenation and upsampling processes to obtain... Second input features ,Will and The third output feature is obtained by concatenating the ODConv and EMA modules in the second-layer decoding module. Repeat the above process until the fourth-layer decoding module outputs a marked image of the bronchoscope channel region. .

[0056] Example 2

[0057] This embodiment, based on Embodiment 1, compares the image segmentation model with the traditional U-Net and Attention U-Net network models.

[0058] This embodiment compares four metrics: Intersection over Union (IoU), Dice similarity coefficient (DSC), Hausdorff distance (HD), and precision (PRE).

[0059] Intersection over Union (IoU) and Precision over Union (PRE) are metrics based on overlap, ranging from [0, 1]. Higher values ​​indicate greater overlap, meaning the predicted results are closer to the true labels. IoU and PRE are expressed as follows:

[0060] ;

[0061] ;

[0062] In this context, TP represents a region that is predicted to be a lesion and is actually a lesion, FP represents a region that is predicted to be a lesion but is not actually a lesion, TN represents a region that is predicted not to be a lesion and is not actually a lesion, and FN represents a region that is predicted not to be a lesion but is actually a lesion.

[0063] Hausdorff distance (HD) measures the maximum mismatch between set A and set B, primarily used to measure the mismatch between edges of identified targets. It is represented as:

[0064] ;

[0065] in, Describes the 2-norm. Let represent the maximum distance from point a in set A to point b in the nearest set B. This represents the maximum distance from point b in set B to point a in the nearest set A.

[0066] Precision (PRE) represents the proportion of samples predicted as lesions that actually contain lesions. It is expressed as:

[0067] .

[0068] Table 1 shows the performance of the image segmentation model compared to the traditional U-Net and Attention U-Net network models on bronchoscopic images.

[0069] Table 1. Comparison of indicators of the three models on bronchoscopic images.

[0070] Model IoU DSC HD PRE U-Net 0. 6964 0.7883 2. 5378 0. 8463 Attention U-Net 0. 6967 0.7915 2. 5142 0. 8548 Image segmentation model 0.7353 0.8314 2.4214 0.8893

[0071] As shown in Table 1, the improved image segmentation model shows improvements across all metrics. In the IoU metric, which reflects overlap, the image segmentation model outperforms the U-Net model by 0.0389 and the Attention U-Net model by 0.0386. In the DSC metric, another indicator of overlap, the image segmentation model outperforms the U-Net model by 0.0431 and the Attention U-Net model by 0.0399, indicating a noticeable improvement in overlap compared to the unimproved traditional model. In the HD metric, which reflects edge mismatch, the image segmentation model reduces the score by 0.1164 compared to the U-Net model and by 0.0928 compared to the Attention U-Net model, both showing reductions and improved matching accuracy. In terms of precision, the image segmentation model outperforms the U-Net model by 0.0430 and the Attention U-Net model by 0.0345, also demonstrating improved precision.

[0072] Example 3

[0073] This embodiment provides a bronchoscopic image segmentation system, including:

[0074] The image acquisition module is used to acquire bronchoscopic images.

[0075] The segmentation module is used to input the bronchoscopic image into a pre-trained image segmentation model to obtain a labeled image of the bronchoscopic aperture region.

[0076] The image segmentation model includes an encoder module, a decoder module, and a skip connection layer.

[0077] The encoder module includes a multi-layer encoding module, and each layer encoding module includes an ODConv module and an EMA module. The bronchoscopic image is input into the multi-layer encoding module, and is convolved by the first layer ODConv module to obtain the first output feature. The first output feature of each layer is input into the EMA module of each layer and the next layer encoding module respectively. Each layer EMA module extracts features from the first output feature to obtain the second output feature. The next layer encoding module downsamples the first output feature of the previous layer and inputs it into the ODConv module of the next layer.

[0078] The skip connection layer is used to input the second output feature of each layer into the decoder module.

[0079] The decoder module includes a multi-layer decoding module, each of which includes an ODConv module and an EMA module. The first-layer decoding module upsamples the second output feature of the last-layer encoding module and splices the second output features of the last two encoding modules to obtain a spliced ​​feature. The spliced ​​feature is then processed by the ODConv module and the EMA module to obtain a third output feature. The remaining layer decoding modules upsample the third output feature of the previous layer and splice the third output feature of the previous layer with the second output feature of the corresponding encoding module until the last layer decoding module outputs a marked image of the bronchoscope channel region.

[0080] Example 4

[0081] This embodiment provides a bronchoscopic image segmentation device, including a processor and a storage medium. The storage medium is used to store instructions. The processor is used to perform operations according to the instructions to execute the steps of the bronchoscopic image segmentation method described in Embodiment 1.

[0082] Example 5

[0083] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the bronchoscopic image segmentation method described in Embodiment 1.

[0084] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A bronchoscope image segmentation method, characterized by, The method comprises the following steps: obtaining a bronchoscope image; inputting the bronchoscope image into a pre-trained image segmentation model to obtain a labeled image of a bronchoscope channel region; wherein the image segmentation model comprises an encoder module, a decoder module and a skip connection layer; the encoder module comprises multiple layers of encoding modules, each layer of encoding module comprising an ODConv module and an EMA module; the bronchoscope image is input into the multiple layers of encoding modules, and a first output feature is obtained through convolution of the first layer of ODConv module; the first output feature of each layer is input into each layer of EMA module and the next layer of encoding module respectively; each layer of EMA module extracts features from the first output feature to obtain a second output feature; and the first output feature of the previous layer is input into the next layer of ODConv module after down-sampling by the next layer of encoding module; the skip connection layer is used for inputting the second output feature of each layer into the decoder module respectively; the decoder module comprises multiple layers of decoding modules, each layer of decoding module comprising an ODConv module and an EMA module; the second output feature of the last layer of encoding module is up-sampled by the first layer of decoding module, and the second output features of the last two layers of encoding module are spliced to obtain a spliced feature; the spliced feature is input into the ODConv module and the EMA module to obtain a third output feature; the third output feature of the previous layer is up-sampled by the remaining layers of decoding module, and the third output feature of the previous layer and the second output feature of the corresponding encoding module are spliced until the last layer of decoding module outputs the labeled image of the bronchoscope channel region; wherein the bronchoscope image is input into the multiple layers of encoding modules, and a first output feature is obtained through convolution of the first layer of ODConv module, comprising: the first layer of ODConv module performs convolution operation on the bronchoscope image, which is represented as: ； wherein, represents a first output feature output by the first layer ODConv module, represents an nth convolution kernel of the first layer ODConv module, represents an attention scalar of the convolution kernel , represents an attention scalar of the convolution kernel in the spatial dimension, represents an attention scalar of the convolution kernel in the input channel dimension, represents an attention scalar of the convolution kernel in the output channel dimension, and X represents a bronchoscopy image. wherein the first output feature of each layer is input into each layer of EMA module and the next layer of encoding module respectively, and the first output feature of the previous layer is input into the next layer of ODConv module after down-sampling by the next layer of encoding module, comprising: The first output feature representation output by the ODConv module in the i-th layer encoding module is represented as The first output feature of the i-th layer encoding module is input into the EMA module in the i-th layer encoding module and the next layer encoding module, respectively. The first output feature of the i-th layer encoding module is input into the EMA module in the i-th layer encoding module and the next layer encoding module, respectively. In addition to the first layer encoding module, the remaining layer encoding modules each include a max-pooling layer for reducing the resolution of the first output feature of the previous layer to half of the original, denoted as the first input feature The i+1th layer ODConv module performs convolution operation on the first input feature , denoted as: ； wherein, represents the first output feature of the i+1th layer ODConv module output, represents the first input feature of the i+1th layer ODConv module .

2. The bronchoscopic image segmentation method of claim 1, wherein, each layer of EMA module extracts features from the first output feature to obtain a second output feature, comprising: The EMA module includes three branches, two branches and one branch; The first branch performs one-dimensional global average pooling operation on the first output feature, The second branch performs convolution on the first output feature, and the output results of the three branches are spliced to obtain a second output feature. The second output feature representation outputted by the EMA module in the i-th layer encoding module is denoted as .

3. The bronchoscopic image segmentation method of claim 1, wherein, the first layer of decoding module up-samples the second output feature of the last layer of encoding module to expand the resolution of the second output feature of the last layer of encoding module to twice the original resolution to obtain a second input feature; each layer of encoding module except the last layer of encoding module corresponds to a skip connection layer, which is used for inputting the second output feature of the encoding module into the corresponding decoding module; the second output feature of the last layer of encoding module is input into the skip connection layer of the previous layer, spliced with the second output feature of the previous layer, and then input into the first layer of decoding module to splice the second input feature to obtain a spliced feature. The method comprises the following steps:

4. A bronchoscope image segmentation system, comprising: an image acquisition module is configured to obtain a bronchoscope image; ​ The segmentation module is configured to input the bronchoscope image into a pre-trained image segmentation model to obtain a labeled image of a bronchoscope channel region. The image segmentation model comprises an encoder module, a decoder module, and a skip connection layer. The encoder module comprises multiple layers of encoding modules, each of which comprises an ODConv module and an EMA module. The bronchoscope image is input into the multiple layers of encoding modules, and the first layer of ODConv modules performs convolution to obtain first output features. The first output features of each layer are input into each layer of EMA modules and the next layer of encoding modules, respectively. Each layer of EMA modules extracts features from the first output features to obtain second output features, and the next layer of encoding modules inputs the first output features of the previous layer into the next layer of ODConv modules after down-sampling. The skip connection layer is configured to input the second output features of each layer into the decoder module. The decoder module comprises multiple layers of decoding modules, each of which comprises an ODConv module and an EMA module. The first layer of decoding modules up-samples the second output features of the last layer of encoding modules and splices the second output features of the last two layers of encoding modules to obtain spliced features. The spliced features are input into the ODConv module and the EMA module to obtain third output features. The remaining layers of decoding modules up-sample the third output features of the previous layer and splice the third output features of the previous layer and the second output features of the corresponding encoding modules until the last layer of decoding modules outputs the labeled image of the bronchoscope channel region. The bronchoscope image is input into the multiple layers of encoding modules, and the first layer of ODConv modules performs convolution to obtain first output features, including: The first layer of ODConv modules performs convolution on the bronchoscope image, which is represented as: ； wherein, represents a first output feature output by the first layer ODConv module, represents an nth convolution kernel of the first layer ODConv module, represents an attention scalar of the convolution kernel , represents an attention scalar of the convolution kernel in the spatial dimension, represents an attention scalar of the convolution kernel in the input channel dimension, represents an attention scalar of the convolution kernel in the output channel dimension, and X represents a bronchoscopy image. The first output features of each layer are input into each layer of EMA modules and the next layer of encoding modules, and the next layer of encoding modules inputs the first output features of the previous layer into the next layer of ODConv modules after down-sampling, including: The first output feature representation output by the ODConv module in the i-th layer encoding module is represented as The first output feature of the i-th layer encoding module is input into the EMA module in the i-th layer encoding module and the next layer encoding module, respectively; In addition to the first layer encoding module, the remaining layer encoding modules each include a max-pooling layer for reducing the resolution of the first output feature of the previous layer to half of the original, denoted as the first input feature The i+1th layer ODConv module performs a convolution operation on the first input feature , denoted as: ； wherein, represents the first output feature of the (i+1)th layer ODConv module output, represents the first input feature of the i-th layer ODConv module .

5. A bronchoscopic image segmentation apparatus characterized by comprising: The system comprises a processor and a storage medium. The storage medium is configured to store instructions. The processor is configured to operate according to the instructions to perform the steps of the bronchoscope image segmentation method of any one of claims 1-3.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the bronchoscope image segmentation method of any one of claims 1-3.

Images