A lung nodule detection method and system based on adaptive multi-scale deformable attention

Abstract

The present application relates to the technical field of lung nodule detection, in particular to a lung nodule detection method and system based on adaptive multi-scale deformable attention. The method comprises the following steps: acquiring a clinical lung nodule dataset; performing data preprocessing on the acquired dataset; constructing a deep network model based on AMDA-YOLO; training the deep network model based on AMDA-YOLO using a linear warm-up strategy; performing lung nodule detection using the trained model; and outputting the detection results. The detection system constructed by the present application has excellent cross-data domain generalization capability and clinical practical value.

Claims

1. A method for detecting lung nodules based on adaptive multi-scale deformable attention, characterized in that, include: Acquire a clinical lung nodule dataset, which includes multi-source clinical lung nodule data, including CT body data, nodule annotation information, and associated examination text information; Data preprocessing is performed on the acquired multi-source text and image datasets; Construct a deep network model based on AMDA-YOLO; A linear preheating strategy was used to train a deep network model based on AMDA-YOLO; Detecting lung nodules using a trained model; Output the detection results; The construction of the deep network model based on AMDA-YOLO includes first building an improved end-to-end lung nodule detection model, AMDA-YOLO, based on the YOLOv12 framework. This model addresses the problem of missed detection of small nodules caused by receptive field fixation and feature information loss during downsampling in traditional convolutional neural networks. It consists of three parts: a deep data representation backbone, an adaptive multi-scale fusion neck section, and a high-precision prediction head. First, the preprocessed two-dimensional slice tensor is input. This study utilizes a variant of the Backbone architecture, alternating between stride convolutions and cross-stage local network (CSP) structures, to generate multi-level representations. The backbone network integrates C3k2 and A2C2f modules. The A2C2f module achieves decoupling and reconstruction of complex semantics through a feature gradient splitting mechanism, targeting input features... The operation process is expressed as follows: ,in, and For the splitting linear mapping operator, The region attention functional is used to recalibrate the spatial weights of nodule candidate regions early in feature propagation. The backbone ultimately outputs feature pyramids with different spatial resolutions. , respectively corresponding to the original input , and Downsampling rate; subsequently, the Neck network adopts a high-performance fusion strategy based on bidirectional path aggregation, and innovatively uses the adaptive multi-scale deformable attention (AMDA) module as the core of the fusion operator. For adjacent scale features of the backbone output, Neck abandons the traditional weighted addition mode and adopts a lossless criterion based on channel-dimensional splicing. Scale feature map Its fusion equation is defined as: ,in For bilinear interpolation upsampling operators, the concatenated high-dimensional feature flow Immediately feed the data into the AMDA module for dynamic deformation reconstruction to capture the boundary details of irregularly shaped nodules, and finally... Convolution Operator Complete channel dimensionality reduction and information integration, namely: Finally, the detection head outputs the target category probability distribution, normalized bounding box coordinate parameters, and target confidence score through the multi-scale prediction branch. It then uses the non-maximum suppression (NMS) operator to perform extreme value filtering on redundant candidate boxes. Let the set of candidate boxes output by the multi-scale prediction branch be... ,in Indicates the first One candidate box, Indicates its target confidence score; first according to Sort by score from highest to lowest, and select the candidate boxes with the highest scores in order. Add to result set For any remaining candidate box ,calculate: ， when At that time, Determined to be with Corresponding to the same nodule target and from Delete it; otherwise, keep it and proceed to the next round of filtering, repeating until... Empty The candidate boxes retained are the final detection results after redundancy removal.

2. The lung nodule detection method based on adaptive multi-scale deformable attention according to claim 1, characterized in that, The data preprocessing of the acquired multi-source text and image datasets includes, to ensure consistency between different CT scan data and eliminate background noise introduced by differences in equipment physical parameters, first processing the tensor of the original three-dimensional CT volume data. Perform HU truncation mapping based on human anatomy priors, defining spatial coordinates as... The voxel has a truncation filter function as follows: ,in, Represents the original 3D CT volume data in spatial coordinates The voxel HU value at the location; The upper limit threshold for high-density truncation is set; The set lower threshold for low-density truncation; This represents the voxel HU value output after upper and lower threshold truncation filtering. Upper and lower thresholds are set based on the density distribution characteristics of lung tissue, and nonlinear truncation suppresses interference from high-density tissue and external air. Subsequently, to adapt the data to the activation space of the deep neural network, the max-min feature scaling operator is used to linearly normalize the truncated volume data. , in, The output voxel values ​​after linear normalization are mapped to a range that is... between; The HU values ​​are the truncated voxels. After mapping, all voxel values ​​are... This improves the numerical stability of the model during gradient backpropagation.

3. The lung nodule detection method based on adaptive multi-scale deformable attention according to claim 2, characterized in that, The data preprocessing of the acquired dataset also includes employing a two-dimensional slice training strategy to address the dimensionality curse and resource consumption issues caused by directly inputting three-dimensional CT volume data into a deep network. First, the standardized three-dimensional voxel tensor of a single patient is divided into a set of independent two-dimensional slices along the axial direction. Each slice Suppose slice The corresponding real lung nodule labeling mask is To construct the training set, and to clearly distinguish between positive and negative sample slices, a set of positive sample slices is defined. With negative sample slice set Represented as: ， ， Finally, to address the class imbalance problem caused by negative sample slices, the negative sample slice set was... An adaptive sampling function based on the complexity of two-dimensional slice data structure is constructed. Specifically, it utilizes the local information entropy of two-dimensional slices. The complexity of tissues in slice data that are easily confused with blood vessels and trachea is expressed as: ,in, This represents the total number of gray levels in the slice. For slices The gray level in the lung mask area is The probability of pixel occurrence is determined, and then a dynamically preserved probability function is constructed. To determine whether the negative sample slice should be included in the final training set: , in, Indicates the first Each negative sample slice retains a dynamic probability of entering the final training set; This represents the local information entropy of the slice. The maximum information entropy value in the set of all negative sample slices; The calibration hyperparameter is used to control the overall proportion of negative samples. It is an extremely small base random sampling rate.

4. The lung nodule detection method based on adaptive multi-scale deformable attention according to claim 3, characterized in that, The construction of the deep network model based on AMDA-YOLO also includes introducing an adaptive multi-scale deformable attention (AMDA) mechanism in the network decoding stage to improve the model's feature decoding capability in multi-scale lung nodule detection tasks. By embedding AMDA in the feature fusion and decoding stages of the YOLO detection network, local detailed features and a larger range of contextual information are adaptively acquired while maintaining low computational cost. First, let the input feature map be... ,in Indicates batch size. Indicates the number of channels. and Let represent the height and width of the feature map, respectively. A 1×1 convolution is used for linear feature mapping, and the projected features are obtained by combining the GELU non-linear activation function. Subsequently, multi-scale features are extracted based on a parallel multi-branch structure, with each branch... Each contains a deformable convolution, and different kernel sizes and dilation rates are configured. To capture receptive fields at different scales, thereby modeling local details and global contextual information separately. The output of each branch is denoted as .

5. The lung nodule detection method based on adaptive multi-scale deformable attention according to claim 4, characterized in that, The construction of the deep network model based on AMDA-YOLO also includes, in order to achieve adaptive fusion of the outputs of multiple scale branches, combining the feature maps of each branch. Channel alignment is performed using AlignConv. Convolution performs a linear mapping on the feature map to unify the channel dimensions of each branch, i.e. ; Then, global average pooling is applied along the spatial dimension to obtain the corresponding feature description vector. The calculation formula is: ,in and The height and width of the feature map are respectively the description vectors of all branches. After cascading, a lightweight perceptron mapping network generates logits for each branch, which are then normalized using a softmax function to obtain the attention weights for each branch. The calculation process culminates in the combination of features from each branch using a weighted fusion method to obtain the final multi-scale fused feature representation, which is defined as follows: The fused features are mapped to a spatial weight map via 1×1 convolution, and then multiplied element-wise with the input features to achieve spatial recalibration, thereby enhancing the response of the target region in the feature map. To integrate with existing backbone networks and various modular structures, AMDA is used for encapsulation in the form of attention sublayers, the complete form of which is: ,in A learnable scaling factor. It is a 1×1 projection mapping.

6. The lung nodule detection method based on adaptive multi-scale deformable attention according to claim 5, characterized in that, The construction of the deep network model based on AMDA-YOLO also includes addressing the limitations of bounding box regression based on the Intersection over Union (IoU) loss function for lung nodule detection tasks. Building upon the traditional IoU regression approach, a scale-aware and shape consistency loss function, SASLose, is proposed to enhance the model's ability to finely regress the bounding boxes of small lung nodules. First, let the model's predicted bounding box be... The real frame is The center width and height representation is converted to the top-left / bottom-right corner representation for calculating the intersection-union ratio (IU). The IU calculation formula between the predicted bounding box and the ground truth bounding box is expressed as: Define the width and height of the minimum envelope box as follows: Its diagonal square is The square of the predicted distance from the center of the reference box is To simultaneously characterize aspect ratio differences, we define: Based on this, a scaling factor is introduced. The effect of adjusting the aspect ratio on the final measurement is expressed as: ,in It is a constant, through scale-shape adaptive dynamic weights. This allows SASLows to perceive the absolute scale of the target while simultaneously evaluating the relative matching quality of the predicted bounding box and the ground truth bounding box in terms of size and shape.

7. The lung nodule detection method based on adaptive multi-scale deformable attention according to claim 6, characterized in that, The scale-shape adaptive dynamic weights It contains three components: basic weights This is used to encode the absolute scale information of the target into the loss function, and is defined as: ,in It's about hyperparameters. It is a scaling factor; a scale-aware factor. , used to measure the area deviation between the predicted bounding box and the ground truth bounding box, is defined as: ,in, and These represent the width and height of the actual lung nodule annotation frame, respectively; and These represent the width and height of the model's predicted bounding box, respectively, and the shape consistency factor. This is used to generate predicted bounding boxes that conform to morphological priors. It calculates the aspect ratio difference between the predicted and ground truth bounding boxes and models this difference using an exponential decay function, thus penalizing aspect ratio deviations. This is expressed as: ,in, Indicates the aspect ratio of the prediction box; This represents the aspect ratio of the true bounding box. By calculating the absolute difference between the two aspect ratios and combining it with exponential decay, predicted boxes that deviate from the prior true shape are quantified and penalized. Then, the three are combined. Then to A two-sided truncation strategy is employed to enhance the stability of the training process, constraining it within a preset effective range: Finally, the proposed definition of SASLoss is: ,in, This is the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. The square of the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; It is the square of the diagonal distance of the smallest closure region that can simultaneously contain both the predicted and ground truth boxes; A parameter used to measure the relative difference in aspect ratio; An adaptive scaling factor to balance the influence of aspect ratio; To control the boundary hyperparameters of the cutoff range; To combine scale-aware and shape consistency factors, dynamic truncation weights are used to adaptively adjust the strength of each penalty term during regression; when the predicted bounding box differs significantly from the true nodule in size or shape, the dynamic adjustment factor... The relative weights of each penalty term will be adaptively adjusted to prioritize correcting the geometric properties with the most prominent deviations.

8. The lung nodule detection method based on adaptive multi-scale deformable attention according to claim 7, characterized in that, The method employs a linear preheating strategy to train a deep network model based on AMDA-YOLO, including using the AdamW optimizer with a weight decay mechanism to update the model's parameters. This is done to suppress overfitting caused by excessive parameters in complex networks. Let the first... The network parameters for the next iteration are: Batch input is The corresponding real label is The class probabilities are obtained through network forward propagation. Candidate boxes and confidence level According to the classification of losses Target confidence loss and bounding box regression loss The weighted sum is used to calculate the total loss: And calculate the gradient: Finally enter First-order and second-order momentum estimations are used, combined with weight decay terms to update parameters, enabling the model to converge simultaneously in three directions: classification confidence, target presence probability, and bounding box geometric consistency. A linear warmup strategy is employed in the early training stages to smoothly increase the learning rate over the first few epochs, avoiding gradient explosion caused by initial random weights. Let the total number of training epochs be... The number of preheating cycles is The base learning rate is The minimum learning rate is The preheating stage is carried out according to... The proportion will steadily increase the learning rate from 0 to After the warm-up, an improved cosine annealing learning rate decay strategy is adopted, which makes the learning rate decrease smoothly and non-linearly with each training epoch. The current learning rate is calculated through the cosine annealing term. : In AMP training, forward computation is employed. Generate a loss and combine it with a dynamic loss scaling factor. Amplify the loss value, backpropagate, and then apply the gradient according to... To balance numerical stability and training efficiency, the model generates candidate detection boxes and predicts the corresponding class probabilities and location parameters in each forward propagation iteration. The total training error of the network is calculated jointly by three parts: binary cross-entropy is used to calculate the classification loss and target confidence loss, scale-aware and shape consistency loss is used to calculate the bounding box regression loss, and the gradient of the joint loss with respect to the network parameters is calculated and updated using the backpropagation algorithm. The total training loss function for a single forward propagation of the network is defined as follows: The weighted sum of the three types of losses: ,in, To predict the binary cross-entropy loss between the class and the true class, To predict the binary cross-entropy loss between confidence level and true target probability, The loss is calculated as the scale perception and shape consistency regression loss. Each factor is assigned a pre-defined loss weight, and the total loss is calculated during the backpropagation phase using the chain rule. Learnable parameters of the network gradient Subsequently, the AdamW optimizer utilizes the modified first-order momentum estimate and second-order momentum estimation Combined with the set weight decay coefficient The parameter is updated using the following formula: , in, The current learning rate is dynamically adjusted after linear preheating and cosine annealing strategies. To prevent the minimum constant with a denominator of zero.

9. A lung nodule detection system based on adaptive multi-scale deformable attention, executing the lung nodule detection method based on adaptive multi-scale deformable attention as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire a clinical lung nodule dataset, which includes multi-source clinical lung nodule data containing CT body data, nodule annotation information, and associated examination text information. The preprocessing module is configured to preprocess the acquired dataset. The model building module is configured to build deep network models based on AMDA-YOLO; The model training module is configured to train a deep network model based on AMDA-YOLO using a linear warm-up strategy. The detection module is configured to use a trained model to detect lung nodules and output the detection results.

Images