Multi-branch collaborative network-based uncertainty perception pulmonary nodule segmentation method

Abstract

The invention provides an uncertainty perception pulmonary nodule segmentation method based on a multi-branch collaborative network, and the method introduces an expert conflict perception mechanism, and enables a model to learn consensus information and divergence features between scorers at the same time through constructing a probability graph soft label and a conflict graph based on multi-scorer labeling. Therefore, the uncertainty existing in clinical practice can be reflected more comprehensively; a multi-branch collaborative segmentation network structure is provided, an encoder-decoder trunk is combined with three functional branches of consensus, conflict and boundary, and adaptive weighted integration is realized through a conflict regulation and control fusion module. The method not only improves the overall precision of the segmentation result, but also can generate probability distribution prediction which better meets the actual diagnosis demand under the condition that the opinions of multiple scorers are inconsistent, and provides a more reliable basis for clinical auxiliary decision making.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular, and particularly relates to an uncertainty-aware lung nodule segmentation method based on a multi-branch collaborative network. BACKGROUND

[0002] Lung cancer is one of the highest incidence and mortality rates of malignant tumors in the world. Clinical practice shows that through imaging screening and early diagnosis and treatment, the risk of lung cancer-related death can be significantly reduced. In the screening process, the detection and evaluation of lung nodules mainly rely on the comprehensive judgment of medical features such as nodule size, shape and location, and lung nodule segmentation as the core step to obtain these features is of great significance for computer-aided diagnosis. Especially in the case of small nodule volume or blurred boundary, accurate segmentation not only helps to distinguish between benign and malignant nodules, but also provides a reliable basis for subsequent diagnosis and treatment planning.

[0003] In the lung nodule segmentation task, due to the differences between doctors in professional experience, diagnosis habits and interpretation standards, the same nodule will often have significant inconsistencies in the boundary and complex morphology area. This expert conflict not only manifests as differences in labeling results, but also reflects the uncertainty in image interpretation. Traditional methods usually integrate multiple annotators' labels into a single reference label through majority voting or label fusion algorithms to avoid expert disagreement, but this approach cannot truly depict the inherent subjective differences between doctors, which limits the performance of the model in high-uncertainty areas. In recent years, some deep learning methods have attempted to introduce uncertainty modeling during segmentation to alleviate the impact of multiple annotator differences. However, these methods mostly stay at the overall statistical level and lack explicit modeling of local structural differences, especially in the expression of uncertainty in lung nodule boundary areas.

[0004] Therefore, the existing lung nodule segmentation methods still have obvious limitations in handling multiple annotator labeling differences, capturing complex nodule features, and improving clinical credibility. There is an urgent need for a new segmentation method and system that can explicitly integrate expert consensus and conflict information through a multi-branch collaborative modeling mechanism, sensitively capture uncertain areas during segmentation, and accurately depict complex nodules, thereby improving the stability of the segmentation results and the clinical application value. SUMMARY

[0005] As described above, the existing lung nodule segmentation method has the following problems in dealing with multi-rater annotation divergence: majority voting or label fusion is difficult to reflect the subjective differences between doctors; the existing uncertainty method is mostly based on overall statistics, and lacks explicit modeling of structural uncertainty regions such as boundaries, and the segmentation stability and accuracy at complex shapes and fuzzy boundaries cannot be guaranteed. How to explicitly integrate the "expert consensus" and "expert conflict" information within the network without changing the clinical workflow, and to structure the uncertain regions, is still a challenging problem.

[0006] In order to make up for the shortcomings of the prior art, the present application provides an uncertainty-aware lung nodule segmentation method based on a multi-branch collaborative network, which realizes sensitive capture and accurate characterization of uncertain regions without relying on a single label through collaborative learning and conflict regulation fusion mechanism of consensus branch-conflict branch-boundary branch.

[0007] The present application is realized by the following technical scheme: an uncertainty-aware lung nodule segmentation method based on a multi-branch collaborative network, specifically comprising the following steps: S1: constructing a data set; the training and test data sets are composed of original CT images containing lung nodules, corresponding probability graph soft labels, conflict graphs, edge graphs and original independent annotations of all radiologists; the probability graph soft label is obtained by averaging the pixel-level masks of multiple raters, and is used to reflect the overall consensus information; the conflict graph is constructed based on the normalized Bernoulli variance function, and is used to quantify and highlight the disagreement degree of different experts in the spatial position; the edge graph is obtained by calculating the gradient amplitude of the probability graph soft label , and is used to represent the outline and boundary information of the nodule; the original independent annotations of all radiologists will be used for multi-annotator performance evaluation in the final test stage; S2: designing a multi-branch collaborative segmentation network model; the model takes Attention U-Net with an encoder-decoder architecture as the backbone network, and introduces a task decoupling mechanism at the back end to derive three branches with different functions: (i) consensus branch: output consensus probability prediction graph ; (ii) conflict branch: output pixel-level conflict saliency map ; (iii) boundary branch: output edge feature map ; Based on this, the outputs of the three branches are jointly input into the conflict control and fusion module. This module dynamically generates fusion weights based on the pre-constructed conflict graph and the uncertainty of the model prediction, and adaptively weights and integrates the results of the three branches to finally obtain the global probability prediction graph. ; S3: Design the post-processing module; to obtain the final binarized segmentation result. In the global probability prediction graph Further thresholding and post-processing mechanisms are introduced. S4: Training the model; end-to-end training is performed using the RMSprop optimizer, combined with cosine annealing SGDR learning rate scheduling; during training, the parameters of the encoder-decoder backbone, consensus branch, conflict branch, boundary branch, and conflict control fusion module are jointly optimized; training stops when the validation set performance converges; the training objective is to minimize the composite loss function consisting of consensus consistency, conflict preservation, and boundary alignment, as shown below: in: Consensus consensus probability prediction graph to account for consensus consistency loss. Consistent with the soft label P in the probability map; To account for conflict losses, weights are introduced. , with conflict diagram Based on parameters By controlling the weight strength of conflict regions in the loss function, the supervisory role of high-conflict regions is strengthened, and the prediction uncertainty is kept consistent with the empirical variance through KL divergence. For boundary alignment loss, the gradient magnitude map of the probabilistic soft label is used. As a supervisor, the boundary branch output is constrained. Align with it; total loss function It is composed of a weighted combination of three sub-losses, among which These are the weighting coefficients; S5: Test Model; During the testing phase, CT images containing lung nodules are input into the trained model, and the final segmentation results are output. Subsequently, the final results are compared with the original independent annotations from all radiologists, and the average value of the final performance index is taken as the comprehensive evaluation result. The segmentation performance is quantitatively evaluated using metrics including Dice coefficient (DICE), intersection-over-union ratio (IoU), sensitivity (SE), specificity (SP), precision (PC), accuracy (ACC), and normalized surface Dice index. S6: using the model; input the real lung nodule CT image into the trained segmentation network model, and the final binary segmentation result can be directly obtained .

[0008] As a preferred solution, step S2 specifically comprises the following steps: S2-1: constructing a multi-branch collaborative segmentation network model; the model takes Attention U-Net with an encoder-decoder architecture as the backbone network, introduces a task decoupling mechanism at the back end, derives three branches with different functions, and finally integrates the outputs of each branch through a conflict regulation fusion module to generate the final probability map prediction result; S2-2: input the original lung nodule CT image into the Attention U-Net backbone network with an encoder-decoder architecture for feature extraction and upsampling to obtain a high-level semantic feature map ; S2-3: introducing a task decoupling mechanism at the end of the decoder, and inputting the high-level semantic feature map as input into three parallel branches for further processing; S2-4: the consensus branch takes the high-level semantic feature map as input, which successively passes through two 3x3 convolution layers, a batch normalization Batch Normalization layer and a ReLU activation function, and finally outputs a consensus probability prediction map through a 1x1 convolution layer and a Sigmoid activation function ; the consensus branch process is defined by the following formula: wherein, represents a convolution layer with a 3x3 convolution kernel, represents a batch normalization layer, represents a convolution layer with a 1x1 convolution kernel, represents a Sigmoid activation function; S2-5: the conflict branch takes the high-level semantic feature map as input, and its uncertainty estimation module consists of two 3x3 atrous convolution (dilation rate = 2) layers, a batch normalization layer and a LeakyReLU activation function, and finally outputs a pixel-level conflict saliency map through a 1x1 convolution and a Sigmoid function . This structure uses atrous convolution to expand the receptive field, fully capturing long-distance dependencies and divergent contexts, and is used to quantify the annotation differences of different experts in nodule boundaries and difficult areas. The conflict branch process is defined by the following formula: wherein, represents a dilated convolution (dilation rate = 2) layer with a 3x3 kernel, represents a batch normalization layer, represents a convolution layer with a 1x1 kernel, represents a Sigmoid activation function; S2-6: The boundary branch takes the high-level semantic feature map as input, and processes it through a gradient feature enhancement layer, which first extracts the gradient feature map in the x and y directions using Sobel operators, concatenates the obtained gradient feature map with the original high-level semantic feature map , and then passes it through a layer of 3x3 convolution, batch normalization, and ReLU activation function, and finally outputs the edge feature map through a 1x1 convolution. This branch explicitly introduces gradient information to enhance the expression of high-frequency details such as nodule contours, burrs, and small protrusions. The boundary branch process is defined by the following formula: wherein, , is the Sobel operator convolution operation in the x and y directions, is the feature concatenation operation along the channel dimension, represents a convolution layer with a 3x3 kernel, represents a batch normalization layer, represents a convolution layer with a 1x1 kernel; S2-7: The consensus branch output , the conflict branch output , the boundary branch output , and the pre-constructed conflict map are all input into the conflict regulation fusion module; S2-8: In the conflict regulation fusion module, based on the input conflict map and the pixel-level conflict saliency map , a spatially adaptive fusion weight map is calculated through a weight generation network composed of two layers of 3x3 convolution layers ; wherein the first layer of convolution is followed by a ReLU activation function, and the second layer of convolution is followed by a Sigmoid activation function, and the calculation process is represented as: , and represents a channel splicing operation; S2-9: using the generated weight map to adaptively weight the outputs of the three branches to generate a final global probability prediction map . The specific fusion formula is: , wherein represents pixel-by-pixel multiplication, and are preset hyperparameters, and take values of 0.6 and 0.4, respectively.

[0009] Further, in step S2-9, and take values of 0.6 and 0.4, respectively.

[0010] As a preferred solution, step S3 specifically comprises the following steps: S3-1: when the pixel value in the global probability prediction map is greater than or equal to a threshold value , it is determined as a nodule region, otherwise it is determined as a background region; wherein the threshold value is initially set to 0.5, and the final threshold value is determined by grid search on the validation set to make the Dice similarity coefficient maximum; S3-2: on the basis of thresholding, the stability and clinical usability of the results are improved through a post-processing step.

[0011] Further, step S3-2 specifically comprises the following steps: S3-2-1: using three-dimensional connected component analysis to remove small false positive regions with a number of voxels less than a preset threshold value; S3-2-2: smoothing the segmentation edge through morphological closing operation.

[0012] As a preferred solution, in step S4, , , , , .

[0013] Compared with the prior art, the present application has the following beneficial effects: the present application mainly includes two aspects, one is to introduce an expert conflict perception mechanism, by constructing a probability graph soft label based on multi-rater labeling and a conflict graph, the model can learn the consensus information and divergence features between raters at the same time, so as to more comprehensively reflect the uncertainty existing in clinical practice, and the other is to propose a multi-branch collaborative segmentation network structure, which combines the encoder-decoder main body with the consensus, conflict and boundary three functional branches, and realizes adaptive weighted integration through the conflict regulation fusion module. The design can fully utilize the accurate prediction of the consensus branch on the stable area, the explicit modeling of the conflict branch on the high uncertainty area and the enhanced expression of the boundary branch on the details and contours, so as to better depict the spatial consistency and difference of the nodule area in the lung nodule segmentation task. Compared with the prior art, the present application not only improves the overall accuracy of the segmentation result, but also generates a more actual diagnosis demand probability distribution prediction under the condition that the opinions of multiple raters are inconsistent, thereby providing a more reliable basis for clinical auxiliary decision-making.

[0014] Additional aspects and advantages of the application will become apparent from the following description, or will be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0015] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 As a data set form; Figure 2 As a method flowchart of the present application; Figure 3 As a network structure schematic diagram of the present application. DETAILED DESCRIPTION

[0016] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0017] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0018] The following will be described in combination with Figures 1 to 3 The multi-branch collaborative network based uncertainty perception lung nodule segmentation method of the embodiments of the present application will be specifically described.

[0019] This invention addresses the problem that existing segmentation methods struggle to effectively model consensus and conflicting information among multiple raters when lung nodules exhibit significant differences in annotations by different experts, complex nodule morphology, and ambiguous boundaries. It employs multi-branch collaborative learning and conflict adaptive fusion techniques to explicitly characterize the structural uncertainty of nodules, generating segmentation results that more closely resemble the annotation styles of the experts. Figure 2 As shown, this invention proposes an uncertainty-aware lung nodule segmentation method based on a multi-branch collaborative network, specifically including the following steps: S1: Construct the dataset, such as Figure 1 As shown; both the training and testing datasets consist of original CT images containing lung nodules, corresponding probabilistic soft labels, conflict maps, edge maps, and original independent annotations from all radiologists; probabilistic soft labels Obtained by averaging pixel-level masks from multiple raters, used to reflect overall consensus information; Conflict graph It is constructed based on the normalized Bernoulli variance function to quantify and highlight the degree of disagreement among different experts regarding spatial location; marginal plot Then, by using soft labels on the probability map The gradient magnitude is calculated to characterize the contour and boundary information of the nodules; the original independent annotations from all radiologists will be used for multi-annotator performance evaluation in the final testing phase. S2: Design a multi-branch collaborative segmentation network model, such as Figure 3 As shown; the model uses the Attention U-Net with an encoder-decoder architecture as the backbone network, and introduces a task decoupling mechanism at its backend, resulting in three branches with different functions: (i) Consensus Branch: Output consensus probability prediction graph This is used to characterize areas with high consistency among multiple raters, ensuring that the model has stable recognition capabilities in the core nodule region; (ii) Conflict Branch: Output pixel-level conflict saliency map It focuses on modeling expert disagreements and ambiguous areas, and improves the model's adaptability to complex and difficult-to-determine regions by explicitly characterizing uncertainty; (iii) Boundary branch: Output edge feature map It is used to capture the high-frequency texture and directional details of nodules, thereby enhancing the ability to express nodule edges and small spinous processes; Based on this, the outputs of the three branches are jointly input into the conflict control and fusion module. This module dynamically generates fusion weights based on the pre-constructed conflict graph and the uncertainty of the model prediction, and adaptively weights and integrates the results of the three branches to finally obtain the global probability prediction graph. ; Specifically, the following steps are included: S2-1: Construct a multi-branch collaborative segmentation network model; the model takes Attention U-Net with an encoder-decoder architecture as the backbone network, introduces a task decoupling mechanism at the back end, derives three branches with different functions, and finally integrates the outputs of each branch through a conflict regulation fusion module to generate the final probability map prediction result; S2-2: Input the original lung nodule CT image into the Attention U-Net backbone network with an encoder-decoder architecture for feature extraction and upsampling to obtain a high-level semantic feature map ; S2-3: Introduce a task decoupling mechanism at the end of the decoder, and input the high-level semantic feature map into three parallel branches for further processing; S2-4: The consensus branch takes the high-level semantic feature map as input, which is sequentially passed through two 3x3 convolution layers, a batch normalization layer and a ReLU activation function, and finally output a consensus probability prediction map through a 1x1 convolution layer and a Sigmoid activation function ; This branch enhances the feature expression ability through dense convolution operation, and is used to reliably represent the core region of the nodule with high consistency of multiple raters; the consensus branch process is defined by the following formula: wherein, represents a convolution layer with a 3x3 convolution kernel, represents a batch normalization layer, represents a convolution layer with a 1x1 convolution kernel, represents a Sigmoid activation function; S2-5: The conflict branch takes the high-level semantic feature map as input, and its uncertainty estimation module consists of two 3x3 atrous convolution (dilation rate = 2) layers, a batch normalization layer and a LeakyReLU activation function, and finally outputs a pixel-level conflict saliency map through a 1x1 convolution and a Sigmoid function . This structure uses atrous convolution to expand the receptive field, fully capturing long-distance dependencies and divergent contexts, and is used to quantify the annotation differences of different experts in the nodule boundary and difficult areas. The conflict branch process is defined by the following formula: wherein, denotes a 3x3 dilated convolution (dilation rate = 2) layer, denotes a batch normalization layer, denotes a 1x1 convolution layer, denotes a Sigmoid activation function; S2-6: The boundary branch takes the high-level semantic feature map as input, and processes it through a gradient feature enhancement layer, which first extracts a high-level semantic feature map of gradient magnitude in the x and y directions, concatenates the resulting gradient feature with the original high-level semantic feature map , and then passes it through a layer of 3x3 convolution, batch normalization, and ReLU activation function, and finally outputs an edge feature map through a 1x1 convolution . This branch explicitly introduces gradient information to enhance the expression of high-frequency details such as nodule contours, burrs, and small protrusions. The boundary branch process is defined by the following formula: wherein, , is the Sobel operator convolution operation in the x and y directions, is the feature concatenation operation along the channel dimension, denotes a 3x3 convolution layer, denotes a batch normalization layer, denotes a 1x1 convolution layer; S2-7: The outputs of the consensus branch , the outputs of the conflict branch , the outputs of the boundary branch , and the pre-constructed conflict map are all input into the conflict regulation fusion module; S2-8: In the conflict regulation fusion module, based on the input conflict map and the pixel-level conflict saliency map , a spatially adaptive fusion weight map is calculated through a weight generation network composed of two layers of 3x3 convolution layers; the first layer uses ReLU activation function after convolution, and the second layer uses Sigmoid activation function after convolution, and the calculation process is represented as: , and represents the channel concatenation operation; S2-9: Utilize the generated weight map Adaptive weighted fusion of the outputs of the three branches to generate the final global probability prediction map The specific fusion formula is: Wherein represents pixel-by-pixel multiplication, and are preset hyperparameters, respectively taking values of 0.6 and 0.4.

[0020] S3: Design a post-processing module, as shown in Figure 3 To obtain the final binary segmentation result , further introduce thresholding and post-processing mechanism on the global probability prediction map ; Specifically, the following steps are included: S3-1: When the pixel value in the global probability prediction map is greater than or equal to the threshold value , it is determined as a nodule region, otherwise it is determined as a background region; wherein the threshold value is initially set to 0.5, and the final threshold value is determined by grid search on the validation set to maximize the Dice similarity coefficient, thereby obtaining the optimal segmentation performance; S3-2: On the basis of thresholding, the stability and clinical usability of the results are improved through post-processing steps, which include the following steps: S3-2-1: Remove small false positive regions with a number of voxels less than a preset threshold using three-dimensional connected component analysis, thereby reducing the false detection rate; S3-2-2: Smooth the segmentation edges by morphological closing operation to eliminate small cracks and ensure the integrity of the nodule contour, making the final binary segmentation result more stable, accurate, and providing a reliable basis for subsequent lung nodule detection and clinical diagnosis.

[0021] S4: Train the model; use the RMSprop optimizer for end-to-end training, and cooperate with the cosine annealing SGDR learning rate scheduling; In the training process, the parameters of the encoder-decoder backbone, the consensus branch, the conflict branch, the boundary branch, and the conflict regulation fusion module are jointly optimized; Stop training with the performance convergence of the validation set as the criterion; The training target is to minimize the composite loss function composed of consensus consistency, conflict preservation, and boundary alignment, as shown below: Wherein: is the consensus consistency loss, which constrains the consensus probability prediction map to be consistent with the probability map soft label Keep consistent; Introduce weights for conflict loss , based on conflict map , control the weight intensity of the conflict area in the loss function through parameters , so as to strengthen the supervision of high conflict areas, and keep the prediction uncertainty consistent with the empirical variance through KL divergence; For boundary alignment loss, the gradient amplitude map of the probability map soft label is used As supervision, constrain the output of the boundary branch Align with it; the total loss function Is composed of three sub-losses weighted combination, wherein Is the weight coefficient, the present application can take , , , , ; S5: test the model; in the test stage, input the CT image containing lung nodules into the trained model, and output the final segmentation result; then, compare the final result with the original independent labeling of all radiologists, and take the average value of the final performance index as the comprehensive evaluation result; quantitative evaluation of segmentation performance is carried out by using indexes including Dice coefficient DICE, intersection over union IoU, sensitivity SE, specificity SP, precision PC, accuracy ACC and normalized surface Dice (NSD) and the like; S6: use the model; input the real lung nodule CT image into the trained segmentation network model, and the final binary segmentation result can be directly obtained .

[0022] In the description of the present application, the term "a plurality of" refers to two or more, unless otherwise explicitly limited, the orientation or position relationship indicated by the terms "upper", "lower" and the like is based on the orientation or position relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation on the present application; the terms "connection", "installation", "fixation" and the like should be understood in a broad sense, for example, "connection" can be fixed connection, or detachable connection, or integral connection; it can be directly connected, or indirectly connected through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0023] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "certain embodiments", and the like is intended to indicate that the described implementation, feature, structure, material or characteristic is included in at least one embodiment or example of the application. The illustrative representations of the above terms in the specification are not necessarily referring to the same embodiment or example. Moreover, the described implementation, feature, structure, material or characteristic can be combined in any one or more embodiments or examples in a suitable manner.

[0024] The above only is the preferred embodiment of the present application, and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for uncertainty-aware lung nodule segmentation based on multi-branch cooperative network, characterized in that ,Specifically comprising the following steps: S1: Dataset Construction; Both the training and testing datasets consist of original CT images containing lung nodules, corresponding probabilistic soft-labels, conflict maps, edge maps, and original independent annotations from all radiologists; probabilistic soft-labels Obtained by averaging pixel-level masks from multiple raters, used to reflect overall consensus information; Conflict graph It is constructed based on the normalized Bernoulli variance function to quantify and highlight the degree of disagreement among different experts regarding spatial location; edge plot Then, by using soft labels on the probability map The gradient magnitude is calculated to characterize the contour and boundary information of the nodules; the original independent annotations from all radiologists will be used for multi-annotator performance evaluation in the final testing phase. S2: design a multi-branch collaborative segmentation network model; the model takes Attention U-Net with an encoder-decoder architecture as the backbone network, and introduces a task decoupling mechanism at the back end to derive three branches with different functions: (i) consensus branch: output consensus probability prediction map ; (ii) conflict branch: output pixel-level conflict saliency map ; (iii) boundary branch: outputting edge feature maps ; On this basis, the three-branch outputs are jointly input to a conflict regulation fusion module; the module dynamically generates fusion weights according to a pre-constructed conflict graph and model-predicted uncertainty, and adaptively weights and integrates the three-branch results, to finally obtain a global probability prediction graph ; S3: design post-processing module; to obtain the final binary segmentation result In the global probability prediction map Further introduce thresholding and post-processing mechanism; S4: train the model; use the RMSprop optimizer for end-to-end training, and cooperate with the cosine annealing SGDR learning rate scheduling; during the training process, the parameters of the encoder-decoder backbone, the consensus branch, the conflict branch, the boundary branch, and the conflict regulation fusion module are optimized jointly; the training is stopped when the performance of the validation set converges; the training target is to minimize the composite loss function composed of consensus consistency, conflict preservation, and boundary alignment, as follows: wherein: is a consensus consistency loss, constraining the consensus probability prediction map with the probability map soft label to keep consistent; is a conflict loss, introducing a weight to the conflict map as the basis, controlling the weight intensity of the conflict area in the loss function through the parameter , so as to strengthen the supervision of the high conflict area, and make the prediction uncertainty consistent with the experience variance through the KL divergence; is a boundary alignment loss, taking the gradient amplitude map of the probability map soft label as supervision, constraining the boundary branch output to align with it; the total loss function is composed of three sub-losses weighted combination, wherein is a weight coefficient; S5: test the model; in the test stage, input the CT image containing lung nodules into the trained model to output the final segmentation result; then, compare the final result with the original independent annotation of all radiologists, and take the average value of the final performance indicators as the comprehensive evaluation result; use the Dice coefficient DICE, the intersection over union IoU, the sensitivity SE, the specificity SP, the precision PC, the accuracy ACC, and the normalized surface Dice index to quantitatively evaluate the segmentation performance; S6: using the model; inputting the real lung nodule CT image into the trained segmentation network model, so as to directly obtain the final binary segmentation result .

2. The method of claim 1, wherein the method is based on a multi-branch cooperative network. ,The step S2 specifically comprises the following steps: S2-1: construct a multi-branch collaborative segmentation network model; the model takes Attention U-Net with an encoder-decoder architecture as the backbone network, and introduces a task decoupling mechanism at the back end to derive three branches with different functions, and finally integrates the outputs of each branch through a conflict regulation fusion module to generate the final probability map prediction result; S2-2: input the original lung nodule CT image into an Attention U-Net backbone network with an encoder-decoder architecture for feature extraction and upsampling to obtain a high-level semantic feature map ; S2-3: Introduce task decoupling mechanism at the decoder end, to split the high-level semantic feature maps As input, they are fed into three parallel branches for further processing; S2-4: consensus branch with high-level semantic feature map As input, it sequentially passes through two layers of 3x3 convolutional layers, Batch Normalization layers and ReLU activation functions, and finally passes through a 1x1 convolutional layer and a Sigmoid activation function, outputting a consensus probability prediction map ; the consensus branch process is defined by the following formula: wherein, denotes a convolution layer with a 3x3 convolution kernel, denotes a batch normalization layer, denotes a convolution layer with a 1x1 convolution kernel, denotes a sigmoid activation function; S2-5: conflict branch with high-level semantic feature map For input, its uncertainty estimation module consists of two 3x3 dilated convolution (dilation rate = 2) layers, a batch normalization layer and a LeakyReLU activation function, and finally a 1x1 convolution and a Sigmoid function output pixel-level conflict saliency map ; this structure uses dilated convolution to expand the receptive field, fully capturing long-distance dependencies and divergent contexts, to quantify the annotation differences of different experts in the nodule boundary and difficult areas; the conflict branch process is defined by the following formula: wherein, represents a dilated convolution (dilation rate = 2) layer with a convolution kernel of 3 x 3, represents a batch normalization layer, represents a convolution layer with a convolution kernel of 1 x 1, represents a Sigmoid activation function; S2-6: Boundary branch to high-level semantic feature map For input, it is processed by the gradient feature enhancement layer, which first extracts the high-level semantic feature map using the Sobel operator In the x and y directions, the gradient amplitude is obtained. The resulting gradient features are combined with the original high-level semantic feature map Concatenate, then pass through a layer of 3x3 convolution, batch normalization and ReLU activation function, and finally output the edge feature map through 1x1 convolution ; this branch explicitly introduces gradient information to enhance the expression of high-frequency details such as nodule contours, burrs and small protrusions; the boundary branch process is defined by the following formula: wherein, , is a Sobel operator convolution operation in x direction and y direction, is a feature concatenation operation along the channel dimension, represents a convolution layer with a 3x3 convolution kernel, represents a batch normalization layer, represents a convolution layer with a 1x1 convolution kernel; S2-7: consensus branch output , conflict branch output , boundary branch output and pre-constructed conflict graph are input into the conflict regulation fusion module together; S2-8: In the conflict regulation fusion module, based on the input conflict graph and pixel-level conflict saliency map , a spatial adaptive fusion weight map is calculated by a weight generation network composed of two layers of 3x3 convolution layers ; wherein the first layer adopts ReLU activation function after convolution, and the second layer adopts Sigmoid activation function after convolution, and the calculation process is represented as: , and represents a channel splicing operation; S2-9: Utilize the generated weight map The outputs of the three branches are adaptively weighted and fused to generate the final global probability prediction map The specific fusion formula is: wherein represents pixel-by-pixel multiplication, and are preset hyperparameters, respectively taking values of 0.6 and 0.

4.

3. The method of claim 2, wherein the method is based on a multi-branch cooperative network. , the step S2-9 in and are respectively valued at 0.6 and 0.

4.

4. The method of claim 1, wherein ,The step S3 specifically comprises the following steps: S3-1: When the global probability prediction graph The pixel value is greater than or equal to the threshold. If the condition is met, it is identified as a nodule region; otherwise, it is identified as a background region. The initial threshold is set to 0.5, and the final threshold is determined by grid search on the validation set to maximize the Dice similarity coefficient. S3-2: based on thresholding, improve the stability and clinical usability of the results through post-processing steps.

5. The method of claim 4, wherein the method is based on a multi-branch cooperative network. ,The step S3-2 specifically comprises the following steps: S3-2-1: use three-dimensional connected component analysis to remove small false positive regions with a number of voxels less than a predetermined threshold; S3-2-2: smooth the segmentation edges by morphological closing operation.

6. The method of claim 1, wherein , the step S4 , , , , .

Images