An adaptive multi-scale attention lung cancer detection method and system

Abstract

An adaptive multi-scale attention-based lung cancer detection method and system, belonging to the field of medical imaging technology, addresses the problems of poor adaptability, insufficient reliability, weak interpretability, poor calibration performance, and imbalance between generalization and efficiency in existing models. The method includes: acquiring raw CT images, performing preprocessing to obtain preprocessed images and biomarker features; extracting multi-scale features from the preprocessed images to obtain multi-scale feature maps, performing adaptive weighting to obtain optimized features, performing uncertainty modeling and fusion to obtain fused features, performing hierarchical resolution processing to obtain fine-grained features; aligning biomarker features with fine-grained features to obtain interpretable features; converting interpretable features into token sequences and modeling global dependencies to obtain detection classification results and uncertainty estimates. This invention is applicable to lung nodule detection scenarios.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging technology, specifically relating to an adaptive multi-scale attention lung cancer detection technology. Background Technology

[0002] Lung cancer is the leading cause of cancer death worldwide, accounting for about a quarter of all cancer-related deaths. Early detection through low-dose computed tomography (CT) scans can reduce mortality by up to 20%. However, manually interpreting CT images presents challenges such as inter-observer variability, high false positive rates, and difficulty in identifying tiny or subsolid nodules less than 5 mm in diameter within complex anatomical regions.

[0003] With the development of deep learning technology, Convolutional Neural Networks (CNNs) have revolutionized the field of medical imaging with their feature extraction and classification capabilities, while Visual Transformers (ViTs) based on self-attention mechanisms have global contextual understanding capabilities. Combining the two can further improve model performance.

[0004] Existing hybrid convolutional neural network-visual transformer (CNN-ViT) methods have been applied to lung nodule detection, but these methods still have shortcomings in adapting nodule features, quantifying diagnostic confidence, and integrating radiological knowledge, making it difficult to meet clinical demands for high reliability, interpretability, and robustness in diagnostic models. Related technologies originate from recent patent literature and journal articles in the fields of medical imaging and deep learning, such as research on the application of multi-scale CNNs and ViT in medical imaging.

[0005] Defects and shortcomings of existing technology: (1) Static architecture: The existing hybrid model adopts a fixed structure, which cannot adaptively adjust the receptive field according to the diverse characteristics of nodules such as shape, size, and density, resulting in insufficient adaptability to small subsolid nodules and large spiculated nodules. (2) Lack of uncertainty modeling: Some prediction results do not quantify diagnostic confidence, and cannot distinguish between random uncertainty caused by inherent noise in CT images and cognitive uncertainty caused by insufficient training data, which limits the credibility of clinical applications; (3) Insufficient integration of radiological knowledge: The model does not fully incorporate key biomarker information such as spiculation, ground-glass opacity, and calcifications, resulting in poor interpretability and difficulty in conforming to the knowledge system of radiologists. Summary of the Invention

[0006] This invention provides an adaptive multi-scale attention lung cancer detection method and system. Based on the AMSA-ViT uncertainty-aware lung cancer CT detection model, it solves the technical problems of poor adaptability, insufficient reliability, weak interpretability, poor calibration performance, and imbalance between generalization and efficiency of existing models by dynamically adjusting the receptive field, quantifying the prediction uncertainty, integrating radiological biomarkers, and hierarchically learning multi-resolution features. This achieves high accuracy and high safety in lung cancer CT image detection.

[0007] Firstly, the present invention aims to provide an adaptive multi-scale attention-based lung cancer detection method, comprising the following steps: S1: Acquire raw CT images, perform preprocessing, and obtain preprocessed images and biomarker features; S2: Perform multi-scale feature extraction on the preprocessed image to obtain a multi-scale feature map; S3: Adaptively weight the multi-scale feature maps to obtain optimized features; S4: Perform uncertainty modeling and fusion of the optimized features to obtain fused features; S5: Perform hierarchical resolution processing on the fused features to obtain fine-grained features; S6: Align biomarker features with fine-grained features to obtain interpretable features; S7: Convert interpretable features into token sequences and model global dependencies to obtain detection and classification results, as well as uncertainty estimates.

[0008] Furthermore, a preferred solution is provided: the preprocessing includes: adaptive histogram equalization, multi-scale Gaussian filtering, 3D U-Net anatomical segmentation, and radiomics feature extraction.

[0009] Furthermore, a preferred solution is provided: In S2, three backbone networks are used to extract multi-scale features, namely: EfficientNetB0, DenseNet121, and ResNet50V2. The preprocessed image is input into EfficientNetB0, DenseNet121, and ResNet50V2 respectively to obtain three feature maps.

[0010] Furthermore, a preferred embodiment is provided: S3 includes: S31: The importance score of each feature map is calculated by a learnable scoring function consisting of two fully connected layers and a sigmoid activation function, and then the attention weight corresponding to each feature map is obtained by Softmax normalization. S32: Based on the attention weights corresponding to each feature map, perform a weighted summation of each feature map to obtain the optimized features.

[0011] Furthermore, a preferred embodiment is provided: S4 includes: S41: Capture the inherent noise of the original CT image, resulting in an image containing random uncertainty; S42: Model the uncertainty of the model caused by insufficient training data or knowledge gaps to obtain cognitive uncertainty; S43: The predicted values ​​are weighted and summed by a learnable fusion weight layer to obtain the fusion features and uncertainty coefficients.

[0012] Furthermore, a preferred embodiment is provided: S5 includes: S51: Perform 3D convolutional downsampling on the input features to capture global anatomical context information; S52: Perform deconvolution upsampling on the low-resolution features and concatenate it with the original input features; S53: Perform deconvolution upsampling on medium-resolution stitched features to distinguish fine-grained texture and morphological features; S54: Perform data enhancement and optimization on the original CT images, including random rotation, scale scaling, elastic deformation, and controlled noise injection, to obtain fine-grained features.

[0013] Furthermore, a preferred embodiment is provided: S7 includes: S71: Tokenize the interpretable features to obtain several tokens; S72: The classification results of normal, benign, and malignant diseases, as well as the uncertainty estimate, are obtained by calculating two fully connected layers and the Softmax activation function.

[0014] Secondly, the purpose of this invention is to propose an adaptive multi-scale attention lung cancer detection system, said system being implemented based on an adaptive multi-scale attention lung cancer detection method as described in any one or more of the above-mentioned schemes, said system comprising: Data input and preprocessing module: used to acquire raw CT images, perform preprocessing, and obtain preprocessed images and biomarker features; Multi-scale feature extraction module: used to extract multi-scale features from the preprocessed image to obtain multi-scale feature maps; Dynamic scale-aware attention module: used to adaptively weight multi-scale feature maps to obtain optimized features; Uncertainty Quantification Feature Fusion Module: Used to perform uncertainty modeling and fusion of optimized features to obtain fused features; Progressive multi-resolution learning module: used to perform hierarchical resolution processing on fused features to obtain fine-grained features; Biomarker-guided knowledge distillation module: used to align biomarker features with fine-grained features to obtain interpretable features; The Visual Transformer processing and prediction output module is used to convert interpretable features into token sequences, model global dependencies, obtain detection and classification results, and uncertainty estimates.

[0015] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes an adaptive multi-scale attention lung cancer detection method according to any one or more of the above-described schemes.

[0016] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes an adaptive multi-scale attention lung cancer detection method as described in any one or more of the above-described schemes.

[0017] Compared with the prior art, the advantages of the present invention are: (1) Provide a dynamic adaptation mechanism so that the model can adjust the receptive field according to the nodule characteristics and improve the detection adaptability to different types of nodules.

[0018] This invention introduces a dynamic scale-aware attention mechanism, enabling the model to adaptively adjust its receptive field based on differences in morphology, size, and density of lung nodules. This allows it to simultaneously adapt to the detection needs of both small subsolid nodules and structurally complex nodules, overcoming the problems of fixed model structure and insufficient adaptability in existing technologies.

[0019] (2) The diagnostic performance is stable and meets the high sensitivity requirements of lung cancer screening.

[0020] Experimental results based on the IQ-OTHNCCD lung cancer dataset show that the method of this invention achieves high classification accuracy in the three-class classification task, with 100% sensitivity to malignant samples and high recognition ability for normal samples. It can effectively reduce the risk of missed diagnosis of malignant nodules in actual screening scenarios, while maintaining the stability of overall classification performance.

[0021] (3) Quantify the two types of uncertainty and generate confidence-calibrated prediction results to improve the reliability of clinical decision-making.

[0022] This invention uses an uncertainty quantification feature fusion module to model and fuse random uncertainty caused by inherent noise in CT images and cognitive uncertainty caused by insufficient training data. This enables the model to provide corresponding uncertainty information while outputting classification results, which helps to improve the credibility of prediction results in clinical applications.

[0023] (4) The prediction confidence is calibrated to reduce the clinical risk of false positives.

[0024] This invention introduces a calibration strategy after model training to maintain a high consistency between the predicted confidence level and the actual correct probability. Its expected calibration error is significantly lower than the safety threshold commonly used in clinical applications, thereby effectively reducing the risk of unnecessary intervention in clinical procedures caused by false positive predictions while maintaining high sensitivity.

[0025] (5) Integrate radiological biomarkers to enhance the clinical interpretability of the model.

[0026] This invention introduces radiomics features and prior radiological knowledge into the model training process through a biomarker-guided knowledge distillation mechanism. This enables the model to learn feature expressions consistent with clinical diagnostic logic, which helps radiologists understand the basis of the model's predictions and improves the acceptability of the results interpretation.

[0027] (6) A multi-resolution hierarchical learning strategy is adopted to balance detection accuracy and computational efficiency.

[0028] This invention employs a progressive multi-resolution learning framework to perform layered processing on CT images. While acquiring global structural information, it retains local fine-grained features, achieving a balance between detection accuracy and computational efficiency, and improving the feasibility of the model in practical applications.

[0029] This invention is applicable to the detection of lung nodules. Attached Figure Description

[0030] 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 the structures shown in these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating the AMSA-ViT model processing described in a specific embodiment of the present invention; Figure 2 This is a diagram illustrating the preprocessed image effect described in a specific embodiment of the present invention. Figure 3 The present invention provides receiver operating characteristic (ROC) curves for the AMSA-ViT model in three types of samples: normal tissue, benign nodules, and malignant nodules, as described in a specific embodiment of the invention. Detailed Implementation

[0032] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0033] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0035] Implementation Method 1 Reference Figure 1 This implementation method is described below.

[0036] This embodiment proposes an adaptive multi-scale attention-based lung cancer detection method, including the following steps: S1: Acquire raw CT images, perform preprocessing, and obtain preprocessed images and biomarker features; S2: Perform multi-scale feature extraction on the preprocessed image to obtain a multi-scale feature map; S3: Adaptively weight the multi-scale feature maps to obtain optimized features; S4: Perform uncertainty modeling and fusion of the optimized features to obtain fused features; S5: Perform hierarchical resolution processing on the fused features to obtain fine-grained features; S6: Align biomarker features with fine-grained features to obtain interpretable features; S7: Convert interpretable features into token sequences and model global dependencies to obtain detection and classification results, as well as uncertainty estimates.

[0037] Implementation Method 2 Reference Figure 2 This implementation method is described below.

[0038] This embodiment is a further illustrative example of the preprocessing in S1 of the adaptive multi-scale attention lung cancer detection method described in Embodiment 1.

[0039] The preprocessing described in this embodiment includes: adaptive histogram equalization, multi-scale Gaussian filtering, 3D U-Net anatomical segmentation, and radiomics feature extraction.

[0040] Furthermore, the specific process of the preprocessing is as follows: S11, Adaptive Histogram Equalization (AHE), is expressed by the following formula:

[0041] in, Original pixel intensity For local minimum intensity, For local maximum intensity, The grayscale level is used in this embodiment. .

[0042] S12, multi-scale Gaussian filtering, is expressed by the following formula: , in, For Gaussian smoothed images, For the input image, Represents convolution operation. With a mean of 0 and a variance of Gaussian kernel.

[0043] S13, 3D U-Net Anatomy Segmentation: In this step, the encoding path contains 4 convolutional blocks, each of which contains 2 3D convolutional layers, a BatchNorm layer, and a ReLU activation function. The decoding path contains 4 deconvolutional blocks, and local features are fused through skip connections.

[0044] S14, Imagemics Feature Extraction, including texture features, intensity features, and morphological features. Texture features are extracted using gray-level co-occurrence matrix and gray-level run-length matrix parameters; intensity features include mean, variance, and kurtosis; morphological features include volume, surface area, and sphericity.

[0045] Figure 2 The images processed using the preprocessing flow described in this embodiment are: (a) the original CT image, (b) the image after Gaussian filtering and smoothing, (c) the image after contrast-limited adaptive histogram equalization (CLAHE) enhancement, (d) the binary mask generated by Otsu's thresholding method, (e) morphological operations, and (f) the final processed image.

[0046] The preprocessing process described in this embodiment is a key preparation step for the input data, as detailed below: (1) Import raw CT images in DICOM format ( Figure 2 (a)), with a resolution of 512×512; (2) Noise was removed using a Gaussian filter with variance σ = 1.0, resulting in... Figure 2 (b) shows the smoothed image; (3) Perform adaptive histogram equalization (AHE) to enhance local contrast, resulting in Figure 2 (c) shows the enhanced image; (4) Segment the lung region using 3D U-Net and generate... Figure 2 (d) shows a binary mask that focuses the region of interest of the nodule; (5) Extract texture, intensity, and morphological features from the masked region to form a biomarker vector and output it. Figure 2 (f) shows the final processed image with a resolution of 128×128×128.

[0047] Implementation Method 3 This embodiment is a further illustrative example of S2 and S3 in the adaptive multi-scale attention lung cancer detection method described in Embodiment 1.

[0048] In step S2, three backbone networks are used to extract multi-scale features: EfficientNetB0, DenseNet121, and ResNet50V2. The preprocessed image is input into EfficientNetB0, DenseNet121, and ResNet50V2 respectively to obtain three feature maps.

[0049] The S3 is implemented through a dynamic scale-aware attention (DSA) module and a feature weighted fusion module, including: S31: The importance score of each feature map is calculated by a learnable scoring function consisting of two fully connected layers and a sigmoid activation function, and then the attention weight corresponding to each feature map is obtained by Softmax normalization. Furthermore, in this step, the DSA module adaptively selects the receptive field based on the size and density of the nodules, enabling the model to simultaneously capture small subsolid nodules and large spiculated nodules. The formula for calculating the attention weights of the feature maps is as follows: , in For the first The attention weights corresponding to each feature / scale represent the importance that the model assigns to that feature / scale, and all of them The sum is 1; For features / scale The original fractions before normalization; It is a learnable scoring function implemented by the Dynamic Scale Aware Attention (DSA) module, whose role is to calculate the relative importance of each scale and generate a score. Subsequently, through execution Normalization operation transforms the original score Convert to probability value This allows features with higher scores to receive more attention weight.

[0050] S32: Based on the attention weights corresponding to each feature map, perform a weighted summation of each feature map to obtain the optimized features.

[0051] The optimized feature dimension obtained in this embodiment is 256×64×64, where the complex nodule structure corresponds to higher weights in ResNet50V2 (average weight 0.350) and DenseNet121 (average weight 0.605), and the simple feature corresponds to EfficientNetB0 (average weight 0.044).

[0052] This implementation provides a dynamic adaptation mechanism, enabling the model to adjust its receptive field according to nodule characteristics, thereby improving its adaptability to the detection of different types of nodules.

[0053] Implementation Method 4 This embodiment is a further illustrative example of S4 in the adaptive multi-scale attention lung cancer detection method described in Embodiment 1.

[0054] S4 in this embodiment includes: S41: Capture the inherent noise of the original CT image, resulting in an image containing random uncertainty. ; S42: Model the uncertainty caused by insufficient training data or knowledge gaps to obtain cognitive uncertainty. ; S43: The three predicted values ​​are weighted and summed using a learnable fusion weight layer to obtain the fusion feature (dimension 256×64×64) and the uncertainty coefficient. The range is 0 to 0.1), and the formula for weighted summation is: , in For fusion features with uncertainty quantification, To be allocated to each Learnable weights These are the model predictions that contribute to the final output.

[0055] This implementation quantifies two types of uncertainty, generates confidence-calibrated predictions, and improves the reliability of clinical decision-making.

[0056] Implementation Method 5 This embodiment is a further illustrative example of S5 in the adaptive multi-scale attention lung cancer detection method described in Embodiment 1.

[0057] S5 in this embodiment includes: S51, Low-resolution processing: Perform 3D convolution downsampling with stride 2 on the input features to 32×32×32 to capture global anatomical context information; S52, Medium resolution processing: Deconvolution upsampling is performed on low-resolution features to 64×64×64, and then concatenated with the original input features to optimize the morphology and boundaries of nodule regions; S53, high-resolution processing: deconvolution upsampling is performed on medium-resolution stitched features to 128×128×128 to distinguish fine-grained texture and morphological features; S54, Perform data augmentation and optimization on the original CT images: During the model training phase, perform the following operations on the input CT images: Random rotation: Press Randomly adjust the image angle; Scale: Randomly scale the image by a ratio of 0.8 to 1.2; Elastic deformation: The image is elastically distorted by a coefficient of 0.1; Controllable noise injection: Gaussian noise with a mean of 0 and a variance of 0.01 is added to obtain a high-resolution fine-grained feature map, which balances detection accuracy, computational efficiency and generalization ability.

[0058] This implementation achieves hierarchical and efficient learning of multi-resolution features, balancing detection accuracy and computational efficiency.

[0059] Implementation Method Six This embodiment is a further illustrative example of S6 in the adaptive multi-scale attention lung cancer detection method described in Embodiment 1.

[0060] S6 in this embodiment includes: S61, Biomarker Extraction: Extract texture, intensity and morphological features from nodule regions after anatomical segmentation by 3D U-Net to form a set of biomarkers; S62, Bioguided Feature Generation: Converting extracted biomarkers into bioguided feature representations based on prior radiological knowledge. ; S63, Network Feature Extraction: Obtaining features from the visual Transformer processing module. L Feature representation of a selected layer ; S64, Distillation loss calculation, expressed by the formula: , S65, Feature Alignment Optimization: [This section appears to be incomplete and requires further context.] By incorporating the model's total loss function and optimizing the network parameters through backpropagation, and Align the features and output a clinically interpretable feature map, meaning the features can be interpreted.

[0061] This implementation integrates radiological biomarkers to enhance model interpretability, reduce false positive rates and the risk of missed diagnoses, and optimize model calibration performance.

[0062] Implementation Method Seven This embodiment is a further illustrative example of S7 in the adaptive multi-scale attention lung cancer detection method described in Embodiment 1.

[0063] S7 in this embodiment includes: S71: Tokenize the interpretable features to obtain several tokens, as shown in the formula: , The batch size is bs=16, each sample generates N=128×128=16384 tokens, and the token embedding dimension is d=256.

[0064] S72: Employs a Transformer encoder, using two fully connected layers (512 hidden layers) and a Softmax activation function to calculate the predicted probabilities for normal, benign, and malignant cases. ).

[0065] The Transformer encoder is represented as: , in Input sequence for tokens, For Transformer processing modules; The formula for self-attention is expressed as: , Where Q, K, and V are the query, key, and value matrices, and V is the number of attention heads.

[0066] Finally, the classification head outputs three prediction results: normal, benign, and malignant, along with corresponding uncertainty estimates to assist in subsequent diagnostic decisions.

[0067] Implementation Method Eight Reference Figure 1 , Figure 2 , Figure 3 This implementation method is described below.

[0068] This embodiment is a further illustrative example of the adaptive multi-scale attention lung cancer detection method described in any one or more of the embodiments from Embodiment 1 to Embodiment 7.

[0069] Preprocessed image ( Figure 2 (f) Features are extracted through a multi-scale backbone network, and after being fused with attention weighting and uncertainty, 256-dimensional fine-grained features are output; The fine-grained features are processed by a 6-layer encoder (8-head self-attention) to obtain three initial prediction probabilities; The probability is calibrated using a temperature scaling method (temperature coefficient T=1.2), and the final diagnostic result and confidence level are output (confidence level ≥0.95 indicates high reliability). Based on 330 samples from the test set. Figure 3 The ROC curves shown indicate that the AUC for all three classes of samples is ≥0.999, demonstrating that the model's diagnostic accuracy meets clinical needs.

[0070] Implementation Method Nine This embodiment proposes an adaptive multi-scale attention-based lung cancer detection system. The system is implemented based on the adaptive multi-scale attention-based lung cancer detection method described in any one or more of embodiments one through seven. The system includes: Data input and preprocessing module: performs quality optimization, region segmentation and biomarker extraction on the input CT images, and outputs the preprocessed images and biomarker features; Multi-scale feature extraction module: Receives the preprocessed image, extracts multi-scale features through three backbone networks, and outputs multi-dimensional feature representation; Dynamic scale-aware attention module: Adaptively weights multi-scale features and outputs optimized key features; Uncertainty Quantification Feature Fusion Module: Performs uncertainty modeling and fusion on optimized features, and outputs fused features with confidence scores; Progressive multi-resolution learning module: performs hierarchical resolution processing on fused features and outputs fine-grained features; Biomarker-guided knowledge distillation module: Aligns preprocessed biomarker features with fine-grained features to output clinically interpretable features; The visual Transformer processing module converts the above features into token sequences and models global dependencies. Prediction output module: Outputs three categories: normal / benign / malignant, along with an estimate of uncertainty.

[0071] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An adaptive multi-scale attention-based lung cancer detection method, characterized in that, Includes the following steps: S1: Acquire raw CT images, perform preprocessing, and obtain preprocessed images and biomarker features; S2: Perform multi-scale feature extraction on the preprocessed image to obtain a multi-scale feature map; S3: Adaptively weight the multi-scale feature maps to obtain optimized features; S4: Perform uncertainty modeling and fusion of the optimized features to obtain fused features; S5: Perform hierarchical resolution processing on the fused features to obtain fine-grained features; S6: Align biomarker features with fine-grained features to obtain interpretable features; S7: Convert interpretable features into token sequences and model global dependencies to obtain detection and classification results, as well as uncertainty estimates.

2. The adaptive multi-scale attention lung cancer detection method according to claim 1, characterized in that, The preprocessing includes: adaptive histogram equalization, multi-scale Gaussian filtering, 3D U-Net anatomical segmentation, and radiomics feature extraction.

3. The adaptive multi-scale attention lung cancer detection method according to claim 1, characterized in that, In step S2, three backbone networks are used to extract multi-scale features: EfficientNetB0, DenseNet121, and ResNet50V2. The preprocessed image is input into EfficientNetB0, DenseNet121, and ResNet50V2 respectively to obtain three feature maps.

4. The adaptive multi-scale attention lung cancer detection method according to claim 3, characterized in that, S3 includes: S31: The importance score of each feature map is calculated by a learnable scoring function consisting of two fully connected layers and a sigmoid activation function, and then the attention weight corresponding to each feature map is obtained by Softmax normalization. S32: Based on the attention weights corresponding to each feature map, perform a weighted summation of each feature map to obtain the optimized features.

5. The adaptive multi-scale attention lung cancer detection method according to claim 1, characterized in that, S4 includes: S41: Capture the inherent noise of the original CT image, resulting in an image containing random uncertainty; S42: Model the uncertainty of the model caused by insufficient training data or knowledge gaps to obtain cognitive uncertainty; S43: The predicted values ​​are weighted and summed by a learnable fusion weight layer to obtain the fusion features and uncertainty coefficients.

6. The adaptive multi-scale attention lung cancer detection method according to claim 1, characterized in that, S5 includes: S51: Perform 3D convolutional downsampling on the input features to capture global anatomical context information; S52: Perform deconvolution upsampling on the low-resolution features and concatenate it with the original input features; S53: Perform deconvolution upsampling on medium-resolution stitched features to distinguish fine-grained texture and morphological features; S54: Perform data enhancement and optimization on the original CT images, including random rotation, scale scaling, elastic deformation, and controlled noise injection, to obtain fine-grained features.

7. The adaptive multi-scale attention lung cancer detection method according to claim 1, characterized in that, S7 includes: S71: Tokenize the interpretable features to obtain several tokens; S72: The classification results of normal, benign, and malignant diseases, as well as the uncertainty estimate, are obtained by calculating two fully connected layers and the Softmax activation function.

8. An adaptive multi-scale attention-based lung cancer detection system, characterized in that, The system is implemented based on an adaptive multi-scale attention lung cancer detection method as described in any one of claims 1-7, and the system includes: Data input and preprocessing module: used to acquire raw CT images, perform preprocessing, and obtain preprocessed images and biomarker features; Multi-scale feature extraction module: used to extract multi-scale features from the preprocessed image to obtain multi-scale feature maps; Dynamic scale-aware attention module: used to adaptively weight multi-scale feature maps to obtain optimized features; Uncertainty Quantification Feature Fusion Module: Used to perform uncertainty modeling and fusion of optimized features to obtain fused features; Progressive multi-resolution learning module: used to perform hierarchical resolution processing on fused features to obtain fine-grained features; Biomarker-guided knowledge distillation module: used to align biomarker features with fine-grained features to obtain interpretable features; The Visual Transformer processing and prediction output module is used to convert interpretable features into token sequences, model global dependencies, obtain detection and classification results, and uncertainty estimates.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes an adaptive multi-scale attention lung cancer detection method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes an adaptive multi-scale attention lung cancer detection method according to any one of claims 1-7.

Images