Two-stage pulmonary nodule detection method and system based on cross-layer attention fusion

Abstract

The invention relates to a two-stage pulmonary nodule detection method and system based on cross-layer attention fusion, and belongs to the technical field of medical image processing, and the method comprises the steps: carrying out the preprocessing of a three-dimensional chest CT image, segmenting a pulmonary parenchyma region, and carrying out the standardization of the pulmonary parenchyma region, and obtaining a to-be-detected image; inputting the to-be-detected image into the candidate nodule detection network, and generating a position frame of a candidate nodule; extracting a three-dimensional image region corresponding to the candidate nodule position frame, inputting the three-dimensional image region into a false positive suppression network for classification, and filtering out false positive candidates; and outputting a final pulmonary nodule detection result confirmed by the false positive suppression network. Through the design of dual-stage task decoupling, an innovative space and channel residual module and a multi-scale progressive sensing network module and a self-adaptive training strategy, the core problems of sensitivity and specificity imbalance, weak feature discrimination ability, poor multi-scale adaptability and unstable training in intelligent detection of pulmonary nodules are systematically solved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a two-stage lung nodule detection method and system based on cross-layer attention fusion. Background Technology

[0002] Lung cancer is one of the leading causes of cancer-related morbidity and mortality worldwide, and early detection and intervention are crucial for improving patient prognosis. Lung nodules, as an early imaging manifestation of lung cancer, are primarily screened through large-scale low-dose chest computed tomography (CT) scans. However, the traditional manual interpretation of images by radiologists has inherent limitations: First, each patient's CT images contain hundreds of cross-sections, making manual analysis labor-intensive, inefficient, and prone to causing physician fatigue; second, lung nodules exhibit significant heterogeneity in size, shape, density, and anatomical location, making it easy to miss small or atypical nodules; finally, diagnostic results are highly dependent on the physician's personal experience, resulting in a high subjective false-positive rate.

[0003] To address these challenges, computer-aided detection systems have emerged. Most existing intelligent lung nodule detection methods are based on 3D CT images for identification; however, these methods often struggle to fully exploit and utilize the contextual information in 3D space during feature extraction and discrimination. Existing methods frequently employ ordinary 3D convolution operations in the candidate nodule detection stage, which has limited ability to model multi-scale feature fusion and long-range spatial dependencies. This results in insufficient sensitivity to small nodules, apical nodules, and irregularly shaped nodules, leading to a high false negative rate. Furthermore, in the false positive suppression stage, most methods rely on single-scale or simply aggregated features for discrimination, failing to fully consider the large differences in nodule size and morphology, resulting in limited ability to distinguish between real nodules and artifacts, and an overall high false positive rate. Therefore, existing methods struggle to achieve a balance between high sensitivity and low false positive rate in clinical applications, limiting their reliability and applicability in actual lung nodule screening. Summary of the Invention

[0004] To address the aforementioned problems, the purpose of this invention is to provide a two-stage lung nodule detection method and system based on cross-layer attention fusion.

[0005] A two-stage lung nodule detection method based on cross-layer attention fusion includes:

[0006] S110: Acquire 3D chest CT images;

[0007] S120: The three-dimensional chest CT image is preprocessed to segment the lung parenchyma region, and the lung parenchyma region is standardized to obtain the image to be detected;

[0008] S130: Input the image to be detected into the candidate nodule detection network to generate the location bounding boxes of candidate nodules; the candidate nodule detection network is an encoder-decoder architecture, with attention-guided spatial and channel residual modules as feature extraction units;

[0009] S140: Extract the 3D image region corresponding to the candidate nodule location box, input it into the false positive suppression network for classification, and filter out false positive candidates; the false positive suppression network is a multi-scale progressive perception network, which is used to process candidate nodule images of different scales in parallel and perform cross-scale feature fusion.

[0010] S150: Outputs the final lung nodule detection results confirmed by the false positive suppression network.

[0011] Preferably, in S120, the three-dimensional chest CT image is preprocessed to segment the lung parenchyma region, including:

[0012] Gray-scale thresholding is performed on 3D chest CT images to extract voxels within a preset threshold range and generate an initial lung region mask.

[0013] The initial lung region mask was sequentially expanded and eroded to obtain a morphologically processed mask.

[0014] Three-dimensional connected component analysis was performed on the morphologically processed mask to retain the two connected regions with the largest number of voxels, which were used as the left and right lung lobe regions, respectively.

[0015] The lung parenchyma was obtained by filling the internal cavities in the left and right lung lobes.

[0016] Preferably, in S130, an attention-guided spatial and channel residual module is embedded at a preset level of the encoder; and the final output feature of the attention-guided spatial and channel residual module is represented as follows:

[0017]

[0018]

[0019]

[0020]

[0021] in, The final output feature representation of the attention-guided spatial and channel residual module; residual operator It consists of a 3×3×3 convolutional layer and the PReLU activation function. For feature maps, they can usually be denoted as: Where C is the number of channels, and D, H, and W are the depth, height, and width of the three-dimensional volume data, respectively; For 3D convolution, For three-dimensional average pooling, For three-dimensional max pooling, To modify the activation function of the linear unit, This represents the Sigmoid activation function. , These are learnable attention weight parameters.

[0022] Preferably, in S130, the training process of the candidate nodule detection network is as follows:

[0023] Obtain a training dataset, which contains multiple lung CT scans and corresponding lung nodule annotation information;

[0024] The candidate nodule detection network is trained using the training dataset; wherein the loss function during training is:

[0025]

[0026]

[0027]

[0028] in, and These represent the number of positive and negative samples, respectively. α is the predicted probability of sample i, representing the network's prediction that the sample belongs to a lung nodule. α is the weighting factor in the focus loss, γ = 2 is the adjustment factor for the focus loss, and λ is the dynamic coefficient used to balance the weights between classification loss and regression loss. and Let represent the predicted and true values ​​of the regression parameters for the k-th bounding box, respectively, and represent the location and size of the nodule in the CT image. Focus loss function. By paying more attention to difficult-to-classify samples (such as small nodules), the model's sensitivity to nodules is improved. Smoothing L1 loss is also employed. Used in regression tasks to precisely adjust the position of nodule bounding boxes, thereby improving the accuracy of nodule localization. Total loss function. By combining the losses from classification and regression tasks, the network achieves good performance in both detection and localization.

[0029] Preferably, in S140, the three-dimensional image region corresponding to the prediction result of the trained candidate nodule detection network on the training set data is used as the basis and labeled as a true positive sample or a false positive sample; the false positive sample and the false positive sample are used to train the false positive suppression network; wherein, the cross-entropy loss function is used and the network parameters are optimized by the Adam optimizer during the training process.

[0030] This invention also provides a two-stage lung nodule detection system based on cross-layer attention fusion, comprising:

[0031] The data acquisition module is used to acquire three-dimensional chest CT images;

[0032] The standardization module is used to preprocess the three-dimensional chest CT image to segment the lung parenchyma region, and standardize the lung parenchyma region to obtain the image to be detected.

[0033] The detection module is used to input the image to be detected into the candidate nodule detection network to generate the location bounding boxes of the candidate nodules; the candidate nodule detection network is an encoder-decoder architecture, with attention-guided spatial and channel residual modules as feature extraction units;

[0034] The classification module is used to extract the three-dimensional image region corresponding to the candidate nodule location box, input it into the false positive suppression network for classification, and filter out false positive candidates; the false positive suppression network is a multi-scale progressive sensing network, which is used to process candidate nodule images of different scales in parallel and perform cross-scale feature fusion.

[0035] The test result output module is used to output the final lung nodule detection results confirmed by the false positive suppression network.

[0036] The present invention also provides an electronic device, including a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected via the bus, characterized in that the computer program, when executed by the processor, implements the steps in the above-described two-stage lung nodule detection method based on cross-layer attention fusion.

[0037] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps in the above-described two-stage lung nodule detection method based on cross-layer attention fusion.

[0038] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0039] This invention relates to a two-stage lung nodule detection method based on cross-layer attention fusion. Compared with existing technologies, this invention achieves decoupling and synergistic optimization of performance through a two-stage architecture of "high-sensitivity coarse screening + high-specificity fine judgment". The first stage focuses on maximizing nodule recall, significantly improving the network's feature capture ability for difficult nodules by introducing attention-guided spatial and channel residual modules. The second stage focuses on efficiently identifying true and false nodules by adaptively fusing multi-scale information through a multi-scale progressive perception network, greatly enhancing the model's ability to distinguish between true and false nodules. In addition, a dynamic weighted training strategy ensures stable model convergence and achieves optimal performance balance.

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is an overall flowchart of the intelligent lung nodule detection method provided in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the candidate nodule detection network in an embodiment of the present invention;

[0044] Figure 3 Detailed structural breakdown diagram of the spatial and channel residual module for attention guidance;

[0045] Figure 4 This is a schematic diagram of the false positive suppression network in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the data preprocessing process in an embodiment of the present invention;

[0047] Figure 6 Grad-CAM heatmap showing the detection effect of the intelligent lung nodule detection system provided in this embodiment of the invention. Detailed Implementation

[0048] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0049] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0050] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0051] Please see Figure 1-6 A two-stage lung nodule detection method based on cross-layer attention fusion includes:

[0052] S110: Acquire 3D chest CT images;

[0053] The specific steps for acquiring and processing three-dimensional chest CT images include:

[0054] S111: Acquire CT image sequence data containing pulmonary nodules.

[0055] S112, The CT image data is processed according to preset rules to extract three-dimensional image data containing complete lung tissue.

[0056] In this embodiment, thin-slice chest CT DICOM data of the patient is acquired through a medical image archiving and communication system. Preset rules typically refer to automatically loading the entire scan sequence while preserving the original spatial information. The core processing object of this method is three-dimensional volumetric data containing both lungs.

[0057] S120: The three-dimensional chest CT image is preprocessed to segment the lung parenchyma region, and the lung parenchyma region is standardized to obtain the image to be detected;

[0058] S130: The image to be detected is input into the candidate nodule detection network to generate the location bounding boxes of the candidate nodules; the candidate nodule detection network is an encoder-decoder architecture, with attention-guided spatial and channel residual modules as feature extraction units to enhance feature discrimination capabilities;

[0059] The training process of the candidate nodule detection network is as follows:

[0060] S210, Obtain the initial training dataset, which contains multiple lung CT scans and corresponding lung nodule annotation information.

[0061] S220 processes all training images based on a unified preprocessing strategy to obtain a standardized training dataset.

[0062] S230, For the standardized dataset, construct data samples for two-stage network training by combining the annotation information.

[0063] S240 divides the constructed dataset into a training set and a validation set.

[0064] S250, the two-stage network is trained or sequentially trained using the training set, and then validated and optimized using the validation set to obtain the final model that meets the performance requirements.

[0065] With the network model already built, the model is trained and optimized using a large amount of labeled data to obtain a high-performance detection model.

[0066] In step S210, an initial training dataset containing multiple CT scans and precise nodule annotations is obtained. To ensure the model's generalization ability, the dataset should cover lung nodules of different sizes, morphologies (solid, ground-glass opacities, partially solid), and locations (central, subpleural). In a preferred implementation of this embodiment, the publicly available LUNA16 dataset is used. This dataset, derived from LIDC-IDRI, contains 888 CT scans and 1186 nodule annotations ≥3mm, and provides expert-consensus nodule coordinates and segmentation masks.

[0067] Step S220 involves preprocessing each training sample to improve data quality and consistency. For example... Figure 5 As shown, this embodiment employs a comprehensive preprocessing strategy, which specifically includes the following steps:

[0068] First, lung parenchyma segmentation is performed. Based on the low-density distribution of lung tissue in CT images, grayscale thresholding is applied to the original CT image to extract voxels within a preset threshold range, generating an initial lung region mask. Then, a three-dimensional morphological closing operation is performed on the initial lung region mask, including dilation and erosion operations, to eliminate noise areas in the segmentation result and fill in small fractures within the lung parenchyma. Next, three-dimensional connected component analysis is performed on the morphologically processed mask. Based on the number of voxels contained in each connected region, the two connected regions with the largest number of voxels are retained as the left and right lung lobe regions, respectively, while the remaining connected regions are discarded. Finally, the internal cavities of the left and right lung lobe regions are filled to obtain a complete and continuous lung parenchyma mask.

[0069] Subsequently, based on the lung parenchyma mask, voxel data located within the lung parenchyma region are extracted from the original CT image and used as input data for subsequent candidate nodule detection and false positive suppression network training.

[0070] Next, data standardization will be performed:

[0071] First, the gray values ​​of CT voxels located in the lung parenchyma area are restricted by range. Gray values ​​less than the preset lower limit (-1200HU) are uniformly set as the lower limit value, and gray values ​​greater than the preset upper limit (600HU) are uniformly set as the upper limit value, thereby constraining all voxel gray values ​​within the gray range corresponding to the typical window width and window level of the lung.

[0072] Spatial normalization is then performed: using cubic spline interpolation, the image is resampled to an isotropic voxel spacing (1mm×1mm×1mm) to ensure that CT images obtained under different scanning devices and scanning parameters maintain consistent spatial resolution, thereby ensuring the consistency of spatial scale in subsequent candidate nodule detection and false positive suppression network training.

[0073] These operations effectively standardize input data specifications and eliminate interference from irrelevant organizations, providing high-quality input for the model.

[0074] In step S230, samples are constructed according to the requirements of the two-stage networks. For the candidate detection network, the entire preprocessed lung parenchyma region or a fixed-size 3D image patch cropped from it is used as input, and corresponding nodule centers and bounding box labels are generated. For the false positive suppression network, multi-scale image patches are cropped as input based on the candidate boxes generated by the first-stage network, and are labeled as positive or negative samples according to their overlap with the true annotations.

[0075] Step S240 divides the processed dataset into training and validation sets. A 6-fold cross-validation strategy is adopted, in which the data is randomly divided into 6 parts, and 5 parts are used for training and 1 part for validation in turn, so as to fully evaluate the model performance and stability and prevent overfitting.

[0076] The candidate nodule detection network is trained using an end-to-end multi-task learning approach, which includes:

[0077] S251, the preprocessed 3D training image samples are input into the network. The network's forward propagation sequentially performs multi-scale feature extraction and enhancement through the encoder and its integrated ASCRM modules, and then performs feature fusion and spatial information recovery through the decoder and its integrated ASCRM modules, finally outputting the predicted nodule center and bounding box.

[0078] S252 calculates the multi-task loss function between network predictions and ground truth labels. The proposed loss function consists of two parts: classification loss and regression loss, optimized for multiple tasks through a dynamic weighting mechanism. For the classification task, an improved binary focal loss function is used, with its core formula defined as:

[0079]

[0080] in and These represent the number of positive and negative samples, respectively. Indicates the predicted probability of a sample (for positive samples). Negative samples are The hyperparameter α = 0.6 is used to adjust the weights of positive and negative samples, and γ = 2 alleviates the class imbalance problem by suppressing the loss weights of easily classified samples. The regression task uses Smooth L1 Loss, which is mathematically expressed as:

[0081]

[0082] The total loss function, through adaptive weight balancing of the two tasks, has the following complete form:

[0083]

[0084] The dynamic coefficient λ adaptively adjusts the contribution of the classification loss based on the sample distribution, preventing the regression task from being dominated by sample sparsity. This design improves classification accuracy while ensuring the stability of localization regression.

[0085] The key innovation is the introduction of dynamic weighting coefficients, which are automatically calculated and adjusted based on the ratio of positive to negative samples in the current training batch, in order to achieve adaptive balance optimization between classification and regression tasks.

[0086] S253, based on the aforementioned loss, uses the backpropagation algorithm to calculate the gradient and employs the Adam optimizer to update all parameters of the entire candidate nodule detection network at once, achieving end-to-end joint training. During training, an online hard example mining strategy can be incorporated to improve the model's ability to identify difficult samples.

[0087] S140: Extract the 3D image region corresponding to the candidate nodule location box, input it into the false positive suppression network for classification, and filter out false positive candidates; the false positive suppression network is a multi-scale progressive perception network, which is used to process candidate nodules of different scales in parallel and fuse cross-scale features;

[0088] For training the false positive suppression network, a supervised classification method based on multi-scale input is adopted, and the specific process is as follows:

[0089] S254, Construct a training sample set. Based on the prediction results of the pre-trained candidate nodule detection network on the training set data, collect the generated candidate boxes and label them as true positive samples or false positive samples according to their overlap with the ground truth annotations. To optimize the learning process, the ratio of positive to negative samples will be manually balanced.

[0090] S255 performs multi-scale training. For each candidate sample, multiple 3D image patches with different spatial resolutions are generated as input to the false positive suppression network.

[0091] S256, Optimize network parameters. The network uses the standard cross-entropy loss function and is trained with the Adam optimizer to learn to distinguish between true nodules and various false positive structures. Furthermore, data augmentation techniques such as random rotation and flipping can be used in the training of both networks to further improve the model's generalization ability and robustness.

[0092] S150: Outputs the final lung nodule detection results confirmed by the false positive suppression network.

[0093] It should be noted that before training begins, there is also the step of building the network model:

[0094] S201 establishes the initial architecture of the candidate nodule detection network, including an encoder, an attention guidance module (ASCRM), and a decoder.

[0095] S202 constructs the encoder's basic feature extraction path using 3D convolution and downsampling layers.

[0096] S203, embed the attention-guided spatial and channel residual module (ASCRM) at the key level of the encoder.

[0097] S204, a false positive suppression network was constructed as a multi-scale progressive sensing network, including a parallel multi-branch structure and cross-scale feature interaction pathways.

[0098] The construction and operation of the false positive suppression network includes the following steps:

[0099] S310: Crop the candidate nodule block into three-dimensional image blocks of large, medium and small sizes;

[0100] S320: Image patches of different resolutions are input into multiple parallel feature extraction branches, each branch containing a series of three-dimensional convolutional layers to extract hierarchical features from image patches of the corresponding scale.

[0101] S330: Through the cross-scale feature interaction pathway, the high-level semantic features extracted from the branch processing larger-scale image patches are transferred and fused into the branch processing smaller-scale image patches to achieve top-down semantic guidance.

[0102] S340: Through feature aggregation and classification layers, the high-level semantic features output from all branches are aggregated, and the probability of the candidate nodule being a true nodule is calculated via a fully connected layer.

[0103] The method for candidate nodule detection networks includes the following steps:

[0104] S410: Construct a multi-task learning framework whose overall loss function consists of two parts: a classification sub-loss and a regression sub-loss;

[0105] S420: The sub-loss of the classifier adopts the focal loss function to alleviate the extreme imbalance between foreground (nodules) and background voxels in the training data;

[0106] S430: The contributions of the classifier loss and the regressor loss to the total loss are balanced by a dynamically adjusted weight coefficient; the dynamic weight coefficient is calculated in real time based on the ratio of positive samples (nodules) to negative samples (background) in the current training batch, enabling the network to adaptively adjust its learning focus on the classification task and the regression task.

[0107] Model building needs to be completed before actual training. Traditional 3D detection networks have shortcomings in feature fusion and long-range dependency modeling. This embodiment systematically improves performance by introducing the ASCRM module and MPPN network.

[0108] The attention-guided spatial and channel residual module is one of the core components of this invention. The overall structure of ASCRM is as follows: Figure 3As shown, based on the traditional residual structure, this module integrates dense skip connections and a convolutional block attention module (CBAM) to enhance multi-scale semantic representation capabilities and improve the network's attention to key spatial regions.

[0109] Specifically, ASCRM constructs deep residual paths by stacking multiple 3×3×3 convolutional layers, while introducing long-range cross-layer connections to progressively transfer high-resolution structural features from early shallow layers to deeper network layers. This design effectively preserves the fine-grained texture information and boundary details of the nodules. The update form of each layer's features in the residual path is as follows:

[0110]

[0111] in, This represents the residual unit corresponding to the k-th 3×3×3 convolutional block. This is the output feature map for the current stage. To further facilitate information interaction between different feature levels, the ASCRM module introduces a cross-layer residual fusion mechanism. This mechanism integrates feature representations from different depths of the network. and To integrate, among which Typically, higher resolution texture details are preserved, while This contains richer high-level semantic information. To achieve more effective feature integration, a nonlinear residual refinement operation is applied to the summed features to obtain an enhanced aggregated feature representation:

[0112]

[0113] Among them, residual operators It consists of a 3×3×3 convolutional layer and a PReLU activation function. This design not only enhances the expressive power of the fused features but also ensures the stable propagation of information across layers. By mitigating the representation degradation problem that may occur during multi-scale feature fusion, this residual fusion structure effectively improves the joint modeling capability of fine-grained structural cues and high-level semantic information, providing a more discriminative feature foundation for the subsequent attention module.

[0114] Building upon this foundation, to further enhance feature discriminative power, the ASCRM module introduces CBAM, constructing a dual attention mechanism that operates simultaneously on both the spatial and channel dimensions. The channel attention branch captures global response features through average pooling and max pooling operations, generating channel attention weights; the spatial attention branch aggregates multi-scale pooled features to generate a spatial response map that highlights key regions. Its computational form is as follows:

[0115]

[0116]

[0117] in, This represents the Sigmoid activation function. This is an element-wise multiplication operation. , These are learnable attention weight parameters.

[0118] The final output features of the ASCRM module are represented as follows:

[0119]

[0120] Through the above structural design, the ASCRM module achieves efficient fusion of shallow and deep features, while strengthening the representation of key information in both spatial and channel dimensions. When this module is used in multiple stacked layers in the network, it can improve the accuracy and robustness of lung nodule candidate detection, especially when facing nodule lesions with complex morphology, it exhibits stronger detection capabilities.

[0121] The multi-scale progressive sensing network is another core component of this invention. For example... Figure 4 As shown, MPPN classifies candidate nodules into three scale categories based on nodule diameter and extracts three-dimensional voxel blocks of sizes 32×48×48, 16×24×24, and 8×12×12 respectively to construct an adaptive multi-scale input representation. This design avoids the spatial distortion and semantic information degradation problems commonly encountered in unified resampling, effectively improving the network's ability to model the geometric structure and contextual information of nodules at different scales. For inputs of different scales, the network sets independent three-dimensional convolutional branches, forming a multi-scale feature extraction path from coarse to fine.

[0122] To alleviate the semantic isolation problem between features at different scales, a top-down feature injection mechanism is introduced in the mesoscale and small-scale branches, enabling high-level semantic information from the large-scale branch to guide the identification process of small nodules. Furthermore, each branch integrates a Convolutional Block Attention (CBAM) module to enhance the model's ability to focus on discriminative regions. High-level semantic features from each scale branch are concatenated along the channel dimension and obtained through global average pooling to achieve a compact global feature representation. This feature representation is then recalibrated by attention and input to a fully connected layer, ultimately outputting the probability that each candidate region is a true nodule.

[0123] Therefore, combining Figure 2 and Figure 4The network structure shown, through the integration of an ASCRM module detection network and an MPPN network with cross-scale interaction, constructs a complete model system capable of efficiently and accurately completing the lung nodule detection task.

[0124] This invention provides a two-stage intelligent lung nodule detection method based on cross-layer attention fusion and multi-scale progressive perception. This method, through a collaborative architecture of "high-sensitivity candidate detection" and "high-specificity false-positive suppression," combined with an innovative attention module and multi-scale network design, achieves automatic and accurate identification of lung nodules from three-dimensional chest CT images. While maintaining a high nodule detection rate, this method significantly reduces false alarms, effectively solving the bottleneck problem of existing single-network models struggling to balance sensitivity and specificity. It provides an efficient and reliable computer-aided diagnostic tool for early lung cancer screening.

[0125] Specifically, the beneficial effects of the present invention are reflected in the following aspects:

[0126] 1. Superior Overall Detection Performance: The proposed two-stage framework achieved leading test results on the authoritative public dataset LUNA16. In the first stage, the candidate nodule detection network achieved a high sensitivity of 98.5%. Under the complete two-stage workflow, the system achieved a clinically high sensitivity of 94.9%, with an average of only 0.55 false positives per scan, significantly outperforming most existing methods. This is attributed to the decoupled design of the two-stage tasks, allowing the model to optimize both the "complete detection" and "accurate diagnosis" objectives separately, ultimately resulting in a significant improvement in overall performance (CPM score).

[0127] 2. Enhanced Feature Learning and Discrimination Capabilities: By introducing attention-guided spatial and channel residual modules, the network can adaptively focus on key feature channels and spatial regions related to nodules. The cross-layer residual fusion mechanism of this module promotes effective interaction between deep and shallow layer features. Combined with dual attention calibration, it greatly enhances the characterization ability and detection sensitivity of difficult nodules such as small nodules, ground-glass opacities, and nodules adhered to the pleura / blood vessels, providing a foundation for high recall.

[0128] 3. Excellent robustness and multi-scale adaptability: To address the challenge of large differences in lung nodule size, this invention designs a multi-scale progressive sensing network for false positive suppression. This network processes candidate regions of different scales through parallel branches and utilizes an innovative top-down feature injection pathway to obtain semantic context guidance from large-scale analysis when identifying small nodules. This design significantly improves the model's discrimination accuracy for nodules of different sizes, thereby effectively suppressing false positives caused by structures such as vascular sections and bronchi in complex backgrounds.

[0129] 4. Stable and efficient model training mechanism: To address the extreme class imbalance problem in the training data, the candidate detection network employs a dynamically weighted multi-task loss function. The weights in this loss function can be adjusted in real time according to the ratio of positive and negative samples in the training batch. Combined with the focus loss function, this enables the model to adaptively balance the learning priorities of classification and regression tasks, and focus on difficult samples, ensuring the stability of the training process and the final generalization performance of the model.

[0130] 5. Significant Clinical Practicality and Translational Value: This invention achieves fully automated analysis, rapidly processing entire CT scans and outputting intuitive structured reports and 3D localization results. This significantly reduces the workload of radiologists, improves screening efficiency and diagnostic consistency, and reduces missed diagnoses and misdiagnoses caused by fatigue and experience differences. Its high accuracy and low false positive rate make it highly suitable for large-scale lung cancer screening scenarios, possessing the potential to be directly translated into clinical auxiliary diagnostic products, contributing to the early diagnosis and treatment of lung cancer.

[0131] Compared with existing methods for detecting lung nodules, the intelligent detection framework based on two-stage attention and multi-scale perception proposed in this invention demonstrates significant advantages in architecture design, feature learning, multi-scale processing, and training strategies. A detailed comparison is as follows:

[0132] 1. Compared to traditional single-stage detection methods:

[0133] Traditional single-stage detectors typically perform candidate nomination and classification simultaneously within a single network, making it difficult to balance extremely high sensitivity with extremely low false positive rates. Increasing recall often leads to a surge in false positives, while rigorous filtering of false positives can result in missed atypical nodules. This invention employs a clearly defined two-stage cascaded architecture, decoupling "high-sensitivity screening" and "high-specificity identification" into two specially optimized sub-networks. The first stage prioritizes high recall, while the second stage focuses on precise discrimination, thus achieving optimal synergy between the two at the system level, ensuring a high detection rate while keeping false positives at an extremely low level.

[0134] 2. Compared to traditional 3D detection networks and their variants:

[0135] Traditional 3D CNN detection networks primarily rely on stacked convolutions for feature extraction, which has limitations in modeling dependencies between feature channels and long-range spatial context, and is insufficient in discriminative power for small, low-contrast nodules. This invention innovatively integrates attention-guided spatial and channel residual modules into the detection network. This module not only preserves multi-scale details through cross-layer residual fusion, but also enables the network to adaptively calibrate and focus on key features through parallel channel and spatial attention mechanisms, significantly enhancing the network's feature representation ability and detection sensitivity for complex morphologies and difficult-to-detect nodules.

[0136] 3. Compared to other state-of-the-art models that employ attention or cascade strategies:

[0137] While some existing methods incorporate attention mechanisms or employ a two-stage approach, they often rely on simple image scaling or single-path analysis when dealing with extreme differences in nodule size, failing to fully leverage the semantic connections between multi-scale contexts. This invention's false positive suppression stage utilizes a multi-scale progressive perceptual network, its core innovation being the construction of a top-down cross-scale feature injection pathway. This design allows the branch processing small nodules to directly receive high-level semantic guidance from large-scale analysis, achieving true multi-scale semantic fusion rather than simple feature splicing. This significantly improves the identification accuracy for nodules of different sizes, especially small nodules, effectively suppressing multiple types of false positives.

[0138] 4. Regarding the overall architecture and training strategy:

[0139] Most existing methods use fixed loss weights or simple combinations of loss functions to train multi-task detection networks. When faced with extreme foreground-background imbalance in CT data, the optimization process is prone to instability, and the model tends to favor easier samples. This invention designs a dynamically weighted multi-task loss function for the candidate detection network, where the weights balancing classification and regression losses can be adjusted in real-time and adaptively based on the sample distribution of the current batch of data. This innovative training strategy ensures that the network consistently focuses on key tasks and difficult samples during training, thereby learning a more robust detection model and improving the generalization performance and stability of the method.

[0140] In summary, this invention systematically solves the core challenges of intelligent lung nodule detection, including the imbalance between sensitivity and specificity, weak feature discrimination, poor multi-scale adaptability, and unstable training, through two-stage task decoupling, innovative ASCRM and MPPN module design, and adaptive training strategies. Compared to existing technologies, this invention achieves superior overall performance in publicly available benchmark tests, demonstrating greater clinical applicability and translational value.

[0141] The present invention also provides an electronic device, including a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor. The transceiver, the memory, and the processor are connected via the bus. The computer program, when executed by the processor, implements the steps of the aforementioned two-stage lung nodule detection method based on cross-layer attention fusion. Compared with the prior art, the beneficial effects of the electronic device provided by the present invention are the same as those of the aforementioned two-stage lung nodule detection method based on cross-layer attention fusion, and will not be elaborated upon here.

[0142] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps in the above-described two-stage lung nodule detection method based on cross-layer attention fusion. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present invention are the same as the beneficial effects of the two-stage lung nodule detection method based on cross-layer attention fusion described in the above-described technical solution, and will not be repeated here.

[0143] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A two-stage lung nodule detection method based on cross-layer attention fusion, characterized in that, include: S110: Acquire 3D chest CT images; S120: The three-dimensional chest CT image is preprocessed to segment the lung parenchyma region, and the lung parenchyma region is standardized to obtain the image to be detected; S130: Input the image to be detected into the candidate nodule detection network to generate the location bounding boxes of candidate nodules; the candidate nodule detection network is an encoder-decoder architecture, with attention-guided spatial and channel residual modules as feature extraction units; S140: Extract the 3D image region corresponding to the candidate nodule location box, input it into the false positive suppression network for classification, and filter out false positive candidates; the false positive suppression network is a multi-scale progressive sensing network, which is used to process candidate nodule images of different scales in parallel and perform cross-scale feature fusion. S150: Outputs the final lung nodule detection results confirmed by the false positive suppression network.

2. The two-stage lung nodule detection method based on cross-layer attention fusion according to claim 1, characterized in that, In S120, the three-dimensional chest CT image is preprocessed to segment the lung parenchyma region, including: Gray-scale thresholding is performed on 3D chest CT images to extract voxels within a preset threshold range and generate an initial lung region mask. The initial lung region mask was sequentially expanded and eroded to obtain a morphologically processed mask. Three-dimensional connected component analysis was performed on the morphologically processed mask to retain the two connected regions with the largest number of voxels, which were used as the left and right lung lobe regions, respectively. The lung parenchyma was obtained by filling the internal cavities in the left and right lung lobes.

3. The two-stage lung nodule detection method based on cross-layer attention fusion according to claim 2, characterized in that, In step S130, attention-guided spatial and channel residual modules are embedded at a preset level of the encoder; and the final output features of the attention-guided spatial and channel residual modules are represented as follows: in, The final output feature representation of the attention-guided spatial and channel residual module; residual operator It consists of a 3×3×3 convolutional layer and the PReLU activation function. For feature maps, they can usually be denoted as: Where C is the number of channels, and D, H, and W are the depth, height, and width of the three-dimensional volume data, respectively; For 3D convolution, For three-dimensional average pooling, For max pooling, To modify the activation function of the linear unit, This represents the Sigmoid activation function. , These are learnable attention weight parameters.

4. The two-stage lung nodule detection method based on cross-layer attention fusion according to claim 3, characterized in that, In S130, the training process of the candidate nodule detection network is as follows: Obtain a training dataset, which contains multiple lung CT scans and corresponding lung nodule annotation information; The candidate nodule detection network is trained using the training dataset; wherein the loss function during training is: in, and These represent the number of positive and negative samples, respectively. Let α represent the predicted probability of the sample, γ = 2, and λ be the dynamic coefficient. Let be the predicted value of the regression parameter for the k-th bounding box. For the true value of the regression parameter of the k-th bounding box, For the total loss function, For a binary focus loss function, This is the smoothed L1 loss function.

5. The two-stage lung nodule detection method based on cross-layer attention fusion according to claim 4, characterized in that, In S140, the three-dimensional image region corresponding to the prediction result of the trained candidate nodule detection network on the training set data is used as the basis to mark it as a true positive sample or a false positive sample. The false positive suppression network was trained using both true positive and false positive samples; during training, the cross-entropy loss function was used and the network parameters were optimized using the Adam optimizer.

6. A two-stage lung nodule detection system based on cross-layer attention fusion, characterized in that, include: The data acquisition module is used to acquire three-dimensional chest CT images; The standardization module is used to preprocess the three-dimensional chest CT image to segment the lung parenchyma region, and standardize the lung parenchyma region to obtain the image to be detected. The detection module is used to input the image to be detected into the candidate nodule detection network and generate the location bounding boxes of the candidate nodules; The candidate nodule detection network is an encoder-decoder architecture, with attention-guided spatial and channel residual modules as feature extraction units. The classification module is used to extract the three-dimensional image region corresponding to the candidate nodule location box, input it into the false positive suppression network for classification, and filter out false positive candidates; the false positive suppression network is a multi-scale progressive sensing network, which is used to process candidate nodule images of different scales in parallel and perform cross-scale feature fusion. The test result output module is used to output the final lung nodule detection results confirmed by the false positive suppression network.

7. An electronic device comprising a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected via the bus, characterized in that, When the computer program is executed by the processor, it implements the steps in the two-stage lung nodule detection method based on cross-layer attention fusion as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps in the two-stage lung nodule detection method based on cross-layer attention fusion as described in any one of claims 1-5.

Images