A method, device, equipment, and medium for 3D lung nodule identification based on Mamba.
By preprocessing and enhancing low-dose chest CT images, combined with iterative training of the Mamba state-space multi-directional scanning recognition network, the problem of insufficient accuracy in lung nodule recognition in existing technologies has been solved, and the accuracy and stability of lung nodule benign and malignant identification have been improved.
Patent Information
- Application Number
- CN202611332380.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
Existing lung nodule identification methods based on three-dimensional convolutional neural networks are unable to effectively model long-range dependencies, resulting in insufficient accuracy in identifying benign and malignant nodules.
By performing tissue library construction, nodule cropping, spatial resampling, and CT value standardization on low-dose chest CT images, a standardized nodule set is generated. The pose distribution of the nodule samples is expanded by multi-axial random rotation and spherical random translation enhancement methods. A Mamba state space multi-directional scanning recognition network is built and iteratively trained to construct a target 3D nodule recognition model.
While maintaining low computational complexity, this study effectively models the long-range dependencies between pulmonary nodule voxels, improving the accuracy and stability of identifying benign and malignant pulmonary nodules in low-dose chest CT images.
Smart Images

Figure CN123118539A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, device and medium for 3D lung nodule identification based on Mamba. Background Technology
[0002] Lung cancer is one of the leading causes of death and morbidity among malignant tumors worldwide, and early prediction is crucial for improving survival rates. Pulmonary nodules are common lung abnormalities in clinical imaging, and their benign or malignant determination is essential for early screening and treatment pathway selection in lung cancer. Low-dose computed tomography (LDCT) is increasingly becoming the preferred method for early lung cancer screening. Compared to conventional CT, LDCT can significantly reduce radiation dose while obtaining clearer information about lung structures, making it effective for routine screening of high-risk populations. However, LDCT images suffer from relatively low image quality, high noise levels, and blurred lesion margins in practical applications, posing significant challenges for physicians in interpreting the benign or malignant nature of pulmonary nodules. Traditional manual image interpretation not only relies on the professional experience of radiologists, is highly subjective and has poor repeatability, but is also inefficient when facing large-scale screening tasks, failing to meet the needs of precision medicine.
[0003] To improve the automation and diagnostic efficiency of benign and malignant lung nodule identification, existing technologies typically employ lung nodule identification schemes based on three-dimensional convolutional neural networks (CNNs). Models such as 3D CNN, ResNet, and DenseNet are applied to feature extraction and benign / malignant classification of lung nodules in low-dose chest CT images. However, existing CNN-based lung nodule identification methods are limited by the local receptive field of the convolution operator, making it difficult to effectively model the long-range dependencies between voxels in lung nodules in low-dose chest CT images. Their ability to capture the overall three-dimensional structure and morphological evolution of lung nodules is insufficient. Furthermore, stacking network layers to expand the receptive field drastically increases computational cost and memory usage, thus limiting the model's ability to fully express the multi-level semantic features of lung nodules. Consequently, the accuracy of benign / malignant lung nodule identification in low-dose chest CT images is difficult to further improve. Summary of the Invention
[0004] This invention provides a 3D lung nodule identification method, device, equipment, and medium based on Mamba, to solve the technical problem that existing lung nodule identification methods are unable to effectively model long-range dependencies, resulting in insufficient accuracy in identifying benign and malignant nodules.
[0005] Firstly, a 3D lung nodule identification method based on Mamba is provided, including: Low-dose chest CT images were acquired and tissue libraries were constructed by comparing them with corresponding benign and malignant identification information to obtain a set of labeled lung nodules. The lung nodule sample set was subjected to nodule block trimming, spatial resampling, and CT value standardization to obtain a standardized nodule block set. The standardized nodule block set is enhanced by multi-axial random rotation and spherical random translation to obtain the lung nodule training tensor set; Based on the lung nodule training tensor set, a Mamba state space multi-directional scanning recognition network was built to obtain an initial 3D nodule recognition model. The lung nodule training tensor set is input into the initial 3D nodule recognition model for iterative training to obtain the target 3D nodule recognition model. The CT image to be identified is input into the target 3D nodule recognition model for recognition, and the lung nodule recognition result is output.
[0006] Secondly, a Mamba-based 3D lung nodule recognition device is provided, comprising: The sample acquisition module is used to acquire low-dose chest CT images and construct a tissue library with the corresponding benign and malignant identification information to obtain an labeled lung nodule sample set. The sample processing module is used to perform nodule block trimming, spatial resampling and CT value standardization on the lung nodule sample set to obtain a standardized nodule block set. The sample augmentation module is used to perform multi-axial random rotation and spherical random translation augmentation on the standardized nodule block set to obtain the lung nodule training tensor set. The model building module is used to build a Mamba state-space multi-directional scanning recognition network based on the training tensor set of the lung nodules to obtain an initial 3D nodule recognition model. The model training module is used to input the lung nodule training tensor set into the initial 3D nodule recognition model for iterative training to obtain the target 3D nodule recognition model. The result recognition module is used to input the CT image to be recognized into the target 3D nodule recognition model for recognition and output the lung nodule recognition result.
[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described Mamba-based 3D lung nodule identification method.
[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described Mamba-based 3D lung nodule identification method.
[0009] The above-mentioned Mamba-based 3D lung nodule recognition method, device, equipment, and medium include: acquiring low-dose chest CT images and constructing a tissue library with corresponding benign / malignant identification information to obtain an annotated lung nodule sample set; performing nodule block clipping, spatial resampling, and CT value standardization on the lung nodule sample set to obtain a standardized nodule block set; enhancing the standardized nodule block set with multi-axial random rotation and spherical random translation to obtain a lung nodule training tensor set; constructing a Mamba state-space multi-directional scanning recognition network based on the lung nodule training tensor set to obtain an initial 3D nodule recognition model; inputting the lung nodule training tensor set into the initial 3D nodule recognition model for iterative training to obtain a target 3D nodule recognition model; inputting the CT image to be recognized into the target 3D nodule recognition model for recognition, and outputting the lung nodule recognition result. In this invention, to address the problem that existing lung nodule identification methods struggle to effectively model long-range dependencies, leading to insufficient accuracy in benign / malignant nodule identification, a standardized set of nodule blocks with uniform scale and grayscale is provided for subsequent model training by performing tissue library construction, nodule block cropping, spatial resampling, and CT value standardization on the acquired low-dose chest CT images. This is achieved through multi-axial random rotation and spherical random translation enhancement methods to expand the pose distribution and spatial diversity of the nodule samples, resulting in a lung nodule training tensor set with strong generalization capabilities. Based on this training tensor set, a Mamba state-space multi-directional scanning identification network is constructed. Leveraging the Mamba state-space multi-directional scanning mechanism, long-range dependencies between lung nodule voxels are effectively modeled while maintaining low computational complexity, fully capturing the overall three-dimensional structure and morphological evolution characteristics of the lung nodules. Iterative training yields a target 3D nodule identification model. Finally, the CT image to be identified is input into the target 3D nodule identification model for recognition, thereby improving the accuracy and stability of benign / malignant lung nodule identification in low-dose chest CT images. Attached Figure Description
[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention 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.
[0011] Figure 1 This is a schematic diagram of an application environment for a 3D lung nodule identification method based on Mamba, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating a 3D lung nodule identification method based on Mamba in one embodiment of the present invention. Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S10; Figure 4 yes Figure 2 A schematic diagram of a specific implementation method for step S20; Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 6 yes Figure 2 A schematic diagram of a specific implementation of step S40; Figure 7 yes Figure 2 A schematic diagram of a specific implementation method for step S50; Figure 8 yes Figure 2 A schematic diagram of a specific implementation method for step S60; Figure 9 This is a schematic diagram of a 3D lung nodule identification device based on Mamba according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 11 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] The Mamba-based 3D lung nodule identification method provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can obtain input conditions from the client, namely, acquiring low-dose chest CT images and constructing a tissue library with corresponding benign / malignant identification information to obtain an annotated lung nodule sample set; the lung nodule sample set is then processed by nodule block cropping, spatial resampling, and CT value standardization to obtain a standardized nodule block set; the standardized nodule block set is then enhanced by multi-axial random rotation and spherical random translation to obtain a lung nodule training tensor set; a Mamba state-space multi-directional scanning recognition network is built based on the lung nodule training tensor set to obtain an initial 3D nodule recognition model; the lung nodule training tensor set is input into the initial 3D nodule recognition model for iterative training to obtain a target 3D nodule recognition model; the CT image to be recognized is input into the target 3D nodule recognition model for recognition, and the lung nodule recognition result is output. In this invention, to address the problem that existing lung nodule identification methods struggle to effectively model long-range dependencies, leading to insufficient accuracy in benign / malignant nodule identification, a standardized set of nodule blocks with uniform scale and grayscale is provided for subsequent model training by performing tissue library construction, nodule block cropping, spatial resampling, and CT value standardization on the acquired low-dose chest CT images. This is achieved through multi-axial random rotation and spherical random translation enhancement methods to expand the pose distribution and spatial diversity of the nodule samples, resulting in a lung nodule training tensor set with strong generalization capabilities. Based on this training tensor set, a Mamba state-space multi-directional scanning identification network is constructed. Leveraging the Mamba state-space multi-directional scanning mechanism, long-range dependencies between lung nodule voxels are effectively modeled while maintaining low computational complexity, fully capturing the overall three-dimensional structure and morphological evolution characteristics of the lung nodules. Iterative training yields a target 3D nodule identification model. Finally, the CT image to be identified is input into the target 3D nodule identification model for recognition, thereby improving the accuracy and stability of benign / malignant lung nodule identification in low-dose chest CT images. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.
[0014] Please see Figure 2 As shown, Figure 2 A flowchart illustrating a Mamba-based 3D lung nodule identification method provided for an embodiment of the present invention includes the following steps: S10: Acquire low-dose chest CT images and construct a tissue library with the corresponding benign and malignant identification information to obtain a labeled lung nodule sample set.
[0015] Low-dose chest CT images from clinical screening scenarios were collected, and the collected low-dose chest CT images were associated with their corresponding benign and malignant identification information and unified into a library. This resulted in a lung nodule sample set labeled with benign and malignant characteristics, providing a data foundation for subsequent model training and validation.
[0016] Combination Figure 3 As shown, step S10 specifically includes: S101: Obtain the original DICOM image sequence of low-dose chest CT and integrate it to obtain the original CT image set.
[0017] DICOM (Digital Imaging and Communications in Medicine) is an internationally recognized standard format for storing and transmitting medical images and related information. After a low-dose chest CT scan, the acquired chest tomographic images are saved frame-by-frame as DICOM files, with each subject corresponding to a set of DICOM image sequences arranged in slice order. In this step, the original low-dose chest CT DICOM image sequences of the subjects are acquired in batches from the medical image archiving and communication system or clinical screening workstation. The original DICOM image sequences are then collected and organized using the subject number, scan date, and scan sequence number as indexes. All DICOM slices belonging to the same scan are integrated into a complete chest image unit, thus obtaining an original CT image set composed of image units from multiple subjects, providing the original data source for subsequent analysis and 3D reconstruction.
[0018] S102: Perform DICOM head analysis and volume data 3D reconstruction on the original CT image set to obtain the CT image set to be labeled.
[0019] The DICOM header refers to the metadata field in the DICOM file located before the pixel data, used to describe image acquisition and geometric parameters. It includes key parameters such as pixel spacing, slice spacing, scan orientation, origin and orientation of the user coordinate system, reslope, and reintercept. The volume data refers to the three-dimensional voxel data formed by stacking several two-dimensional tomographic slices generated from the same scan according to their spatial position. In this step, the DICOM header of each image unit in the original CT image set is parsed, the aforementioned geometric parameters and grayscale mapping parameters are extracted, and based on the extracted slice spacing and scan orientation, all two-dimensional slices belonging to the same image unit are stacked and geometrically aligned according to their actual spatial position in the user coordinate system. This completes the three-dimensional reconstruction of the volume data, resulting in a set of annotated CT images that can be observed and positioned in any section in three-dimensional space, on a patient-by-patient basis.
[0020] S103: The set of CT images to be labeled is associated with the preset benign and malignant information in the pathology report to build a library, thereby obtaining a set of CT images with labels.
[0021] The benign or malignant information in the pathology reports refers to the diagnostic information issued by the clinical pathology department based on puncture biopsy or surgical resection specimens, which clearly concludes whether the lung nodules are benign or malignant. This serves as the gold standard for identifying the benign or malignant nature of lung nodules. In this step, benign or malignant information from pathology reports corresponding to the set of CT images to be labeled is collected in advance. Using the patient's unique identifier and examination time as the association key, each CT image to be labeled is matched and bound to its corresponding benign or malignant information from the pathology report. If a patient has multiple lung nodules, the corresponding benign or malignant conclusion is bound according to the nodule number. After completing the association and matching, the CT images to be labeled and their benign or malignant labels are included in a unified image database for structured storage and management, thereby obtaining a labeled CT image set.
[0022] S104: The labeled CT image set is reviewed for nodule coordinate annotation and low-quality images are screened out to obtain a qualified labeled image set.
[0023] The nodule coordinate annotation refers to the annotation information recorded in the three-dimensional volume data as coordinate points or bounding boxes in the user's coordinate system, used for locating and cropping the nodule region in subsequent steps. In this step, the nodule coordinate annotation of each labeled CT image in the labeled CT image set is reviewed. Annotators with experience in image diagnosis verify and correct the position and range of the nodule coordinates and their consistency with the corresponding benign or malignant labels, eliminating samples with misaligned annotations, missing coordinates, or inconsistencies with pathological conclusions. At the same time, the labeled CT image set is screened for low-quality images, removing images that do not meet the requirements for subsequent processing due to severe artifacts, motion blur, missing tomography, incomplete scanning range, or abnormal contrast. Finally, samples with accurate nodule coordinate annotations and satisfactory image quality are retained to obtain a qualified labeled image set.
[0024] S105: Divide the qualified labeled image set into layers according to a preset ratio to obtain a labeled lung nodule sample set.
[0025] The stratified partitioning refers to dividing the sample subsets so that the proportion of benign and malignant samples in each subset is substantially consistent with the overall benign and malignant proportion in the qualified labeled image set, thus avoiding class distribution imbalance caused by random partitioning. In this step, the partitioning ratios corresponding to the training set, validation set, and test set are pre-set, and based on the benign / malignant label, all samples in the qualified labeled image set are stratified and partitioned non-overlapping according to the subject dimension, forming a training set for model parameter learning, a validation set for monitoring parameter tuning during training, and a test set for final performance evaluation. The training set, validation set, and test set together constitute the labeled lung nodule sample set, providing a well-organized and reliably labeled sample foundation for subsequent standardization processing and model training.
[0026] S20: Perform nodule block trimming, spatial resampling, and CT value standardization on the lung nodule sample set to obtain a standardized nodule block set.
[0027] Each low-dose chest CT image in the lung nodule sample set was cropped to focus the cropped nodule region on the lung nodule body; the cropped nodule blocks were then spatially resampled to unify the voxel scale under different devices and scanning parameters; and the CT values of the voxels were standardized to eliminate differences in grayscale distribution, finally obtaining a standardized nodule block set with uniform scale and consistent grayscale.
[0028] Combination Figure 4 As shown, step S20 specifically includes: S201: Using the marked nodule center coordinates as a reference, perform three-dimensional block trimming on the lung nodule sample set to obtain the original 3D nodule block set.
[0029] The nodule center coordinates refer to the coordinate points used to describe the spatial location of the lung nodule in the three-dimensional volume data, which have been verified in step S104. In this step, for each labeled CT image in the lung nodule sample set, the corresponding nodule center coordinates are used as the geometric reference. A bounding box is defined along the X, Y, and Z axes in the three-dimensional neighborhood of the nodule center coordinates according to a preset size. The voxel region covered by the bounding box is completely cut out from the labeled CT image to form a three-dimensional voxel block centered on the lung nodule, retaining the nodule body and a small amount of surrounding contextual tissue. The preset size is preferably 128×128×64 voxels, so as to minimize the interference of irrelevant tissues on subsequent recognition tasks while ensuring that the nodule region is completely covered. After performing the above three-dimensional block cropping on all labeled CT images in the lung nodule sample set, an original 3D nodule block set composed of several nodule voxel blocks is obtained.
[0030] S202: Extract the origin, voxel spacing and transformation matrix element information from the original 3D nodule block set to obtain a nodule block set with element information.
[0031] The origin refers to the physical spatial coordinates of the upper left front voxel corresponding to the 3D nodule block in the user coordinate system, used to describe the absolute position of the nodule block in the original scan space; the voxel spacing refers to the actual physical distance between the centers of adjacent voxels along the X, Y, and Z axes in the 3D nodule block, reflecting the spatial sampling density of the original scan; the transformation matrix refers to the spatial transformation matrix used to describe the geometric mapping relationship between the voxel index coordinate system and the user coordinate system, which can accurately convert the voxel index into physical coordinates in the user coordinate system. In this step, the above-mentioned origin, voxel spacing, and transformation matrix metadata are extracted from the DICOM header information to which each nodule block belongs in the original 3D nodule block set, and the extracted metadata is bound and stored with the corresponding nodule block to obtain a nodule block set with metadata, so that the subsequent resampling process can perform accurate spatial coordinate transformation and geometric alignment based on the above metadata.
[0032] S203: The set of nodule blocks with meta-information is subjected to isotropic voxel resampling and trilinear interpolation alignment to obtain a resampled nodule block set.
[0033] The isotropic voxel resampling refers to making the voxel spacing of the three-dimensional nodal block equal in the X, Y, and Z axes through resampling, so as to eliminate the geometric distortion caused by the difference in sampling density in each axis in the original scan; the trilinear interpolation refers to an interpolation method that estimates the gray value of the target sampling point by linearly weighting the gray values of the eight adjacent voxels around the target sampling point in the three-dimensional voxel grid according to the distance weight between the target sampling point and each adjacent voxel. In this step, based on the origin, voxel spacing, and transformation matrix element information corresponding to each nodule in the nodule block set with element information, the voxel spacing of each nodule block is uniformly adjusted to a preset isotropic voxel spacing, preferably 0.78×0.78×0.78mm. The resampled nodule blocks are then aligned to a preset standard size in terms of spatial dimensions, preferably 64×64×64 voxels. During the adjustment process, for each target voxel located on the new voxel grid after resampling, a trilinear interpolation method is used to perform weighted calculations from the gray values of adjacent voxels at the corresponding positions in the original nodule block to obtain the gray value of the target voxel, thereby obtaining a resampled nodule block set with consistent voxel spacing and uniform spatial resolution.
[0034] S204: Perform lung parenchyma masking and background voxel suppression on the resampled nodule block set to obtain the lung parenchyma nodule block set.
[0035] The lung parenchyma mask refers to a binary mask used to identify lung parenchyma regions within a 3D nodule block. Voxels representing lung parenchyma regions are assigned valid identifiers, while voxels representing background regions such as the chest wall, mediastinum, and external air are assigned invalid identifiers. Background voxel suppression involves setting the grayscale values of voxels within the non-lung parenchyma regions identified by the lung parenchyma mask to a uniform low response value, thereby reducing the interference of background tissue on lung nodule feature extraction. In this step, for each nodule block in the resampled nodule block set, lung parenchyma regions are extracted based on their grayscale distribution characteristics and the typical CT value range of lung parenchyma. Morphological closing operations are used to fill the cavities within the lung parenchyma formed by nodules, blood vessels, or tracheal walls, thereby generating a lung parenchyma mask of the same size as the nodule block. The lung parenchyma mask is then multiplied voxel-by-voxel by the corresponding nodule block, and background suppression is applied to voxels outside the mask, thus preserving voxel information containing lung nodules within the lung parenchyma and reducing the background voxel response outside the mask, resulting in a lung parenchyma nodule block set.
[0036] S205: Perform CT value window truncation and linear mapping normalization on the lung parenchymal nodule block set to obtain a standardized nodule block set.
[0037] The CT value refers to a physical quantity that measures the degree of attenuation of X-rays by a voxel, expressed in Hounsfield Units (HU). Different tissues have a relatively fixed range of CT value distribution in CT images. The CT value window truncation refers to limiting the CT value of a voxel to a preset upper and lower limit range. Voxel CT values exceeding this range are truncated to the upper and lower limits to suppress the adverse effects of extreme gray levels on model training. The linear mapping normalization refers to mapping the truncated CT values to a preset standard numerical range according to a linear relationship, so that the gray level distribution of different samples is within the same dimension and the same value range. In this step, for each lung parenchymal nodule in the lung parenchymal nodule set, the original CT values of each voxel are truncated within the range of [-1000, 600] HU, so that the CT values of voxels within this range can be completely preserved, and the CT values of voxels outside the range are truncated to -1000 HU or 600 HU respectively; then the truncated voxel CT values are mapped to the [0, 1] interval according to the linear mapping relationship, thereby completing the voxel grayscale normalization process, and obtaining a standardized nodule block set with uniform scale and consistent grayscale, which provides a standardized input basis for subsequent multi-axial random rotation and spherical random translation enhancement.
[0038] Furthermore, after obtaining the standardized nodule block set, the following optimization steps can be performed on the standardized nodule block set to improve the signal-to-noise ratio, localization accuracy, and feature separability of the standardized nodule blocks: S2051: Perform three-dimensional nonlocal mean denoising filtering on the standardized nodule block set to obtain a denoised standardized nodule block set.
[0039] The aforementioned three-dimensional nonlocal mean denoising filter refers to a denoising method that, for three-dimensional voxel data, uses the local neighborhood of the target voxel as a reference block, searches within a search window for several similar blocks that have a high degree of similarity in grayscale distribution to the reference block, and then performs a weighted average based on the grayscale value of the central voxel of each similar block according to similarity weights. This suppresses random noise and preserves the edges of tissue structures. Considering that the standardized nodule block set originates from low-dose chest CT images, voxel grayscale values usually retain significant quantum noise and electronic noise, which interfere with the fine structure of lung nodule edges and internal textures. In this step, a cubic neighborhood reference block and a search window are constructed along the X, Y, and Z axes for each standardized nodule block in the standardized nodule block set. The gray-level similarity between each candidate block and the reference block within the search window is calculated on a voxel-by-voxel basis. Based on the gray-level similarity, a weighted weight is generated to sum the gray-level values of the central voxel of the candidate block, thereby obtaining the denoised gray-level value of the target voxel. After completing the denoising calculation for all voxels, a denoised standardized nodule block set with clear structural edges and significantly reduced noise levels is obtained.
[0040] S2052: Based on the gray-scale centroid, refine and reposition the nodule center of the denoised and standardized nodule block set to obtain a center-corrected nodule block set.
[0041] The gray-scale centroid refers to the geometric center position obtained by weighting the spatial coordinates of each voxel within the effective gray-scale region of the 3D nodule block, using the gray-scale values of each voxel as weights. It reflects the actual mass distribution center of the nodule region in the 3D voxel space. Considering the manual annotation deviation of the nodule center coordinates reviewed in step S104, which may cause the geometric center of the cropped nodule block to deviate from the actual center of the nodule body, in this step, for each denoised and normalized nodule block in the denoised and normalized nodule block set, firstly, candidate nodule voxel regions are extracted within it based on the typical gray-scale range of the nodule. Then, the gray-scale centroid coordinates are calculated using the gray-scale values of each voxel within the candidate nodule voxel region as weights. Finally, the denoised and normalized nodule block is translated within a preset range using the gray-scale centroid coordinates as the new nodule center, so that the nodule body is distributed as centrally as possible in the 3D voxel space, thereby obtaining a center-corrected nodule block set with more accurate nodule center positioning.
[0042] S2053: Calculate the three-dimensional gradient magnitude of the centrally corrected nodule block set and use it as an auxiliary channel to splice it with the original channel to obtain a multi-channel nodule block set.
[0043] The three-dimensional gradient magnitude refers to the scalar value obtained by calculating the gray-level change rate along the X, Y, and Z axes of the three-dimensional voxel data and synthesizing the gray-level change rate along the three axes according to the Euclidean norm. It is used to describe the intensity of gray-level change at each voxel position and can highlight the edge contour and internal texture structure of the lung nodule. In this step, the three-dimensional difference operator is applied to each centrally corrected nodule block in the centrally corrected nodule block set along the X, Y, and Z axes to calculate a three-dimensional gradient magnitude volume of the same size as the centrally corrected nodule block. Then, the three-dimensional gradient magnitude volume is used as an auxiliary channel and spliced with the original gray-level channel of the centrally corrected nodule block along the channel dimension to form a multi-channel nodule block set that carries additional edge structure information in addition to the gray-level channel, thereby providing a richer input representation for the subsequent model to capture the edge and morphological features of the lung nodule.
[0044] S2054: Perform nodule quality scoring and outlier sample removal on the multi-channel nodule block set to obtain an optimized and standardized nodule block set.
[0045] The nodule block quality score refers to the comprehensive scoring of the overall quality of each multi-channel nodule block based on preset quality assessment indicators. These indicators include, but are not limited to, the effective voxel ratio of the nodule region, the coverage ratio of the lung parenchyma region, the noise residual level, and the edge gradient intensity distribution. The abnormal sample removal refers to removing multi-channel nodule blocks with quality scores below a preset threshold as abnormal samples that do not meet the requirements for subsequent training. In this step, a comprehensive quality score is calculated for each multi-channel nodule block in the multi-channel nodule block set based on the aforementioned quality assessment indicators. Multi-channel nodule blocks with comprehensive quality scores below the preset threshold are removed from the sample set, retaining only those with quality scores that meet the preset requirements. This results in an optimized and standardized nodule block set with accurate nodule localization, low noise levels, significant edge features, and reliable overall quality, providing a better input foundation for the multi-axial random rotation and spherical random translation enhancement in subsequent step S30.
[0046] S30: Perform multi-axial random rotation and spherical random translation enhancement on the standardized nodule block set to obtain the lung nodule training tensor set.
[0047] The standardized nodule blocks in the standardized nodule block set are rotated at random angles along multiple axes, and translated with random directions and distances within the spherical region to expand the pose distribution and spatial diversity of the lung nodule samples. The enhanced nodule blocks are then encapsulated into a lung nodule training tensor set that can be used for model training.
[0048] Combination Figure 5 As shown, step S30 specifically includes: S301: The standardized nodule block set is randomly rotated within a preset angle range along the X, Y, and Z axes to obtain a multi-axis rotating nodule block set.
[0049] The random rotation transformation refers to a geometric transformation method that uses the geometric center of the three-dimensional nodule block as the rotation center, independently samples rotation angles around the X, Y, and Z coordinate axes within a preset angle range, and constructs a three-dimensional rotation transformation matrix based on the sampled rotation angles to remap the voxel coordinates in the nodule block. This method is used to simulate the posture differences of lung nodules under different user postures and scanning orientations. In this step, for each standardized nodule block in the standardized nodule block set, random angles are independently sampled along the X, Y, and Z axes within a preset ±20 degree angle range. Based on the sampled three rotation angles, rotation components around the X, Y, and Z axes are constructed sequentially and synthesized into a total rotation transformation matrix. Then, the spatial coordinates of each voxel in the nodule block are remapped using the total rotation transformation matrix. For the target voxel located at the new position after mapping, the gray value is estimated from the original voxel gray value using trilinear interpolation, thereby completing the multi-axis random rotation of the standardized nodule block and obtaining a multi-axis rotated nodule block set with diverse posture distributions.
[0050] S302: The multi-axis rotating nodule block set is randomly oriented and translated within a preset radius spherical region to obtain a spherical translation nodule block set.
[0051] The spherical region refers to a three-dimensional spherical region defined by the geometric center of the three-dimensional nodule block as the sphere center and a preset distance as the radius. The random direction displacement sampling refers to randomly sampling the displacement direction and displacement amplitude within the spherical region. The sampled displacement direction covers any orientation in three-dimensional space, and the sampled displacement amplitude does not exceed the radius of the spherical region, so that the geometric center after displacement is still located within the spherical region. In this step, for each multi-axis rotating nodule block in the multi-axis rotating nodule block set, a spherical region is defined by its geometric center as the sphere center and a preset radius of 2.5 mm. A translation vector is randomly sampled within this spherical region. Then, the multi-axis rotating nodule block is translated as a whole along the X, Y, and Z axes according to the translation vector. The gray value of the target voxel located at the new position after translation is estimated using trilinear interpolation, thereby realizing the random direction displacement of the lung nodule in three-dimensional space and obtaining a spherical translation nodule block set with spatial diversity in geometric center distribution.
[0052] S303: Perform elastic deformation perturbation and voxel grayscale random jitter processing on the spherical translational nodal block set respectively to obtain the deformation perturbation nodal block set.
[0053] The elastic deformation perturbation refers to applying small-amplitude non-rigid perturbations driven by random displacement to the spatial position of each voxel in the three-dimensional nodule block, causing the nodule block to undergo local bending, stretching, or compression while maintaining the overall identifiable shape. This is used to simulate the slight morphological differences exhibited by lung nodules under different physiological states. The voxel grayscale random jitter refers to applying random perturbation to the grayscale value of each voxel within a preset amplitude range. This is used to simulate the grayscale fluctuations introduced by different scanning devices and acquisition parameters. In this step, for each spherical translation nodule in the set of spherical translation nodules, a three-dimensional random displacement field with the same size as the nodule is first generated, and the three-dimensional random displacement field is smoothed to ensure the spatial continuity of deformation. Then, based on the smoothed displacement field, the spatial position of each voxel in the nodule is slightly perturbed, and the gray value of the target voxel at the new position after perturbation is estimated by trilinear interpolation. Then, random jitter is applied to the gray value of each voxel in the nodule after elastic deformation perturbation within a preset range, thereby obtaining the set of deformed perturbed nodules.
[0054] S304: Perform boundary mirror filling and central region clipping restoration on the deformed disturbance nodule block set respectively to obtain a size-aligned nodule block set.
[0055] The boundary mirror filling refers to expanding the boundaries of each side of the 3D nodular block outward in a mirror-symmetrical manner along the corresponding directions using voxel distribution. This expands the nodular block in size, preventing invalid filling or edge artifacts caused by voxels near the boundaries being mapped outside the original nodular block range during the aforementioned random rotation, spherical translation, and elastic deformation perturbation. The central region trimming and restoration refers to cutting out a central region from the expanded nodular block according to a preset standard size, using the geometric center of the 3D nodular block as a reference, thus restoring it to the standard size required in the previous steps. In this step, for each deformable perturbation nodular block in the set of deformable perturbation nodular blocks, first, boundary mirror filling is performed along the X, Y, and Z axes on its six side boundaries to obtain an expanded nodular block. Then, the central region is trimmed according to a preset standard size using the geometric center of the expanded nodular block as a reference, thereby eliminating boundary artifacts and realigning the nodular blocks to a uniform size, resulting in a set of size-aligned nodular blocks with consistent dimensions and smooth boundaries.
[0056] S305: Add a single-channel dimension to the size-aligned nodule block set and encapsulate it with a tensor to obtain a lung nodule training tensor set.
[0057] The single-channel dimension refers to an additional dimension added to the three spatial dimensions of the three-dimensional voxel data to describe the number of image channels. Since low-dose chest CT images are grayscale images and each voxel corresponds to only one grayscale value, the added channel dimension has a length of one. Tensor encapsulation refers to structurally organizing the size-aligned nodule blocks according to the tensor format required by the deep learning model, giving them a uniform tensor shape in both channel and spatial dimensions that can be directly received by the model. In this step, for each size-aligned nodule block in the set, a single-channel dimension of length one is added before its three-dimensional spatial dimensions, and the nodule block after adding the single-channel dimension is encapsulated and stored in a tensor shape of (1, 64, 64, 64), where the first dimension corresponds to the added single-channel dimension, and the latter three dimensions correspond to the number of voxels along the X, Y, and Z spatial axes, respectively. After completing the addition of the single-channel dimension and tensor encapsulation for all size-aligned nodule blocks, a lung nodule training tensor set with a uniform format, which can be directly input into the subsequent Mamba state-space multi-directional scanning recognition network, is obtained.
[0058] S40: Based on the lung nodule training tensor set, construct the Mamba state space multi-directional scanning recognition network to obtain the initial 3D nodule recognition model.
[0059] Based on the input dimension corresponding to the lung nodule training tensor set and the lung nodule benign / malignant identification task, a recognition network based on the Mamba state space multi-directional scanning mechanism is constructed. The Mamba state space multi-directional scanning mechanism is used to model the long-range dependency relationship between lung nodule voxels with low computational complexity, and the network parameters are initialized and encapsulated to obtain an initial 3D nodule recognition model that has not yet been trained.
[0060] Combination Figure 6 As shown, step S40 specifically includes: S401: Input the lung nodule training tensor set into a two-layer three-dimensional convolution module for shallow feature encoding to obtain a first-scale feature map set.
[0061] The dual-layer 3D convolution module refers to a feature encoding module composed of two cascaded 3D convolutional layers. Each 3D convolutional layer performs sliding window convolution operations on the input tensor along the X, Y, and Z axes using a 3D convolutional kernel, and then uses batch normalization and nonlinear activation functions to normalize and nonlinearly transform the convolution results. The shallow feature encoding refers to extracting low-level feature representations, mainly voxel grayscale transitions, edge directions, and local textures from the input tensor through a limited number of convolutional operations at the network entry point, providing a preliminary abstract feature foundation for subsequent high-level modeling. In this step, each lung nodule training tensor in the lung nodule training tensor set is input into the two-layer three-dimensional convolution module in a preset batch. The tensors are then sequentially processed by a cascaded sliding window convolution operation through the first and second three-dimensional convolutional layers. The effective receptive field is expanded layer by layer by the two three-dimensional convolutional layers, and the number of feature channels is increased in the channel dimension. This encodes a shallow feature representation containing several feature channels from the single-channel grayscale information of the input tensor. After completing the above shallow feature encoding for all lung nodule training tensors in the lung nodule training tensor set, a first-scale feature map set is obtained that is basically aligned with the input tensor in the spatial dimension and has undergone dimensionality increase in the channel dimension.
[0062] S402: Perform three-dimensional block embedding and positional encoding injection on the first scale feature map set to obtain a three-dimensional feature token sequence.
[0063] The 3D block embedding refers to dividing the 3D feature map into several non-overlapping 3D feature blocks according to a preset cube size along the X, Y, and Z spatial dimensions, and mapping each 3D feature block to a feature vector of a preset dimension through linear projection transformation. The feature vector is called a token and is used to carry the feature information of each local region in the 3D feature map in the form of a sequence. The position encoding refers to attaching an encoding vector to each token in the sequence to describe its corresponding spatial position in the original 3D feature map, so that the subsequent Mamba state space multi-directional scan modeling process can retain the 3D spatial relative relationship between each token while modeling the sequence. The position encoding injection refers to fusing the position encoding with the corresponding token in a dimensional alignment manner. In this step, each first-scale feature map in the first-scale feature map set is divided into three-dimensional blocks along the X, Y, and Z axes according to a predetermined block size. The resulting three-dimensional feature blocks are then mapped to tokens of a preset dimension through linear projection transformation. Next, a corresponding position code is generated based on the spatial position of the three-dimensional feature block corresponding to each token in the first-scale feature map. The position code is then fused and injected with the corresponding token, thereby converting the spatial features originally carried in the form of three-dimensional feature maps into a sequence of three-dimensional feature tokens that carry three-dimensional spatial position information.
[0064] S403: Use the three-dimensional feature token sequence to perform Mamba state space multi-directional scan modeling to obtain a multi-directional context feature set.
[0065] The Mamba state space refers to a sequence modeling framework based on a state space model, which introduces a selective scanning mechanism to allow state transition parameters to adaptively adjust according to input features. It can effectively model the dependencies between tokens that are far apart in the sequence while maintaining linear computational complexity. The multi-directional scanning refers to expanding and traversing the three-dimensional feature token sequence according to multiple preset scanning orders. The multiple scanning orders cover at least forward and reverse traversal along the X, Y, and Z axes, so that each token can receive context information from other tokens in the sequence in different directions. The context features refer to the feature representations carried by each token after Mamba state space multi-directional scanning modeling, which incorporate long-range dependency information from other tokens in the sequence. In this step, the three-dimensional feature token sequence is expanded according to various preset scanning orders, and the expanded sequences obtained under each scanning order are sequentially input into the Mamba state space model for selective scanning modeling. The Mamba state space model models the long-range dependencies between tokens under each scanning order and outputs the context features in the corresponding direction. Then, the context features corresponding to each scanning order are fused according to the token position to obtain a multi-directional context feature set that can simultaneously reflect the long-range dependencies in multiple directions in three-dimensional space.
[0066] S404: Perform hierarchical downsampling fusion and residual connection processing on the multi-directional context feature set to obtain a multi-scale nodule feature set.
[0067] The hierarchical downsampling refers to the progressive spatial compression of features along the X, Y, and Z spatial dimensions with a preset step size, thereby progressively reducing the spatial resolution and progressively increasing the dimensionality of the features in the channel dimension, thus forming feature representations with different spatial resolutions and receptive fields at different levels. The hierarchical downsampling fusion refers to the channel-level merging of multi-directional contextual features from different scanning directions during the hierarchical downsampling process, enabling the features of each level to simultaneously fuse multi-directional long-range dependency information. The residual connection refers to the introduction of a direct mapping path between adjacent levels, adding the features of the previous level to the features of the next level element-wise using an identity mapping method, in order to alleviate the gradient decay problem during the training process of deep networks. In this step, the multi-directional context feature set is sequentially input into several hierarchical downsampling units according to the hierarchical structure. Each hierarchical downsampling unit performs channel-level fusion of the multi-directional context features while compressing the spatial dimension and increasing the channel dimension according to a preset step size, thereby obtaining multi-level features from shallow to deep, with progressively decreasing spatial resolution and progressively increasing semantic abstraction. Then, residual connection paths are introduced between adjacent levels, and the shallow features are added element-wise with the deep features after necessary dimensional alignment, thereby obtaining a multi-scale nodal feature set that retains both shallow local details and covers deep global semantics.
[0068] S405: Perform global pooling and classification header parameter initialization and encapsulation on the multi-scale nodule feature set respectively to obtain the initial 3D nodule recognition model.
[0069] The global pooling refers to the overall aggregation of the deepest-level multi-scale nodule features along the three spatial dimensions of X, Y, and Z, compressing them into a single-point feature of length one in the spatial dimension. The single-point feature retains the same number of feature channels in the channel dimension, and is used to carry the global feature representation of the entire lung nodule region in the form of a single feature vector. The classification head refers to several fully connected layers connected after the global feature vector obtained by the global pooling, which are used to map the global feature vector to the prediction score corresponding to the number of benign and malignant categories. The classification head parameter initialization refers to assigning initial values to the weights and bias parameters of each fully connected layer of the classification head according to a preset initialization strategy before the network training begins, so as to ensure stable convergence of the training process. In this step, global pooling is performed on the multi-scale nodule features located at the deepest level in the multi-scale nodule feature set, aggregating them in three-dimensional space into a global feature vector carrying the global features of lung nodules. Then, a classification head with a preset structure is connected after the global feature vector, and the parameters of each fully connected layer of the classification head are initialized according to a preset initialization strategy. Then, the two-layer three-dimensional convolution module, the three-dimensional block embedding and position encoding injection unit, the Mamba state space multi-directional scanning modeling unit, the hierarchical downsampling fusion and residual connection unit, and the classification head are uniformly encapsulated into a complete network structure in sequence, thereby obtaining an initial 3D nodule recognition model that has not yet undergone iterative training, providing a model structure foundation for the iterative training in the subsequent step S50.
[0070] S50: Input the lung nodule training tensor set into the initial 3D nodule recognition model for iterative training to obtain the target 3D nodule recognition model.
[0071] The training tensor set of lung nodules is input into the initial 3D nodule recognition model in batches for forward inference, and the network parameters are iteratively updated repeatedly based on the deviation between the prediction results and the benign and malignant labels. After the training process converges or the preset iteration termination condition is met, the solidified model is used as the target 3D nodule recognition model that can be used for actual recognition.
[0072] Combination Figure 7 As shown, step S50 specifically includes: S501: The training tensor set of lung nodules is sampled and encapsulated into queues according to the proportion of benign and malignant categories to obtain a training batch tensor stream.
[0073] The ratio of benign to malignant categories refers to the numerical distribution ratio of benign and malignant lung nodule training tensors in the lung nodule training tensor set. Given that the sample size of benign and malignant lung nodules often differs significantly in clinical screening scenarios, if random sampling is performed directly according to the original sample distribution, the proportion of samples of each category in each training batch may be severely unbalanced, causing the model to favor the category with a larger sample size during training. The queue encapsulation refers to organizing the sampled lung nodule training tensors into a first-in-first-out batch queue according to a preset batch capacity, so that the tensors of each batch can be delivered to the model for training in a predetermined order. In this step, the sampling quantity of each category in each training batch is first determined based on the ratio of benign to malignant categories in the lung nodule training tensor set. Then, benign and malignant lung nodule training tensors are randomly sampled according to the determined sampling quantity to ensure that the ratio of benign to malignant categories in each training batch meets the preset requirements. Next, the sampled lung nodule training tensors are encapsulated in a queue according to the preset batch capacity to ensure that the lung nodule training tensors in each training batch are consistent in the channel dimension and the three-dimensional spatial dimension and arranged in batch order, thereby obtaining a training batch tensor stream that can be continuously received by the model.
[0074] S502: Input the training batch tensor stream into the initial 3D nodule recognition model for forward propagation inference to obtain the batch prediction probability set.
[0075] The forward propagation inference refers to the process of sequentially passing the input tensor through the two-layer three-dimensional convolution module, the three-dimensional block embedding and positional encoding injection unit, the Mamba state space multi-directional scanning modeling unit, the hierarchical downsampling fusion and residual connection unit, and the classification head along the network structure from front to back. This allows the input tensor to undergo feature transformation layer by layer in each operation unit according to the current parameter values, and finally outputs the prediction result corresponding to the benign or malignant category. The batch prediction probability refers to the probability value obtained after normalizing the prediction scores corresponding to the number of benign or malignant categories output by the classification head after the forward propagation inference process, where the sum of the prediction confidence of each category is one. In this step, each training batch in the tensor stream of the training batch is input into the initial 3D nodule recognition model according to the training batch order. The initial 3D nodule recognition model performs forward propagation inference on all lung nodule training tensors in the training batch, and sequentially completes shallow feature encoding, 3D block embedding and positional encoding injection, Mamba state space multi-directional scanning modeling, hierarchical downsampling fusion and residual connection, global pooling and classification head mapping through each operation unit. The predicted score output by the classification head is normalized to obtain the benign and malignant prediction probabilities corresponding to each lung nodule training tensor in the training batch. After performing the above forward propagation inference on all training batches in the tensor stream of the training batch, a batch prediction probability set corresponding to each training batch is obtained.
[0076] S503: Calculate the weighted cross-entropy loss based on the batch prediction probability set and the preset benign / malignant labels to obtain the backpropagation gradient set.
[0077] The benign / malignant label refers to the standard information used to characterize the true benign / malignant category of each lung nodule training tensor in step S103, which is bound to each lung nodule training tensor in the lung nodule training tensor set. The weighted cross-entropy loss is a loss function constructed by introducing differentiated weight coefficients for different benign / malignant categories based on the standard cross-entropy loss. The weight coefficients are set in reverse according to the relative number of each category in the sample set, so that the category with fewer samples gets a relatively larger weight in the loss calculation, thereby suppressing the adverse effect of class imbalance on model training. The backpropagation refers to the process of calculating the partial derivatives of each layer parameter with respect to the loss by applying the chain rule layer by layer from the end of the network to the input end based on the weighted cross-entropy loss. The backpropagation gradient refers to the gradient information obtained through backpropagation, which is used to guide the update direction and magnitude of each layer parameter. In this step, firstly, weight coefficients are determined for each benign and malignant category based on the proportion of benign and malignant categories in the lung nodule training tensor set. Then, based on the weight coefficients, a weighted cross-entropy loss is calculated between the batch prediction probability corresponding to the lung nodule training tensor in each training batch and the preset benign and malignant labels, thereby obtaining the loss value of that training batch. Then, starting from the loss value, backpropagation is performed sequentially along each operation unit of the initial 3D nodule recognition model, and the partial derivatives of the trainable parameters in each operation unit with respect to the loss are calculated layer by layer, thereby obtaining the backpropagation gradient set corresponding to each training batch.
[0078] S504: Adaptively update the parameters of the initial 3D nodule identification model using the backpropagation gradient set to obtain an iteratively updated model.
[0079] The adaptive parameter update refers to adjusting parameters not only based on the backpropagation gradient obtained from the current training batch, but also adaptively adjusting the update step size for each parameter based on the first and second moment estimates of the historical gradients of each parameter. This allows for a smaller update step size for parameters with large historical gradient fluctuations and a larger update step size for parameters with relatively stable historical gradients, thereby improving the stability and convergence speed of the training process. The iterative update model refers to a network model obtained by repeatedly applying the above adaptive parameter update to multiple training batches based on the initial 3D nodule recognition model, where the parameter values are adjusted accordingly in each iteration. In this step, based on the backpropagation gradients corresponding to each training batch in the backpropagation gradient set, the trainable parameters of each operational unit in the initial 3D nodule recognition model are adaptively updated sequentially according to the training batch order. In each update process, the adaptive update step size is calculated by combining the current backpropagation gradient with the historical first and second moments corresponding to each parameter, and the trainable parameters are slightly adjusted accordingly. As the training batches continue to advance and the training rounds accumulate, the parameter values of the initial 3D nodule recognition model gradually approach the direction that minimizes the weighted cross-entropy loss, thereby obtaining an iteratively updated model that continuously evolves during the training process.
[0080] S505: The optimal weights are solidified in the iterative update model to obtain the target 3D nodule recognition model.
[0081] The optimal weight refers to the network parameter value that achieves the best evaluation result on the validation set corresponding to the lung nodule training tensor set during the training and evolution of the iterative update model, based on the preset model performance evaluation index. Weight solidification refers to fixing and saving the network parameter value corresponding to the optimal weight in a state where further training is not possible, thus freezing the iterative update model into a stable model that can be directly used for inference and recognition under this parameter value. In this step, during the training and evolution of the iterative update model, its performance evaluation index on the validation set is calculated and recorded according to a preset monitoring cycle, and the network parameters and performance evaluation index corresponding to each monitoring moment are saved together. After the training process meets the preset iteration termination condition, the network parameter value corresponding to the best performance evaluation index is selected from all recorded monitoring moments and determined as the optimal weight. Then, the optimal weight is loaded into the iterative update model, and its parameter value is solidified and saved, freezing each computational unit into a stable structure under this parameter value, thereby obtaining a target 3D nodule recognition model that can be directly used for inference and recognition of the CT image to be recognized, providing a trained model foundation for the inference and recognition in the subsequent step S60.
[0082] S60: Input the CT image to be identified into the target 3D nodule recognition model for recognition, and output the lung nodule recognition result.
[0083] The low-dose chest CT image to be identified is provided as inference input to the target 3D nodule recognition model, which then identifies the benign or malignant lung nodules and outputs the lung nodule recognition result containing benign or malignant category information, thereby completing the automated lung nodule recognition of the CT image to be identified.
[0084] Combination Figure 8 As shown, step S60 specifically includes: S601: Receives the low-dose chest CT image to be identified and performs nodule center localization and three-dimensional target area cropping to obtain the nodule mass to be identified.
[0085] The low-dose chest CT image to be identified refers to the low-dose chest CT image data that needs to be identified as benign or malignant by the target 3D nodule recognition model in the actual application stage. Its data source and acquisition method are consistent with the original DICOM image sequence of low-dose chest CT obtained in step S101. The nodule center localization refers to the process of determining the center coordinates of the lung nodule in the three-dimensional voxel space after receiving the low-dose chest CT image to be identified. The determined center coordinates are used as the geometric reference for the three-dimensional target area clipping in subsequent processing. The three-dimensional target area clipping refers to the process of defining a three-dimensional bounding box along the X, Y, and Z axes according to a preset size based on the center coordinates obtained by the nodule center localization and completely cutting out the voxel area covered by the three-dimensional bounding box from the low-dose chest CT image to be identified. Its processing method is consistent with the three-dimensional block clipping performed on the lung nodule sample set in step S201 in terms of geometric reference and clipping size. In this step, the low-dose chest CT image to be identified is first received and the center of the nodule is located to obtain the center coordinates of the lung nodule in the image to be identified. Then, the three-dimensional target area is cropped along the X, Y, and Z axes according to a preset size using the center coordinates as a geometric reference. The voxel area covered by the three-dimensional target area is completely cut out from the low-dose chest CT image to be identified, thereby obtaining the nodule block to be identified with the lung nodule as the center and retaining the nodule body and a small amount of surrounding context tissue.
[0086] S602: Perform voxel resampling, CT value normalization, and tensor dimension encapsulation on the nodule block to be identified to obtain the input tensor to be identified.
[0087] The voxel resampling refers to the process of uniformly adjusting the voxel spacing of the nodule block to be identified along the X, Y, and Z axes to be consistent with the isotropic voxel spacing determined in step S203, so that the voxel spacing of the nodule block to be identified is consistent with the standardized nodule block corresponding to the lung nodule training tensor set. The CT value normalization refers to performing the same CT value window truncation and linear mapping normalization process as in step S205 on the CT values of each voxel in the nodule block to be identified, so that the grayscale distribution of each voxel in the nodule block to be identified is consistent with the standardized nodule block corresponding to the lung nodule training tensor set in terms of dimensions and value range. The tensor dimension encapsulation refers to the process of adding a single-channel dimension of length one before the three-dimensional spatial dimension of the nodule block to be identified and organizing it in a structured manner according to the tensor shape required by the target 3D nodule identification model, so that the tensor shape of the nodule block to be identified is consistent with the lung nodule training tensor obtained in step S305. In this step, the nodule block to be identified is subjected to voxel resampling, CT value normalization and tensor dimension encapsulation in sequence, so that its voxel spacing, gray level distribution and tensor shape are consistent with the lung nodule training tensor corresponding to the lung nodule training tensor set, thereby obtaining the input tensor to be identified that can be directly received by the target 3D nodule recognition model.
[0088] S603: The input tensor to be identified is subjected to enhanced flipping and batch stacking processing to obtain a multi-view test tensor set.
[0089] The enhanced flipping refers to the process of mirror flipping the input tensor to be identified along one or more of the X, Y, and Z axes according to a preset flipping combination. The preset flipping combination includes various flipping methods such as no flipping, single-axis flipping, and multi-axis combination flipping, enabling a single input tensor to generate several flipped copies with different spatial orientations. The viewing angle refers to the spatial observation orientation corresponding to each flipped copy, determined by the flipping combination; each flipped copy corresponds to one viewing angle. The batch stacking refers to the process of sequentially stacking the aforementioned flipped copies along the batch dimension and organizing them into a unified multi-view tensor set. In this step, the input tensor to be identified is first mirror flipped according to the preset flipping combination, resulting in several flipped copies with different spatial orientations. These flipped copies, together with the original input tensor to be identified, constitute a multi-view sample set corresponding to the nodule to be identified. Then, the multi-view sample set is batch stacked along the batch dimension, ensuring that the samples at each viewpoint are consistent in the channel dimension and three-dimensional spatial dimension and arranged in viewpoint order, thereby obtaining a multi-view test tensor set capable of simultaneously carrying lung nodule feature information from multiple views.
[0090] S604: Input the multi-view test tensor set into the target 3D nodule recognition model for forward inference and multi-view probability fusion to obtain the benign or malignant prediction probability.
[0091] The forward inference refers to the process of sequentially passing the tensors of each viewpoint in the multi-view test tensor set through the network structure of the target 3D nodule recognition model from front to back through the two-layer three-dimensional convolution module, the three-dimensional block embedding and position encoding injection unit, the Mamba state space multi-directional scanning modeling unit, the hierarchical downsampling fusion and residual connection unit, and the classification head. This allows each viewpoint tensor to undergo feature transformation layer by layer in each operation unit according to the fixed optimal weights, and the prediction score corresponding to the benign or malignant category is output by the classification head and normalized to obtain the viewpoint prediction probability. The multi-view probability fusion refers to the process of summarizing and aggregating the viewpoint prediction probabilities corresponding to each viewpoint in the multi-view test tensor set according to a preset fusion strategy. The preset fusion strategy includes, but is not limited to, taking the arithmetic mean according to the viewpoint dimension, so that the fused probability can simultaneously reflect the benign or malignant prediction information of lung nodules under multiple views, and to a certain extent suppress the prediction fluctuations introduced by the orientation bias under a single viewpoint. In this step, the multi-view test tensor set is first input into the target 3D nodule recognition model, and forward inference is performed on each view tensor. The target 3D nodule recognition model outputs the corresponding view prediction probability for each view tensor. Then, the view prediction probabilities corresponding to each view in the multi-view test tensor set are fused according to a preset fusion strategy to obtain the benign or malignant prediction probability corresponding to the nodule block to be identified.
[0092] S605: Based on the predicted probabilities of benign and malignant lung nodules, a category threshold is determined and the nodule location information is backfilled and encapsulated to obtain the lung nodule identification result.
[0093] The category threshold determination refers to comparing the predicted probabilities of benign and malignant transformation with a preset determination threshold, and determining the benign or malignant category of the lung nodule to be identified based on the comparison result. When the probability of the malignant category in the predicted probabilities of benign and malignant transformation is not lower than the preset determination threshold, the lung nodule is determined to be malignant; otherwise, it is determined to be benign. The nodule location information refers to the central coordinates of the lung nodule in the low-dose chest CT image to be identified and the spatial range information occupied by the cropped three-dimensional target area in the low-dose chest CT image to be identified, as determined in step S601. The backfilling and encapsulation refers to the process of structuring the benign or malignant category obtained by the category threshold determination and the nodule location information according to a preset structure, and encapsulating them together with the predicted probabilities of benign and malignant transformation into an output and displayable identification result record. In this step, the benign and malignant predicted probabilities are first determined by a preset threshold to determine the benign and malignant category corresponding to the nodule to be identified. Then, the determined benign and malignant category, the corresponding benign and malignant predicted probabilities, and the nodule location information determined in step S601 are backfilled and encapsulated according to a preset structure to obtain the lung nodule identification result containing the spatial location of the lung nodule, the benign and malignant category, and the corresponding predicted probability. Thus, the automated benign and malignant identification of lung nodules in low-dose chest CT images to be identified is completed.
[0094] As can be seen, this application, by performing tissue library construction, nodule clipping, spatial resampling, and CT value standardization on the acquired low-dose chest CT images, can provide a standardized nodule block set with uniform scale and consistent grayscale for subsequent model training. This effectively eliminates the differences in voxel spacing and grayscale distribution introduced by different scanning devices and acquisition parameters, ensuring the standardization and consistency of input data from the source. By enhancing the standardized nodule block set with multi-axial random rotation and spherical random translation, the pose distribution and spatial diversity of lung nodule samples can be significantly expanded without introducing additional annotation costs. This makes the resulting lung nodule training tensor set more adaptable to changes in user pose and spatial displacement of lung nodules, thereby improving the model's generalization ability and robustness.
[0095] This application constructs a Mamba state-space multi-directional scanning recognition network based on the lung nodule training tensor set and uses it to build an initial 3D nodule recognition model. By leveraging the Mamba state-space multi-directional scanning mechanism, it effectively models the long-range dependencies between lung nodule voxels while maintaining low computational complexity. This overcomes the inherent limitations of traditional 3D convolutional neural networks, which are limited by the local receptive field of convolution operators and struggle to express long-range dependencies. The constructed recognition network, while controlling computational load and memory usage, possesses the ability to fully capture the overall 3D structure and morphological evolution characteristics of lung nodules. Furthermore, by... The lung nodule training tensor set is input into the initial 3D nodule recognition model for iterative training, enabling the obtained target 3D nodule recognition model to have high prediction accuracy and stability in the task of identifying benign and malignant lung nodules. Furthermore, by inputting the CT image to be identified into the target 3D nodule recognition model for identification and outputting the lung nodule recognition result, the automatic identification of benign and malignant lung nodules in low-dose chest CT images is realized. This effectively reduces the reliance on manual interpretation by physicians and alleviates the problems of strong subjectivity and poor repeatability of manual interpretation, providing efficient and reliable technical support for large-scale early lung cancer screening tasks.
[0096] To further verify the beneficial effects of the embodiments of the present invention, this embodiment performs performance verification on multiple publicly available lung nodule datasets, and compares it with existing representative recognition models such as ResNet50, nnMamba, and ViT in terms of classification accuracy (ACC) and area under the curve (AUC). The comparison results are shown in Table 1 below: Model ACC AUC ResNet50 0.8629 0.8632 nnMamba 0.8889 0.8945 ViT 0.6959 0.8465 This application 0.9189 0.9143 Table 1 The comparison results above show that, compared with traditional 3D convolutional neural networks represented by ResNet50 and Transformer architecture models represented by ViT, the target 3D nodule recognition model provided in this application has achieved significantly superior performance in both classification accuracy and area under the curve (AUC), with a classification accuracy of 0.9189 and an AUC of 0.9143. It also shows further improvement over nnMamba in both metrics, fully verifying the superior performance of this application in the task of identifying benign and malignant lung nodules in low-dose chest CT images. At the same time, the target 3D nodule recognition model provided in this application can still maintain stable recognition performance in scenarios with limited sample size, demonstrating better robustness and generalization ability. This further confirms the practical effect of this application in solving the technical problem that existing lung nodule recognition methods are unable to effectively model long-range dependencies, resulting in insufficient accuracy in benign and malignant nodule recognition.
[0097] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0098] In one embodiment, a Mamba-based 3D lung nodule recognition device is provided, which corresponds one-to-one with the Mamba-based 3D lung nodule recognition method described in the above embodiments. For example... Figure 9 As shown, the Mamba-based 3D lung nodule identification device includes: a sample acquisition module 100, a sample processing module 200, a sample enhancement module 300, a model building module 400, a model training module 500, and a result recognition module 600.
[0099] Detailed descriptions of each functional module are as follows: The sample acquisition module 100 is used to acquire low-dose chest CT images and perform tissue library construction with corresponding benign and malignant identification information to obtain an labeled lung nodule sample set. The sample processing module 200 is used to perform nodule block trimming, spatial resampling and CT value standardization on the lung nodule sample set to obtain a standardized nodule block set. The sample augmentation module 300 is used to perform multi-axial random rotation and spherical random translation augmentation on the standardized nodule block set to obtain the lung nodule training tensor set. The model building module 400 is used to build a Mamba state-space multi-directional scanning recognition network based on the lung nodule training tensor set to obtain an initial 3D nodule recognition model. The model training module 500 is used to input the lung nodule training tensor set into the initial 3D nodule recognition model for iterative training to obtain the target 3D nodule recognition model. The result recognition module 600 is used to input the CT image to be recognized into the target 3D nodule recognition model for recognition and output the lung nodule recognition result.
[0100] In one embodiment, the sample acquisition module 100 is specifically used for: Acquire raw DICOM images of low-dose chest CT and integrate them to obtain the raw CT image set; The original CT image set is subjected to DICOM head analysis and volume data 3D reconstruction to obtain the CT image set to be annotated; The CT image set to be labeled is associated with the benign and malignant information in the preset pathology report to build a library, thereby obtaining a labeled CT image set; The labeled CT image set was reviewed for nodule coordinate annotation and low-quality images were screened out to obtain a qualified labeled image set. The qualified labeled image set is divided into layers according to a preset ratio to obtain a labeled lung nodule sample set.
[0101] In one embodiment, the sample processing module 200 is specifically used for: The lung nodule sample set was cropped into three-dimensional blocks based on the marked nodule center coordinates to obtain the original 3D nodule block set; The origin, voxel spacing, and transformation matrix element information of the original 3D nodule block set are extracted respectively to obtain a nodule block set with element information; The set of nodal blocks with meta-information is aligned by isotropic voxel resampling and trilinear interpolation to obtain a resampled nodal block set. The lung parenchyma nodule set was subjected to lung parenchyma mask segmentation and background voxel suppression respectively to obtain the lung parenchyma nodule set; The CT value window was truncated and the linear mapping was normalized on the lung parenchyma nodule set to obtain a standardized nodule set.
[0102] In one embodiment, the sample enhancement module 300 is specifically used for: The standardized nodule block set is randomly rotated within a preset angle range along the X, Y, and Z axes to obtain a multi-axis rotating nodule block set. By randomly directional displacement sampling and translation of the multi-axis rotating nodular block set within a preset radius spherical region, a spherical translation nodular block set is obtained; Elastic deformation perturbation and voxel grayscale random jitter processing are applied to the spherical translational nodal block set respectively to obtain the deformation perturbation nodal block set; The deformed perturbation nodule block set is respectively subjected to boundary mirror filling and central region clipping and restoration to obtain a size-aligned nodule block set; A single-channel dimension is added to the size-aligned nodule block set and tensor encapsulation is performed to obtain the lung nodule training tensor set.
[0103] In one embodiment, the model building module 400 is specifically used for: The training tensor set of lung nodules is input into a two-layer three-dimensional convolutional module for shallow feature encoding to obtain a first-scale feature map set. The first scale feature map set is subjected to 3D block embedding and positional encoding injection to obtain a 3D feature token sequence; The three-dimensional feature token sequence is used to perform Mamba state space multi-directional scanning modeling to obtain a multi-directional context feature set. The multi-dimensional context feature set is subjected to hierarchical downsampling fusion and residual connection processing respectively to obtain a multi-scale nodule feature set; The multi-scale nodule feature set is subjected to global pooling and classification header parameter initialization encapsulation to obtain an initial 3D nodule recognition model.
[0104] In one embodiment, the model training module 500 is specifically used for: The lung nodule training tensor set is sampled and queued according to the proportion of benign and malignant categories to obtain a training batch tensor stream; The training batch tensor stream is input into the initial 3D nodule recognition model for forward propagation inference to obtain the batch prediction probability set; Based on the batch prediction probability set and the preset benign / malignant labels, a weighted cross-entropy loss is calculated to obtain the backpropagation gradient set; The initial 3D nodule identification model is adaptively updated using the backpropagation gradient set to obtain an iteratively updated model. The optimal weights are solidified in the iterative update model to obtain the target 3D nodule recognition model.
[0105] In one embodiment, the result recognition module 600 is specifically used for: Receive low-dose chest CT images to be identified and perform nodule center localization and three-dimensional target area cropping to obtain the nodule mass to be identified; The nodule block to be identified is subjected to voxel resampling, CT value normalization and tensor dimension encapsulation to obtain the input tensor to be identified; The input tensor to be identified is enhanced by flipping and batch stacking to obtain a multi-view test tensor set. The multi-view test tensor set is input into the target 3D nodule recognition model for forward inference and multi-view probability fusion to obtain the benign and malignant prediction probabilities. Based on the predicted probabilities of benign and malignant lesions, a category threshold is determined and nodule location information is backfilled and encapsulated to obtain the lung nodule identification result.
[0106] Specific limitations regarding the Mamba-based 3D lung nodule identification device can be found in the limitations of the Mamba-based 3D lung nodule identification method described above, and will not be repeated here. Each module in the aforementioned Mamba-based 3D lung nodule identification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0107] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side method for Mamba-based 3D lung nodule identification.
[0108] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 11 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a Mamba-based 3D lung nodule identification method.
[0109] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed, can perform the steps provided in the above embodiments.
[0110] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.
[0111] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0113] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.
[0114] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A 3D lung nodule identification method based on Mamba, characterized in that, include: Low-dose chest CT images were acquired and tissue libraries were constructed by comparing them with corresponding benign and malignant identification information to obtain a set of labeled lung nodules. The lung nodule sample set was subjected to nodule block trimming, spatial resampling, and CT value standardization to obtain a standardized nodule block set. The standardized nodule block set is enhanced by multi-axial random rotation and spherical random translation to obtain the lung nodule training tensor set; Based on the lung nodule training tensor set, a Mamba state space multi-directional scanning recognition network was built to obtain an initial 3D nodule recognition model. The lung nodule training tensor set is input into the initial 3D nodule recognition model for iterative training to obtain the target 3D nodule recognition model. The CT image to be identified is input into the target 3D nodule recognition model for recognition, and the lung nodule recognition result is output. The step of constructing a Mamba state-space multi-directional scanning recognition network based on the lung nodule training tensor set to obtain an initial 3D nodule recognition model includes: inputting the lung nodule training tensor set into a two-layer three-dimensional convolutional module for shallow feature encoding to obtain a first-scale feature map set; performing three-dimensional block embedding and positional encoding injection on the first-scale feature map set to obtain a three-dimensional feature token sequence; using the three-dimensional feature token sequence to perform Mamba state-space multi-directional scanning modeling to obtain a multi-directional context feature set; performing hierarchical downsampling fusion and residual connection processing on the multi-directional context feature set to obtain a multi-scale nodule feature set; and performing global pooling and classification head parameter initialization encapsulation on the multi-scale nodule feature set to obtain the initial 3D nodule recognition model. The process of inputting the CT image to be identified into the target 3D nodule recognition model for identification and outputting lung nodule recognition results includes: receiving a low-dose chest CT image to be identified and performing nodule center localization and 3D target region cropping to obtain a nodule block to be identified; performing voxel resampling, CT value normalization, and tensor dimension encapsulation on the nodule block to be identified to obtain an input tensor to be identified; performing enhanced flipping and batch stacking processing on the input tensor to be identified to obtain a multi-view test tensor set; inputting the multi-view test tensor set into the target 3D nodule recognition model for forward inference and multi-view probability fusion to obtain benign and malignant prediction probabilities; and performing category threshold determination and nodule localization information backfilling and encapsulation based on the benign and malignant prediction probabilities to obtain lung nodule recognition results.
2. The 3D lung nodule identification method based on Mamba according to claim 1, characterized in that, The process involves acquiring low-dose chest CT images and constructing a tissue library based on corresponding benign / malignant identification information to obtain an annotated lung nodule sample set, including: Acquire raw DICOM images of low-dose chest CT and integrate them to obtain the raw CT image set; The original CT image set is subjected to DICOM head analysis and volume data 3D reconstruction to obtain the CT image set to be annotated; The CT image set to be labeled is associated with the benign and malignant information in the preset pathology report to build a library, thereby obtaining a labeled CT image set; The labeled CT image set was reviewed for nodule coordinate annotation and low-quality images were screened out to obtain a qualified labeled image set. The qualified labeled image set is divided into layers according to a preset ratio to obtain a labeled lung nodule sample set.
3. The 3D lung nodule identification method based on Mamba according to claim 1, characterized in that, The process of cropping, spatially resampling, and standardizing CT values in the lung nodule sample set to obtain a standardized nodule set includes: The lung nodule sample set was cropped into three-dimensional blocks based on the marked nodule center coordinates to obtain the original 3D nodule block set; The origin, voxel spacing, and transformation matrix element information of the original 3D nodule block set are extracted respectively to obtain a nodule block set with element information; The set of nodal blocks with meta-information is aligned by isotropic voxel resampling and trilinear interpolation to obtain a resampled nodal block set. The lung parenchyma nodule set was subjected to lung parenchyma mask segmentation and background voxel suppression respectively to obtain the lung parenchyma nodule set; The CT value window was truncated and the linear mapping was normalized on the lung parenchyma nodule set to obtain a standardized nodule set.
4. The 3D lung nodule identification method based on Mamba according to claim 1, characterized in that, The process of enhancing the standardized nodule block set with multi-axial random rotation and spherical random translation to obtain the lung nodule training tensor set includes: The standardized nodule block set is randomly rotated within a preset angle range along the X, Y, and Z axes to obtain a multi-axis rotating nodule block set. By randomly directional displacement sampling and translation of the multi-axis rotating nodular block set within a preset radius spherical region, a spherical translation nodular block set is obtained; Elastic deformation perturbation and voxel grayscale random jitter processing are applied to the spherical translational nodal block set respectively to obtain the deformation perturbation nodal block set; The deformed perturbation nodule block set is respectively subjected to boundary mirror filling and central region clipping and restoration to obtain a size-aligned nodule block set; A single-channel dimension is added to the size-aligned nodule block set and tensor encapsulation is performed to obtain the lung nodule training tensor set.
5. The 3D lung nodule identification method based on Mamba according to claim 1, characterized in that, The step of inputting the lung nodule training tensor set into the initial 3D nodule recognition model for iterative training to obtain the target 3D nodule recognition model includes: The lung nodule training tensor set is sampled and queued according to the proportion of benign and malignant categories to obtain a training batch tensor stream; The training batch tensor stream is input into the initial 3D nodule recognition model for forward propagation inference to obtain the batch prediction probability set; Based on the batch prediction probability set and the preset benign / malignant labels, a weighted cross-entropy loss is calculated to obtain the backpropagation gradient set; The initial 3D nodule identification model is adaptively updated using the backpropagation gradient set to obtain an iteratively updated model. The optimal weights are solidified in the iterative update model to obtain the target 3D nodule recognition model.
6. A 3D lung nodule identification device based on Mamba, characterized in that, include: The sample acquisition module is used to acquire low-dose chest CT images and construct a tissue library with the corresponding benign and malignant identification information to obtain an labeled lung nodule sample set. The sample processing module is used to perform nodule block trimming, spatial resampling and CT value standardization on the lung nodule sample set to obtain a standardized nodule block set. The sample augmentation module is used to perform multi-axial random rotation and spherical random translation augmentation on the standardized nodule block set to obtain the lung nodule training tensor set. The model building module is used to build a Mamba state-space multi-directional scanning recognition network based on the training tensor set of the lung nodules to obtain an initial 3D nodule recognition model. The model training module is used to input the lung nodule training tensor set into the initial 3D nodule recognition model for iterative training to obtain the target 3D nodule recognition model. The result recognition module is used to input the CT image to be recognized into the target 3D nodule recognition model for recognition and output the lung nodule recognition result. Specifically, the model construction module is used to: input the lung nodule training tensor set into a two-layer three-dimensional convolution module for shallow feature encoding to obtain a first-scale feature map set; and perform three-dimensional block embedding and position encoding injection on the first-scale feature map set to obtain a three-dimensional feature token sequence. The Mamba state space multi-directional scanning modeling process is performed using the three-dimensional feature token sequence to obtain a multi-directional context feature set; the multi-directional context feature set is then subjected to hierarchical downsampling fusion and residual connection processing to obtain a multi-scale nodule feature set; the multi-scale nodule feature set is then subjected to global pooling and classification header parameter initialization encapsulation to obtain an initial 3D nodule recognition model. The result recognition module is specifically used for: receiving a low-dose chest CT image to be identified and performing nodule center localization and three-dimensional target region cropping to obtain a nodule block to be identified; performing voxel resampling, CT value normalization, and tensor dimension encapsulation on the nodule block to be identified to obtain an input tensor to be identified; performing enhanced flipping and batch stacking processing on the input tensor to be identified to obtain a multi-view test tensor set; inputting the multi-view test tensor set into the target 3D nodule recognition model for forward inference and multi-view probability fusion to obtain benign and malignant prediction probabilities; and performing category threshold determination and nodule localization information backfilling and encapsulation based on the benign and malignant prediction probabilities to obtain the lung nodule recognition result.
7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the Mamba-based 3D lung nodule identification method as described in any one of claims 1 to 5.
8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the Mamba-based 3D lung nodule identification method as described in any one of claims 1 to 5.
