A fully automatic segmentation system and method for pulmonary nodules based on ultra-high resolution target scanning CT

Abstract

This application relates to the field of medical image processing technology, and discloses a fully automated lung nodule segmentation system and method based on ultra-high resolution target scanning CT. The system includes a data preprocessing module, a 3D target detection module, an adaptive resampling module, a 3D fine segmentation module, and a post-processing and feature extraction module. The detection module outputs a 3D bounding box to provide nodule in-plane size information for adaptive resampling; the resampling module maintains the Z-axis unchanged while adaptively downsampling the XY axes; the segmentation module adopts a dual-cascaded feature pyramid network architecture to achieve two-stage optimization of coarse and fine segmentation; the post-processing module generates a binarized segmentation mask and calculates quantitative features. This application effectively mitigates the impact of some volume effects on the segmentation accuracy of subsolid nodule boundaries.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a fully automated detection and three-dimensional segmentation technology for lung nodules based on ultra-high resolution target scanning computed tomography (CT). Background Technology

[0002] The widespread use of low-dose computed tomography (CT) in lung cancer screening has led to a continuous increase in the detection rate of subsolid nodules (SSNs). SSNs include pure ground-glass opacities and partially solid nodules, and their pathological spectrum covers preinvasive lesions such as atypical adenomatous hyperplasia and carcinoma in situ to invasive lesions such as minimally invasive adenocarcinoma and invasive adenocarcinoma. Accurate subtyping is crucial for developing clinical follow-up strategies and treatment plans. In actual clinical practice, the quantitative assessment of SSNs—including the measurement of parameters such as nodule size, density distribution, and composition—is highly dependent on the accuracy of segmentation results. Inaccurate segmentation can systematically affect the reliability of these quantitative parameters.

[0003] In the first typical scenario, for subsolid nodules growing closely attached to the pleura, the grayscale difference between the nodule edge and the pleural surface is minimal. Existing segmentation methods often mistakenly include part of the pleural tissue in the nodule outline or miss the nodule edge close to the pleura. In the second scenario, for subsolid nodules adhered to intrapulmonary vessels, the density transition area between the nodule and the vessel along the vessel's direction is blurred, and the segmentation boundary tends to extend excessively along the vessel direction. In the third scenario, small-diameter ground-glass nodules, due to their extremely low density contrast with the surrounding lung parenchyma, are particularly affected by partial volume effects on conventional resolution CT images. The boundary region occupies a large proportion of the total nodule volume, making it difficult for voxel intensity-based segmentation methods to accurately define their extent.

[0004] Existing automatic segmentation methods are typically developed based on conventional resolution CT images. When faced with ultra-high resolution target scan CT images, they encounter adaptation challenges such as excessively large data volumes and significant anisotropy in in-plane and inter-slice resolution, lacking dedicated processing strategies for these high-resolution data characteristics. Furthermore, existing methods suffer from insufficient information integration between the detection and segmentation stages, failing to effectively utilize the target size information carried by the detection results to guide parameter adaptation in subsequent processing. Therefore, there is an urgent need for a technical solution that can fully leverage the spatial detail advantages of ultra-high resolution target scan CT to achieve fully automated and accurate detection and 3D segmentation. Summary of the Invention

[0005] The purpose of this application is to provide a fully automated lung nodule segmentation system and method based on ultra-high resolution target scanning CT, so as to solve the problems mentioned in the background art.

[0006] This application discloses a fully automated lung nodule segmentation system based on ultra-high resolution target scanning CT, comprising: The data preprocessing module is configured to acquire ultra-high resolution target scan CT image data of the lungs with physical spacing attributes, perform CT value intensity standardization based on set window width and window level and morphological lung field extraction, and generate complete lung field volume data. The 3D target detection module is configured to input the lung field volume data into an anchor-based 3D target detection network, locate the spatial position of the lung nodules, and output a three-dimensional bounding box containing the lung nodules. An adaptive resampling module is configured to crop the region of interest from the lung field volume data based on the three-dimensional bounding box, and obtain the maximum pixel size within the nodule plane determined by the three-dimensional bounding box; determine the downsampling factor based on the ratio of the maximum pixel size within the nodule plane to the preset cropping size of the model, and perform adaptive dynamic resampling on the region of interest. The adaptive dynamic resampling maintains the physical layer thickness along the Z-axis unchanged, and performs adaptive downsampling along the XY-axis according to the downsampling factor to obtain the resampled region of interest. The 3D fine segmentation module is configured to input the resampled region of interest into a 3D fine segmentation network based on a dual-cascaded feature pyramid network architecture for voxel-level segmentation; the first segmentation stage of the 3D fine segmentation network extracts coarse segmentation features and outputs a preliminary 3D probability map and a preliminary segmentation mask; the second segmentation stage of the 3D fine segmentation network expands the preliminary segmentation mask in 3D space, concatenates it with the resampled region of interest in the feature channel dimension as input, extracts fine segmentation features, and outputs the final 3D probability map. The post-processing and feature extraction module is configured to perform thresholding and morphological optimization on the final three-dimensional probability map, generate the final binary three-dimensional segmentation mask, and separate nodule components and calculate quantitative features based on the three-dimensional segmentation mask and a preset physical CT value threshold.

[0007] In a preferred embodiment, the acquisition matrix of the ultra-high resolution target scan CT image data of the lung is at least 1024×1024, and the Z-axis slice thickness is no more than 1.0 mm. The scanning field of view is focused on the area where the nodule is located, and the standard lung algorithm is used for reconstruction. In the data preprocessing module: The CT value intensity standardization based on the set window width and window level is as follows: using a window width and window level range of [-1500 HU, 400 HU], the original CT value is linearly mapped to a numerical range of 0 to 1 to cover the range from ground glass density to soft tissue density. The morphological lung field extraction is performed as follows: the image data is segmented using -400 HU as a fixed threshold to obtain an initial binary mask of the lung region; morphological closure operation is performed sequentially to fill the trachea and small airways; three-dimensional connected component analysis is performed to retain the two largest three-dimensional connected components as the left and right lungs; and morphological expansion operation is performed to expand the lung field boundary to ensure that subpleural nodules are not clipped, thereby generating the complete lung field volume data; wherein, the morphological closure operation uses a spherical structural element with a radius of 3 voxels, and the morphological expansion operation uses a spherical structural element with a radius of 5 voxels.

[0008] In a preferred embodiment, the rule for determining the downsampling factor in the adaptive resampling module based on the ratio of the maximum pixel size within the nodule surface to the preset clipping size of the model is as follows: When the maximum pixel size within the nodule surface is less than or equal to the preset clipping size of the model, a 1x downsampling is used; When the maximum pixel size within the nodule surface is greater than the preset clipping size of the model but less than or equal to 1.5 times the preset clipping size of the model, a 2x downsampling is used; When the maximum pixel size within the nodule surface is greater than 1.5 times the preset clipping size of the model and less than or equal to 2 times the preset clipping size of the model, verification is performed using 2x and 3x double downsampling respectively; When the maximum pixel size within the nodule plane is greater than twice the preset clipping size of the model, verification is performed using 3x and 4x double downsampling, respectively. The adaptive dynamic resampling uses cubic spline interpolation when interpolating CT images and nearest neighbor interpolation when resampling segmentation masks.

[0009] In a preferred embodiment, the anchor-based 3D target detection network has 1 input channel, 16 basic filters, and a detection range of subsolid nodules with a diameter of 4-30 mm. The 3D target detection network includes a feature pyramid network structure, which contains 8 downsampling levels and 7 upsampling levels. The downsampling levels are constructed using convolutional blocks, residual basic blocks, and max pooling layers, and the upsampling levels are constructed using upsampling blocks. The 3D target detection network also includes a detection head module, which embeds a receptive field enhancement module for fusing and optimizing multi-scale features, predicting the coordinates of the three-dimensional bounding box through a localization head, and determining the nodule category through a classification head; the localization anchor point corresponding to the localization head is set to [3, 3], and the classification anchor point corresponding to the classification head is set to [1, 1, 1, 1].

[0010] In a preferred embodiment, the 3D fine segmentation network based on a dual-cascaded feature pyramid network architecture has 1 input channel and 8 basic filters; the 3D fine segmentation network includes two cascaded feature pyramid modules and two corresponding segmentation modules constructed using 1×1 convolutional layers, wherein: The first feature pyramid module corresponding to the first segmentation stage and the second feature pyramid module corresponding to the second segmentation stage both contain 7 downsampling levels and 6 upsampling levels; the downsampling levels are constructed using an enhanced receptive field module and residual basic blocks with a squeezing and excitation SE attention mechanism; the upsampling levels are constructed using upsampling blocks, and a receptive field enhancement module is embedded at the mid-to-low-level feature output end to enhance the feature extraction capability of the nodule edge region; The dual-cascaded feature pyramid network architecture introduces an attention gating mechanism at skip connections to adaptively suppress feature responses from irrelevant background regions in order to highlight nodule boundary information.

[0011] In a preferred embodiment, the specific execution process of the first segmentation stage is as follows: extracting features of the resampled region of interest through the first feature pyramid module, outputting the preliminary three-dimensional probability map through the first segmentation module, calculating the probability value of each voxel belonging to a nodule through the Sigmoid activation function, and generating the preliminary segmentation mask through thresholding. The specific execution process of the second segmentation stage is as follows: after the preliminary segmentation mask is expanded in three-dimensional space, it is spliced ​​with the resampled region of interest in the feature channel dimension to form a multi-channel input. The feature extraction is further optimized by the second feature pyramid module, and the final three-dimensional probability map is output by the second segmentation module.

[0012] In a preferred embodiment, the post-processing and feature extraction module includes: The thresholding process uses 0.5 as the threshold value to convert the final three-dimensional probability map into a binary mask. The morphological optimization includes filling the internal holes of the binary mask and removing small connected regions with a three-dimensional volume of less than 30 voxels. The separated nodule components are as follows: within the finally determined binary three-dimensional segmentation mask area, with -300 HU as the preset physical CT value threshold, the area with the original CT value less than -300 HU is divided into ground glass components, and the area with the original CT value greater than or equal to -300 HU is divided into solid components. The quantitative features include at least the total volume of the nodule, the volume of the solid component, the average CT value, and the histogram distribution of the CT value.

[0013] In a preferred embodiment, the system further includes a model training module configured as follows: For training the anchor-based 3D object detection network, a weighted sum of Focal Loss and Smooth L1 Loss is used as the loss function, with a weight ratio of 1:1. Focal Loss is used as the classification loss, with a focusing parameter γ set to 2 and a balance factor α set to 0.25. Smooth L1 Loss is used as the regression loss. Positive samples are defined as IoU ≥ 0.5 between the anchor point and the gold standard bounding box, and negative samples are defined as IoU < 0.3, with the rest ignored. The ratio of positive to negative samples is controlled to 1:3 through an online hard example mining strategy. The parameters of the anchor-based 3D object detection network are frozen, and the region of interest cropped and scaled by the gold standard bounding box is used as input sample to train the 3D fine segmentation network to avoid the propagation of detection errors. The combined loss function L used for training the 3D fine segmentation network is: Wherein, Ldice is the standard Dice similarity coefficient loss, used to measure the spatial morphological overlap between the segmentation mask predicted by the model and the ground truth labeled mask; Lce is the standard binary cross-entropy loss, used to measure the difference in classification distribution at the voxel level.

[0014] In a preferred embodiment, the data augmentation strategy introduced during the training phase of the 3D fine segmentation network includes: random flipping along each of the three axes with a 50% probability; random rotation of ±10° around each of the three axes; isotropic scaling with a scaling factor uniformly distributed in [0.9, 1.1]; elastic deformation using B-spline transformation to simulate tissue deformation caused by body position; and random Gaussian noise with a standard deviation σ∈[0, 0.01].

[0015] In a preferred embodiment, in the model training module, both the 3D object detection network and the 3D fine segmentation network are trained using the Adam optimizer with an initial learning rate of 0.001 and a weight decay of 1×10⁻⁴. During the training of the 3D object detection network, the batch size is 4, with each batch containing 4 1283 voxel blocks. The training process employs a cosine annealing learning rate decay strategy, with a minimum learning rate of 1×10⁻⁶ and a maximum training epoch of 150. The early stopping condition is: if the validation set loss does not decrease for 20 consecutive training epochs, then training is terminated.

[0016] In a preferred embodiment, the gold standard bounding box and voxel-level annotations in the model training module are obtained as follows: two thoracic radiologists with more than 10 years of experience in lung CT diagnosis independently perform layer-by-layer three-dimensional manual segmentation, and the consistency between the annotators is measured by the Dice coefficient between the segmentation results of the two doctors, requiring a Dice coefficient ≥ 0.85; for cases with a Dice coefficient < 0.85, a third senior radiologist with more than 15 years of experience arbitrates and determines the final gold standard segmentation; the division of the training set and the validation set is based on patients, ensuring that all nodules of the same patient appear only in the training set or the validation set.

[0017] This application also discloses a fully automated lung nodule segmentation method based on ultra-high resolution target scanning CT, wherein the various steps are executed sequentially and in concert, including the following steps: Step S1: Acquire ultra-high resolution target scan CT image data of the lungs with physical spacing attributes, perform CT value intensity normalization based on set window width and window level and morphological lung field extraction to generate complete lung field volume data; Step S2: Input the lung field volume data generated in step S1 into an anchor-based 3D target detection network to locate the spatial position of the lung nodules and output a three-dimensional bounding box containing the lung nodules. Step S3: Based on the 3D bounding box output in step S2, crop the region of interest from the lung field volume data generated in step S1, and obtain the maximum pixel size within the nodule plane determined by the 3D bounding box; determine the downsampling factor based on the ratio of the maximum pixel size within the nodule plane to the preset cropping size of the model, and perform adaptive dynamic resampling on the region of interest. The adaptive dynamic resampling maintains the physical layer thickness on the Z-axis unchanged, and performs adaptive downsampling on the XY-axis according to the downsampling factor to obtain the resampled region of interest. Step S4: Input the resampled region of interest obtained in Step S3 into a 3D fine segmentation network based on a dual-cascaded feature pyramid network architecture for voxel-level segmentation; the first segmentation stage of the 3D fine segmentation network extracts coarse segmentation features and outputs a preliminary 3D probability map and a preliminary segmentation mask; the second segmentation stage of the 3D fine segmentation network expands the preliminary segmentation mask in 3D space and concatenates it with the resampled region of interest obtained in Step S3 in the feature channel dimension as input to extract fine segmentation features and output the final 3D probability map; Step S5: Threshold and morphologically optimize the final three-dimensional probability map output in step S4 to generate the final binary three-dimensional segmentation mask. Based on the three-dimensional segmentation mask and the original CT value data obtained in step S1, separate the nodule components and calculate quantitative features using a preset physical CT value threshold.

[0018] In a preferred embodiment, the 3D target detection network in step S2 has 1 input channel and 16 basic filters; the 3D target detection network includes a feature pyramid network structure, which includes 8 downsampling levels and 7 upsampling levels; the detection head module embeds a receptive field enhancement module; the localization anchor point is set to [3, 3]; and the classification anchor point is set to [1, 1, 1, 1].

[0019] In a preferred embodiment, the downsampling factor determination rule for adaptive dynamic resampling in step S3 is as follows: when the maximum pixel size within the nodule plane is less than or equal to the preset clipping size of the model, 1x downsampling is used; when it is greater than the preset clipping size of the model but less than or equal to 1.5x, 2x downsampling is used; when it is greater than 1.5x but less than or equal to 2x, 2x and 3x double downsampling are used for verification respectively; when it is greater than 2x, 3x and 4x double downsampling are used for verification respectively; the adaptive dynamic resampling uses cubic spline interpolation when interpolating CT images and nearest neighbor interpolation when resampling the segmentation mask.

[0020] In a preferred embodiment, the number of input channels of the 3D fine segmentation network in step S4 is 1, and the number of basic filters is set to 8; the first feature pyramid module and the second feature pyramid module both contain 7 downsampling levels and 6 upsampling levels. The downsampling levels are constructed using enhanced receptive field modules and residual basic blocks with squeezing and excitation SE attention mechanisms, and the upsampling levels are constructed using upsampling blocks, with receptive field enhancement modules embedded at the mid-to-low-level feature output ends; the segmentation module is constructed using 1×1 convolutional layers.

[0021] In a preferred embodiment, in step S4, the first segmentation stage calculates the probability value of each voxel belonging to a nodule using the Sigmoid activation function, and generates the preliminary segmentation mask after thresholding; the second segmentation stage expands the preliminary segmentation mask in three-dimensional space and splices it with the resampled region of interest in the feature channel dimension to form a multi-channel input.

[0022] In a preferred embodiment, the thresholding in step S5 uses 0.5 as the threshold; the morphological optimization includes filling pores and removing small connected regions with a volume less than 30 voxels; the separation of nodule components uses -300 HU as the preset physical CT value threshold to separate the ground-glass component and the solid component; the quantitative features include at least the total nodule volume, the solid component volume, the average CT value, and the CT value histogram distribution.

[0023] This application also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0024] This application also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0025] The technical solution provided in this application addresses multiple technical challenges faced in the automatic segmentation of subsolid lung nodules using ultra-high resolution target scan CT images, such as excessively large data volume, significant anisotropy of in-plane and inter-layer resolution, and blurred boundaries between nodules and adjacent tissues. It forms an organic and synergistic overall processing solution through a technical chain that connects the modules of data preprocessing, 3D target detection, adaptive resampling, dual-cascade 3D fine segmentation, and post-processing and feature extraction.

[0026] In the data preprocessing stage, the CT value intensity standardization based on the set window width and window level range covers the complete range from ground-glass opacity to soft tissue density, avoiding information truncation of different density components in subsolid nodules. Morphological lung field extraction is carried out through the sequential combination of closing operation to fill airway cavities, connected domain analysis to preserve the left and right lungs, and expansion operation to expand the lung field boundary. This ensures that subpleural nodules are not missed due to overly tight mask boundaries while removing irrelevant tissue, providing a complete input volume for subsequent detection and segmentation.

[0027] At the level of connection between detection and resampling, the 3D target detection network adopts a feature pyramid network structure containing multiple downsampling and upsampling levels. Combined with the receptive field enhancement module embedded in the detection head and the setting of multiple anchor points, it enhances the detection sensitivity of subsolid nodules of different sizes. Crucially, the 3D bounding boxes output by the detection network not only enable spatial localization but also provide prior information on the maximum pixel size within the nodule plane for adaptive resampling. This effectively resolves the logical-temporal contradiction that resampling requires nodule size when the nodule has not yet been detected, allowing for a fully automated processing flow. The adaptive resampling module determines the downsampling factor based on the ratio of the maximum pixel size within the nodule plane to the model's preset clipping size. When at the boundary of the downsampling range, a dual downsampling verification strategy is used to select the optimal segmentation result. This asymmetric strategy, maintaining the Z-axis physical layer thickness while performing downsampling only on the X and Y axes, effectively alleviates the inherent anisotropy problem of target scan CT data, preserving significantly higher in-plane spatial details than conventional CT while controlling the data size. The differential processing of cubic spline interpolation on the CT image and nearest-neighbor interpolation on the segmentation mask balances edge sharpness preservation with mask binary properties maintenance.

[0028] At the segmentation level, the dual-cascade feature pyramid network architecture links coarse and fine segmentation in the same technical chain. The first segmentation stage provides preliminary localization and contour information of the main nodule region. The output preliminary segmentation mask is expanded in three-dimensional space and stitched with the resampled region of interest in the channel dimension, guiding the second segmentation stage to focus on the edge gradient zone for targeted refinement. A synergistic relationship of functional division and information transmission is formed between the two stages. In the downsampling level, the enhanced receptive field module and the residual basic block with squeezing and excitation attention mechanisms work together to expand the effective receptive field and achieve adaptive channel weighting. The attention gating mechanism introduced at the jump connection uses deep semantic information to adaptively suppress the feature response of irrelevant background regions. The combined effect of the three helps to reduce the mis-inclusion of pleural tissue and excessive extension of the vascular direction in complex adjacency scenarios such as close adjacency to the pleura or vascular adhesion. The post-processing module generates the final binarized three-dimensional segmentation mask through thresholding and morphological optimization, and automatically separates the ground-glass components from the solid components and calculates quantitative features using physical CT value thresholds, providing structured data support for clinical quantitative assessment.

[0029] Figure 3 This paper demonstrates the fully automated segmentation results of the proposed system on six representative cases of subsolid pulmonary nodules. Each case is presented in a left-right comparison: the left image is the original ultra-high resolution target scan CT image, and the right image shows the display of the automatic detection and segmentation results of this application superimposed on the original image. In the right image, the red rectangle is the bounding box output by the 3D target detection network, representing the nodule region located during the detection stage; the blue / cyan outline is the segmentation mask boundary output by the attention-guided 3D fine segmentation network, representing the final nodule segmentation range. The six nodules covered different anatomical locations and morphological characteristics: the first and second were solitary nodules (pure ground-glass nodules) within the lung field, with the segmentation contour closely following the nodule edge and clear, complete boundaries; the third and fourth were vascular adhesion nodules, and the segmentation network accurately distinguished the nodules from adjacent vessels, avoiding misclassification of vessels within the segmentation range; the fifth was a nodule closely adhering to the pleura, and the segmentation network correctly separated the nodule from the pleural surface, avoiding misclassification of pleural tissue as nodule components; the sixth was a small, partially solid nodule, with the yellow / green outline area indicating automatic identification of solid components, demonstrating the component separation function of the post-processing module. These results intuitively demonstrate the automatic segmentation accuracy and robustness of the system under different complex anatomical conditions.

[0030] In the model training phase, the strategy of freezing the detection network parameters and training the segmentation network using regions of interest clipped with the gold standard bounding box avoids the propagation and accumulation of detection errors into the segmentation stage, helping to ensure the stability of segmentation performance. In summary, the above technical aspects form a closed, interdependent organic whole around the physical data characteristics of ultra-high resolution target scanning CT and the needs of clinical scenarios, enabling the system to effectively improve the accuracy and robustness of 3D segmentation of subsolid lung nodules in an end-to-end fully automated workflow.

[0031] The specification of this application contains numerous technical features distributed across various technical solutions. Listing all possible combinations of these technical features (i.e., technical solutions) would make the specification excessively lengthy. To avoid this problem, the various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which are considered to have been described in this specification), unless such a combination of technical features is technically infeasible. For example, one example discloses feature A+B+C, and another example discloses feature A+B+D+E. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; they cannot be used simultaneously. Feature E can technically be combined with feature C. Therefore, the solution A+B+C+D should not be considered as described because it is technically infeasible, while the solution A+B+C+E should be considered as described. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the structure of the fully automated lung nodule segmentation system based on ultra-high resolution target scanning CT according to this application.

[0033] Figure 2 This is a flowchart illustrating a fully automated lung nodule segmentation method based on ultra-high resolution target scanning CT according to the second embodiment of this application.

[0034] Figure 3 This is an image showing the automatic segmentation effect of the fully automatic lung nodule segmentation system and method based on ultra-high resolution target scanning CT according to this application on a representative nodule.

[0035] Figure 4 This is another structural schematic diagram of the fully automated lung nodule segmentation system based on ultra-high resolution target scanning CT according to this application.

[0036] Figure 5 This is another flowchart illustrating the fully automated lung nodule segmentation method based on ultra-high resolution target scanning CT according to the second embodiment of this application. Detailed Implementation

[0037] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0038] Explanation of some concepts: Ultra-high resolution target scan CT refers to high-resolution computed tomography images acquired using target scan mode. Its acquisition matrix is ​​at least 1024×1024, the Z-axis slice thickness is no more than 1.0 mm, the scanning field of view is focused on the area where the target nodule is located, and standard lung algorithm is used for reconstruction. The spatial resolution is significantly higher than that of conventional 512×512 matrix CT images.

[0039] Subsolid nodules (SSNs) are pulmonary nodules that appear as non-completely solid densities on CT images. They include pure ground-glass nodules (pGGN) and partially solid nodules, which may contain both ground-glass and solid density components.

[0040] The partial volume effect refers to the physical phenomenon that when a CT image voxel covers a space containing tissues of different densities, the CT value of that voxel is a weighted average of the densities of each tissue, causing the tissue boundaries to appear as a gradual grayscale transition area rather than a clear interface on the image.

[0041] Hounsfield Unit (HU) is a standard unit of measurement used in CT images to quantify tissue density. The density of water is defined as 0 HU, and air is defined as -1000 HU. Different tissues exhibit corresponding HU values ​​based on their absorption characteristics of X-rays.

[0042] Feature Pyramid Network (FPN) is a deep neural network structure that achieves multi-scale feature extraction and fusion through top-down paths and lateral connections, and can simultaneously preserve semantic information and spatial details at different spatial resolution levels.

[0043] The Dice similarity coefficient (DSC) is a similarity metric used to measure the spatial overlap between two sets. It is calculated by dividing twice the prime number of the intersection of the two sets by the sum of the prime numbers of the two sets. The value ranges from 0 to 1, with a higher value indicating a higher degree of overlap.

[0044] The receptive field enhancement module (RfbBlock3d) is a three-dimensional convolutional module that expands the effective receptive field through a multi-branch parallel convolutional structure. It is used to enhance the network's ability to capture features at different scales without significantly increasing the number of parameters.

[0045] Squeeze-and-Excitation (SE) is a channel attention mechanism that adaptively calculates the importance weights of each channel by performing global average pooling, compression, and excitation operations on each channel of the feature map, thereby achieving channel-dimensional feature weighting.

[0046] Attention gating is an attention computation structure introduced at the skip connections of the encoder-decoder network. It uses the deep features of the decoder as gating signals to calculate attention weights at each spatial location of the shallow features of the encoder, thereby adaptively suppressing the feature responses of irrelevant background regions and highlighting the target boundary information.

[0047] A 3D bounding box is the smallest enclosing rectangle used to define the spatial extent of a target object in 3D space, determined by its start and end coordinates along the three coordinate axes.

[0048] In general, in the field of automated segmentation of subsolid pulmonary nodules, existing technologies typically develop segmentation models based on conventional resolution CT (512×512 matrix). Due to partial volume effects, the grayscale transition zone between the nodule and adjacent pleura or blood vessels is difficult to define precisely. Furthermore, directly inputting ultra-high resolution target scan CT (1024×1024 matrix, FOV approximately 180mm, in-plane voxel spacing approximately 0.176mm) into existing networks presents practical obstacles such as data volume far exceeding GPU memory capacity and severe anisotropy in XY and Z axis resolution. Those skilled in the art lack the motivation and a clear technical path to simultaneously resolve these two contradictions. This application creatively constructs a "firstly, a 3D target detection network (... Figure 4 The system outputs a 3D bounding box to determine the spatial location of the nodules and their in-plane pixel size. Based on this, the cropped region of interest is dynamically resampled, with the Z-axis remaining unchanged and the XY-axis adaptively downsampled by 1 to 4 times according to the ratio of the maximum in-plane pixel size of the nodule to the model's preset cropping size. Finally, the resampled region of interest is fed into a 3D fine segmentation network with a dual-cascaded feature pyramid network architecture. Figure 5 This technology chain—in which the 3D bounding box in the detection phase not only performs the localization function, but also provides previously unavailable nodule scale priors for resampling factor determination, thus effectively resolving the logical-temporal contradiction that "resampling requires nodule size but the nodule has not yet been detected," enabling the fully automated processing flow; while the segmentation network ( Figure 5The initial segmentation mask output by the first feature pyramid module (FPN1) and the first segmentation module (Segmentation1) in the first segmentation stage is spliced ​​with the resampled region of interest in the feature channel dimension after being expanded in three-dimensional space. This provides a priori positional constraints for the fine segmentation of the second feature pyramid module (FPN2) and the second segmentation module (Segmentation2). Combined with the attention gating mechanism introduced at the skip connections and the synergistic effect of the residual basic blocks with SE attention and the enhanced receptive field module (RfbeBlock3d) in the downsampling level, the network can perform targeted feature enhancement and background suppression for nodule boundary regions that are close to the pleura or adhere to blood vessels. The above technical features are not a simple stacking of known modules, but rather an organic whole formed around the physical characteristics of ultra-high resolution target scan CT data and the needs of clinical scenarios. This results in a synergistic technical effect that improves the accuracy and stability of three-dimensional segmentation of subsolid nodules in complex adjacent scenarios while effectively controlling the data scale.

[0049] More specifically, the inventors of this application have discovered through long-term and in-depth research that the root cause of the above-mentioned technical problems is not a deficiency in a single link, but rather a technical contradiction that is intertwined among multiple levels, including data characteristics, processing logic, and network architecture adaptation.

[0050] In their practical experience processing a large number of clinical ultra-high resolution target scan CT images, the inventors discovered that there is a significant anisotropic difference between the in-plane resolution and the inter-slice resolution of such images (typical parameters are 1024×1024 matrix, in-plane pixel pitch of about 0.176mm, and slice thickness of 1.0mm). If the entire high-resolution volume data is directly input into the segmentation network, the data size will far exceed the carrying capacity of the GPU memory. If the data is downsampled globally by a fixed factor to adapt to computing resources, the advantage of target scan CT over conventional CT in terms of boundary sharpness will be eliminated, making high-resolution acquisition meaningless. The inventors further recognized that nodules of different sizes have fundamentally different tolerances to downsampling factors—smaller nodules are prone to losing crucial edge details after oversampling, while larger nodules cause severe data redundancy if not downsampled—therefore, an adaptive strategy is needed to differentiate processing based on the actual in-plane dimensions of the nodules. However, the implementation of this strategy presupposes that the spatial location and in-plane dimensions of the nodules are known before resampling. This requires that the detection step be completed before resampling, and that the output of the detection network can directly provide usable size information for determining the resampling factor. Through a thorough analysis of the above logical relationships, the inventors creatively established a technical route with a specific timing arrangement: "first, perform 3D detection on the complete lung field to obtain a 3D bounding box; then, perform adaptive resampling based on the in-plane pixel dimensions determined by the bounding box; and finally, perform fine segmentation on the resampled region of interest."

[0051] Meanwhile, through a systematic analysis of the difficulties in segmenting subsolid nodules, the inventors recognized that single-stage segmentation networks often exhibit instability in boundary accuracy when facing boundary regions where the gray-scale gradient between ground-glass components and lung parenchyma is extremely gradual, due to a lack of prior constraints on the approximate location and morphology of the nodules. Introducing the coarse segmentation results as a mask into the second-stage segmentation as prior guidance allows the fine segmentation stage to focus purposefully on regions with blurred boundaries, rather than indiscriminately allocating computational resources across the entire region of interest. Based on the above in-depth research, the inventors of this application propose a technical solution that organically combines the characteristics of ultra-high resolution target scan CT data, detection-driven adaptive resampling, and two-stage fine segmentation.

[0052] In this embodiment, this application is preferably applied to the automatic detection and fine three-dimensional segmentation of subsolid pulmonary nodules. Unlike general two-dimensional natural image segmentation, the input data processed in this application has medical image physical attributes such as Hounsfield Unit values, voxel physical spacing, slice thickness, slice interval, matrix size, and scanning field of view. Therefore, the various technical features of this application are not simply aimed at abstract data processing procedures, but rather revolve around the preprocessing, spatial localization, region resampling, three-dimensional segmentation, and post-segmentation quantitative calculation of specific medical image data. Furthermore, the output objects of this application are limited to bounding boxes, three-dimensional probability maps, binarized three-dimensional segmentation masks, and quantitative features obtained based on the masks, rather than directly outputting pathological conclusions or diagnostic and treatment decisions.

[0053] Example 1: Fully Automated Lung Nodule Segmentation System Based on Ultra-High Resolution Target Scan CT See Figure 1 , Figure 2 , Figure 4 and Figure 5 This embodiment provides a fully automated lung nodule segmentation system based on ultra-high resolution target scanning CT. The system includes a data preprocessing module, a 3D target detection module, an adaptive resampling module, a 3D fine segmentation module, a post-processing and feature extraction module, and optionally a model training module. These modules are not simply parallel but are sequentially linked, forming a complete and closed technical chain based on the data flow of "complete lung field volume data → bounding box → region of interest → resampled region of interest → probability map → segmentation mask → quantitative features". Each module is described in detail below.

[0054] I. Data Preprocessing Module The data preprocessing module is configured to acquire ultra-high resolution target scan CT image data of the lungs with physical spacing attributes, and perform standardization processing and lung field extraction on the data to generate complete lung field volume data, providing high-quality input for subsequent detection and segmentation.

[0055] Specifically, the ultra-high resolution target scan CT image data of the lungs mentioned in this embodiment refers to high-resolution CT images acquired using target scan mode. Unlike conventional CT images (typically reconstructed using a 512×512 matrix), the ultra-high resolution target scan CT images used in this application have an acquisition matrix of at least 1024×1024, a Z-axis slice thickness of no more than 1.0 mm, and the field of view (FOV) is focused on the region where the nodule is located, reconstructed using a standard lung algorithm. For example, in a specific application scenario, the scanning parameters include a matrix of 1024×1024, a slice thickness of 1.0 mm, a slice interval of 1.0 mm, a tube voltage of 120 kV, and a scanning field of view of approximately 180 mm × 180 mm. When the FOV is 180 mm and the matrix is ​​1024×1024, the in-plane pixel pitch is approximately 0.176 mm, which significantly improves the spatial resolution compared to a conventional 512×512 matrix CT. This high spatial resolution effectively alleviates the boundary blurring problem caused by partial volume effect in conventional CT, allowing the interface between nodules and surrounding tissues (such as blood vessels and pleura) to be presented more clearly in the image. However, this type of data also brings processing challenges such as larger data volume and more obvious inconsistencies in resolution between the XY and Z directions. Therefore, the subsequent modules are not a simple copy of the conventional CT segmentation process, but are specifically adapted for this data characteristic. It should be noted that those skilled in the art can adjust the specific scanning parameters according to the actual equipment, as long as the spatial resolution of the obtained image is significantly higher than that of a conventional 512×512 matrix CT.

[0056] The data preprocessing module first reads the original DICOM file and extracts pixel data and metadata, including parameters such as reconstruction spacing, slice thickness, slice spacing, and window width and level. This metadata has a clear physical meaning and runs through the subsequent processing flow to ensure the physical spatial accuracy of each step.

[0057] Furthermore, the data preprocessing module performs CT value intensity normalization based on a set window width and level. Specifically, using a window width WW=1900HU and a window level WL=-550HU, the original CT values ​​(Hounsfield Units, HU) are linearly mapped to a numerical range of 0 to 1 within the [-1500 HU, 400 HU] range. The selection of this window width and level range has a clear physical basis: it covers the complete density range from lung parenchyma (approximately -800 HU) to soft tissue (approximately +50 HU), simultaneously preserving image details of both the ground-glass opacity component (approximately -700 to -300 HU) and the solid component (approximately -300 to +50 HU), avoiding information truncation. The formula for calculating the normalized mapping is: Wherein, Vnorm is the standardized value, ranging from [0, 1]; VHU is the original CT value; WLmin is the lower limit of the window width and window level interval (-1500 HU in this embodiment); WLmax is the upper limit of the window width and window level interval (400 HU in this embodiment). CT values ​​below WLmin are uniformly mapped to 0, and values ​​above WLmax are uniformly mapped to 1.

[0058] The data preprocessing module also performs morphological lung field extraction, which extracts the lung parenchyma from the complete chest CT volume, removing irrelevant tissues such as the chest wall and bones, thereby defining the volume of interest for subsequent processing. This operation includes the following sub-steps executed sequentially: First, the CT volume data is segmented using a fixed threshold of -400 HU, marking voxels with CT values ​​below this threshold as initial lung regions. The -400 HU threshold lies between the lung parenchyma density range (approximately -800 to -500 HU) and the soft tissue density range (approximately -300 to +50 HU), effectively distinguishing between aerated lung tissue and non-lung tissue. Then, a morphological closing operation (using a spherical structuring element with a radius of 3 voxels) is performed on the obtained initial binary mask to fill the voids created by the trachea, bronchi, and small airways in the binary mask. Next, a 3D connected component analysis is performed to identify all 3D connected components in the binary mask, retaining the two largest connected components, namely the left and right lungs, while excluding other small regions within the body cavity that may have been missegmented. Finally, a morphological expansion operation (using a spherical structuring element with a radius of 5 voxels) is performed to expand the preserved left and right lung masks outward. The purpose of the expansion operation is to ensure that nodules located under the pleura are not clipped due to the mask boundaries being too close to the lung surface. After the above processing, complete lung field volume data is generated as input for the subsequent 3D target detection module.

[0059] II. 3D Target Detection Module The 3D target detection module is configured to receive complete lung field volume data output by the data preprocessing module. Through an anchor-based 3D target detection network, it automatically locates the spatial position of lung nodules within the lung field and outputs a 3D bounding box containing each detected nodule. It should be noted that in the technical flow of this application, the function of the 3D target detection module is not merely to "find whether there are nodules," but more importantly, to provide the subsequent adaptive resampling module with a clearly defined spatially defined and physically traceable region of interest, while also providing nodule in-plane size information as a basis for determining the resampling factor. This "detect first, then resampling" technical approach effectively solves the problem of sourcing nodule size information required for adaptive resampling—the maximum pixel size within the nodule plane is directly determined by the 3D bounding box output by the detection network, without the need for manual provision or pre-estimation, thus ensuring the fully automated nature of the system's processing flow.

[0060] More specifically, the 3D target detection network described in this embodiment is an anchor-based 3D target detection network with one input channel (i.e., single-channel grayscale CT volume) and 16 base filters (nbase_filters). Its detection range covers subsolid nodules with diameters of approximately 4-30 mm, encompassing the vast majority of lung subsolid nodule sizes that require attention in clinical practice. The core architecture of this detection network adopts a Feature Pyramid Network (FPN) structure, which enhances the detection capability for nodules of different sizes through multi-scale feature extraction and fusion. The Feature Pyramid Network structure includes 8 downsampling levels and 7 upsampling levels. In the downsampling path, each level is constructed progressively using convolutional blocks, ResBasicBlock3d, and max-pooling layers. As the levels deepen, the spatial resolution of the feature map gradually decreases, while the number of channels gradually increases, thereby extracting multi-level features from local texture to global semantics. Each level in the upsampling path is constructed through upsampling blocks (UpBlock3d), which gradually upsample deep semantic features and fuse them with shallow high-resolution features to form a multi-scale feature pyramid that combines semantic information and spatial details.

[0061] Furthermore, the 3D target detection network also includes a detection head module. This detection head module embeds a receptive field enhancement module (RfbBlock3d), which expands the effective receptive field without significantly increasing the number of parameters through a multi-branch parallel convolutional structure (containing convolutional kernels of different sizes and dilation rates). This allows for the fusion and optimization of multi-scale features, enhancing the network's ability to perceive targets of different sizes. The detection head has two parallel output branches: a localization head and a classification head. The localization head predicts the 3D bounding box coordinate offset corresponding to each anchor point, while the classification head determines whether a nodule target exists at each anchor point. The localization anchors (locanchors) are set to [3, 3], and the classification anchors (clsanchors) are set to [1, 1, 1, 1]. The use of multiple anchor points allows the detection network to adapt to nodule morphologies with different aspect ratios and sizes, helping to reduce missed detections of small nodules.

[0062] III. Adaptive Resampling Module The adaptive resampling module is configured to crop a region of interest (ROI) centered on a nodule from the lung field volume data based on the 3D bounding box output by the 3D target detection module, and then adaptively and dynamically resample this ROI. This module is a core collaborative step in balancing the preservation of details in ultra-high resolution images with the consumption of computational resources: if downsampling is performed directly at the whole image level, the high-resolution advantage of target scan CT will be lost; if resampling is not performed locally, the ROI of large nodules will seriously exceed the GPU memory limit. Therefore, this application creatively utilizes the prior size information output by the detection network to perform adaptive dynamic resampling on the cropped ROI.

[0063] Specifically, the adaptive resampling module first crops the region of interest containing the target nodule from the lung field volume data based on the 3D bounding box. Simultaneously, the module directly obtains the maximum in-plane pixel size of the nodule defined by the bounding box—that is, the number of pixels corresponding to the maximum dimension of the bounding box in the XY plane (in-plane direction)—from the coordinate range of the 3D bounding box. Since the 3D bounding box has already been output by the detection network, this pixel size information can be directly calculated from the bounding box coordinates without manual intervention.

[0064] Then, the adaptive resampling module determines the downsampling factor based on the ratio of the maximum in-plane pixel size of the nodule to the model's preset clipping size. The model's preset clipping size is the in-plane dimension of the standard input dimension of the segmentation network. The adaptive dynamic resampling maintains the physical slice thickness along the Z-axis unchanged, downsampling only along the XY-axis at the determined downsampling factor. The technical consideration for keeping the Z-axis unchanged is that the anisotropy of ultra-high resolution target scan CT is manifested in that the in-plane resolution along the XY-axis is much higher than the inter-slice resolution along the Z-axis (e.g., 0.176 mm in-plane versus 1.0 mm inter-slice). The Z-axis itself is no longer a high-resolution direction, and keeping it unchanged ensures that the details of the nodule in the inter-slice direction are not lost. This asymmetric resampling strategy effectively alleviates the anisotropy problem of target scan CT data.

[0065] More specifically, the rules for determining the downsampling factor are as follows: When the maximum pixel size within the nodule plane is less than or equal to the model's preset clipping size, 1x downsampling is used, i.e., no downsampling is performed to maintain the original in-plane resolution. This case is suitable for smaller nodules and can completely preserve their fine structure. When the maximum pixel size within the nodule plane is greater than the model's preset clipping size but less than or equal to 1.5 times the model's preset clipping size, 2x downsampling is used for verification, which is a more common case. When the maximum pixel size within the nodule plane is greater than 1.5 times the model's preset clipping size but less than or equal to 2 times the model's preset clipping size, 2x and 3x double downsampling are used for verification respectively. When the maximum pixel size within the nodule plane is greater than 2 times the model's preset clipping size, 3x and 4x double downsampling are used for verification respectively. The "double downsampling verification" refers to performing resampling and subsequent segmentation processes on the same ROI with two different downsampling factors, and finally selecting the better result as the final output based on the quality index of the segmentation results. This strategy is not just about "running the model one more time," but rather about improving segmentation robustness by comparing candidate scales to address issues such as boundary blurring and loss of local details that may occur in cases of large nodules or at the boundaries of segmentation.

[0066] For example, if the original in-plane voxel spacing is 0.176 mm, then the voxel spacing remains at 0.176 mm at 1x downsampling, adjusts to 0.352 mm at 2x downsampling, 0.528 mm at 3x downsampling, and 0.704 mm at 4x downsampling. Even after 4x downsampling, the in-plane resolution is still significantly higher than that of conventional CT (typically about 0.5-0.7 mm / pixel), thus effectively reducing the data size while retaining the high-resolution advantage of ultra-high-resolution target scanning.

[0067] In terms of interpolation method selection, cubic spline interpolation is used when resampling CT image data. This interpolation method is superior to bilinear interpolation in maintaining edge sharpness, which is especially important for preserving the blurred boundaries of ground-glass components in subsolid nodules. When resampling segmentation masks, nearest neighbor interpolation is used to avoid generating intermediate values ​​between 0 and 1, thus ensuring the binary characteristics of the mask.

[0068] IV. 3D Fine Segmentation Module The 3D fine segmentation module is configured to input the resampled region of interest output by the adaptive resampling module into a 3D fine segmentation network based on a dual-cascade feature pyramid (DPPN) architecture for voxel-level segmentation. This module is also one of the key technical features of this application. Unlike ordinary segmentation networks that directly output segmentation results through only one forward propagation, this application adopts a dual-cascade structure, linking coarse and fine segmentation in the same technical chain. This ensures that the second-stage segmentation is not a re-segmentation under unconstrained conditions, but rather utilizes the rough target location and contour information obtained in the first stage to further refine the boundary region. In other words, the first stage plays the role of "roughly identifying the nodules," while the second stage plays the role of "targeted refinement of the boundaries around the initial mask provided by the first stage." The two stages constitute a clear functional division and collaborative relationship.

[0069] The 3D fine-grained segmentation network has one input channel and eight base filters (nbase_filters). The network comprises two cascaded feature pyramid modules (FPN1 and FPN2) and two corresponding segmentation modules (Segmentation1 and Segmentation2), both constructed using 1×1 convolutional layers. The first and second feature pyramid modules are structurally consistent, each containing seven downsampling levels and six upsampling levels. The downsampling levels are constructed using an enhanced receptive field module (RfbeBlock3d) and residual basic blocks (ResBasicBlock3d, se=True) with a squeeze-and-excitation (SE) attention mechanism. The enhanced receptive field module expands the effective receptive field through a parallel structure of multi-scale convolutional branches, enhancing the ability to capture features at different scales. The SE attention mechanism adaptively weights the channel dimensions of the feature map, enabling the network to focus on feature channels relevant to the segmentation target and suppress responses from irrelevant channels. The upsampling layer is constructed using upsampling blocks (UpBlock3d), and a receptive field enhancement module (RfbBlock3d) is embedded in the mid-to-low-level feature output to enhance the feature extraction capability of the nodule edge region and suppress background noise interference.

[0070] Furthermore, the dual-cascaded FPN architecture introduces an attention gating mechanism at skip connections. In a conventional encoder-decoder architecture, skip connections directly pass shallow features from the encoder to the corresponding level of the decoder to supplement spatial details. However, shallow features also contain a large amount of background information unrelated to the segmentation target. The attention gating mechanism utilizes deep features from the decoder as a gating signal to calculate attention weights for each spatial location in the shallow features, adaptively suppressing feature responses from irrelevant background regions and highlighting nodule boundary information. This improves the adaptability of the segmentation network to complex adjacency scenarios (such as nodules adhering to the pleura or attached to blood vessels).

[0071] The specific execution process of the first segmentation stage is as follows: The first feature pyramid module FPN1 receives the resampled region of interest as input, extracts multi-scale features step by step through the encoder path, and then upsamples and fuses the features step by step through the decoder path. The first segmentation module Segmentation1 processes the mid-to-low-level feature outputs of FPN1 to generate a preliminary 3D probability map. The value of each voxel in the probability map is calculated using the Sigmoid activation function, representing the probability that the voxel belongs to a nodule. If the original network output of the i-th voxel is denoted as zi, its probability value pi can be expressed as: Where pi represents the predicted probability that the i-th voxel belongs to a nodule, zi represents the network output value of that voxel before activation, and e is the natural constant. This mapping can compress any real-valued output to the range of 0 to 1, facilitating subsequent thresholding. The preliminary 3D probability map is thresholded (preferably with a threshold of 0.5) to generate a preliminary segmentation mask. This preliminary segmentation mask provides approximate location and morphological information of the nodules, but inaccurate segmentation may occur in the nodule edge regions.

[0072] The second segmentation stage is executed as follows: the initial segmentation mask is expanded in three-dimensional space and then concatenated with the resampled region of interest in the feature channel dimension to form a multi-channel input. "Expanding in three-dimensional space" refers to appropriately expanding the positive region of the initial segmentation mask along three axes, so that the input of the second stage includes contextual information near the nodule mask boundary. The purpose is that the initial mask may underestimate the true boundary of nodules containing extremely low-contrast ground glass components; by appropriately expanding outwards, the "wide boundary band" containing potential edge gray-scale gradient regions is included in the processing range of the second stage. The expanded mask, as an additional channel, is concatenated with the original resampled region of interest. This multi-channel input simultaneously includes the original image information and the prior location information provided by the coarse segmentation, guiding FPN2 to focus on the nodule region determined by the initial segmentation and its surroundings, further optimizing the feature extraction accuracy of the edge region. The second segmentation module, Segmentation2, processes the output of FPN2 to generate the final three-dimensional probability map.

[0073] Therefore, the first and second stages are not simply two repetitive calculations, but rather form a substantial synergy in the input organization, feature focus, and boundary refinement path: the first stage provides global localization and rough outline, while the second stage focuses on edge detail optimization on this basis, so that the final segmentation result can not only ensure complete coverage of the main nodule region, but also achieve accurate characterization of complex boundary regions.

[0074] V. Post-processing and Feature Extraction Module The post-processing and feature extraction module is configured to post-process the final three-dimensional probability map output by the 3D fine segmentation module to generate the final binarized three-dimensional segmentation mask, and automatically calculate the quantitative features of the nodules based on the mask and the original CT value data.

[0075] The processing flow of this module is as follows: First, thresholding is performed on the final 3D probability map using a threshold of 0.5. Voxels with a probability value greater than or equal to 0.5 are marked as nodules (value 1), and the rest are marked as background (value 0), resulting in a preliminary binary mask. Then, morphological optimization is performed on this binary mask, including filling the holes inside the mask to eliminate possible internal cavities in the segmentation result; simultaneously, small connected regions with a 3D volume smaller than a preset threshold (e.g., 30 voxels) are removed to eliminate false positive regions caused by noise or interference. After optimization, the final binary 3D segmentation mask is obtained.

[0076] In terms of component separation, the post-processing and feature extraction module automatically classifies nodules into ground-glass components and solid components within the finally determined binarized 3D segmentation mask area, using -300 HU as a preset physical CT value threshold. Specifically, areas with original CT values ​​less than -300 HU are classified as ground-glass components (GGO), and areas with original CT values ​​greater than or equal to -300 HU are classified as solid components. This threshold is located at the boundary between the ground-glass density range (approximately -700 to -300 HU) and the soft tissue density range (approximately -300 to +50 HU), consistent with the prevailing standard in radiology for classifying subsolid nodule components.

[0077] Based on the final segmentation mask and component separation results, the quantitative features include at least the following: total nodule volume, calculated by multiplying the number of all positive voxels within the mask by the physical volume of a single voxel; solid component volume, calculated by multiplying the number of positive voxels within the mask with a CT value greater than or equal to -300 HU by the physical volume of a single voxel; average CT value, which is the arithmetic mean of the original CT values ​​of all positive voxels within the mask; and CT value histogram distribution. Thus, this application not only achieves boundary segmentation but also provides a structured data foundation for subsequent quantitative evaluation.

[0078] VI. Model Training Module The system in this embodiment also includes a model training module for training the 3D object detection network and the 3D fine segmentation network, respectively.

[0079] In training the 3D object detection network, complete lung field data with bounding box annotations were used as training samples. The loss function used for training was a weighted sum of Focal Loss and Smooth L1 Loss, with a weight ratio of 1:1. Focal Loss, as the classification loss, was used to address the problem of extreme imbalance between positive and negative samples in detection tasks (positive samples typically account for less than 1% of the total number of anchor points). The formula for calculating Focal Loss is: Where pt is the model's predicted probability of the correct class of the sample, αt is the balancing factor (set to 0.25 in this embodiment), and γ is the focusing parameter (set to 2 in this embodiment). When γ>0, the factor (1 - pt)γ suppresses the loss of easily classified samples (those with larger pt values), making the network training more focused on difficult-to-classify samples. Smooth L1 Loss is used as a regression loss for the regression optimization of the bounding box coordinates.

[0080] Positive samples are defined as anchors whose intersection-over-union (IoU) with the gold standard bounding box is greater than or equal to 0.5, and negative samples are defined as anchors whose IoU is less than 0.3. Other anchors are ignored during training. The ratio of positive to negative samples is controlled at 1:3 using the Online Hard Example Mining (OHEM) strategy, meaning that the negative sample with the largest loss in each mini-batch is selected to participate in backpropagation along with all positive samples.

[0081] In training a 3D fine-grained segmentation network, to prevent potential errors from the early stages of the detection network from propagating to the segmentation stage (i.e., error accumulation), the parameters of the already trained 3D object detection network are first frozen, and the regions of interest (ROIs) cropped and scaled from the gold standard bounding boxes are used as training input samples. The combined loss function L used for training the segmentation network is: Where Ldice is the standard Dice similarity coefficient loss, and Lce is the standard binary cross-entropy loss. The formula for calculating the Dice loss is: Where pi is the network's prediction probability for the i-th voxel, gi is the corresponding gold standard label (0 or 1), Σ represents the summation over all voxels, and ε is a smoothing term used to prevent the denominator from being zero. Dice loss primarily measures the consistency between the predicted mask and the ground truth mask in spatial overlap; its normalization properties make it insensitive to the ratio of positive to negative voxels. The formula for calculating the binary cross-entropy loss is: Where N is the total number of voxels. The binary cross-entropy loss measures the distance between the prediction and the label from the perspective of the difference in classification distribution at the voxel level. The combination of the two losses allows the network to be optimized in terms of both global morphological overlap and voxel-level boundary accuracy: the Dice loss takes into account the overall morphology, while the cross-entropy loss provides gradient signals at the voxel level.

[0082] Furthermore, both networks were trained using the Adam optimizer with an initial learning rate of 0.001 and a weight decay of 1×10⁻⁶. -4 The batch size for training the 3D object detection network is 4, with each batch containing 4 128-bit inputs. 3 Voxel blocks. The training process employs a cosine annealing learning rate decay strategy, smoothly reducing the learning rate from its initial value to a minimum of 1×10⁻⁶. -6 The maximum number of training epochs is 150. The early stopping criterion is: if the validation set loss does not decrease for 20 consecutive training epochs, training is terminated to prevent overfitting.

[0083] The data augmentation strategies introduced during the training phase of the 3D fine segmentation network include: random flipping along each of the three axes with a 50% probability; random rotation of ±10° around each of the three axes using trilinear interpolation; isotropic scaling with a scaling factor uniformly distributed in [0.9, 1.1]; elastic deformation using B-spline transformation to simulate slight changes in nodule morphology under different body positions and tissue deformation caused by respiratory movements; and random Gaussian noise with a standard deviation σ∈[0, 0.01] to simulate image quality fluctuations under different scanning conditions. These augmentation strategies work together to effectively increase the diversity of training data and improve the model's generalization ability.

[0084] Regarding annotation quality control, the gold standard bounding box and voxel-level annotations were obtained as follows: Two thoracic radiologists with over 10 years of experience in lung CT diagnosis independently performed layer-by-layer manual 3D segmentation. The Dice coefficient between the two radiologists' segmentation results was used to measure the consistency between the annotators, requiring a Dice coefficient of no less than 0.85. For cases with a Dice coefficient lower than 0.85, a third senior radiologist with over 15 years of experience arbitrated and determined the final gold standard segmentation. The training and validation sets were divided on a patient-by-patient basis to ensure that all nodules from the same patient appeared only in the training or validation set, thereby avoiding data leakage.

[0085] VII. Explanation of Technical Collaboration Relationship As can be seen from the descriptions of the modules above, the main innovation of this application does not originate from a single network module, but rather from the synergistic cooperation of multiple technical features surrounding a specific medical imaging scenario. First, the data characteristics of ultra-high resolution target scanning CT provide a potential foundation for fine boundary characterization, but also bring memory pressure and anisotropy issues. Second, the 3D target detection module first completes lesion localization, establishing a reliable basis for subsequent ROI-level adaptive resampling and eliminating the logical contradiction of "needing nodule size before detection." Third, the adaptive resampling module does not use fixed scaling, but performs differentiated processing based on the in-plane size of the target reflected by the detection bounding box, achieving a better balance between detail preservation and computational cost. Finally, the dual-cascaded 3D fine segmentation module improves the segmentation capability of complex boundary regions through a "coarse segmentation—mask prior guidance—fine segmentation" approach. These technical features form a closed processing chain, jointly serving to improve the accuracy and stability of automatic detection and 3D fine segmentation of subsolid pulmonary nodules.

[0086] Example 2: A fully automated lung nodule segmentation method based on ultra-high resolution target scanning CT See Figure 1 , Figure 2 , Figure 4 and Figure 5This embodiment provides a fully automated lung nodule segmentation method based on ultra-high resolution target scanning CT, with each step executed sequentially and collaboratively. This method can be implemented by the system described in Embodiment 1.

[0087] Step 100: Data Acquisition and Preprocessing. Acquire ultra-high resolution target scan CT images of the lungs with physical spacing attributes. The acquisition matrix of this image data should be at least 1024×1024, and the Z-axis slice thickness should not exceed 1.0 mm. Perform CT value intensity normalization based on a set window width and level, and morphological lung field extraction on the acquired image data to generate complete lung field volume data.

[0088] Step 110: Perform CT value intensity normalization based on the set window width and window level. Using the window width and window level range of [-1500 HU, 400 HU], the CT values ​​are linearly mapped to the range of [0, 1] according to Formula 1.

[0089] Step 120: Perform morphological lung field extraction. A -400 HU threshold segmentation method is used, combined with morphological closing operations (spherical structuring element, radius 3 voxels), connected component analysis (preserving the two largest 3D connected components), and dilation operations (spherical structuring element, radius 5 voxels), to generate complete lung field volume data.

[0090] Step 200: 3D Object Detection. The lung field volume data generated in Step 100 is input into an anchor-based 3D object detection network to automatically locate the spatial position of the lung nodules and output a 3D bounding box containing the lung nodules. The network structure and parameter settings of the 3D object detection network are consistent with those described in the 3D object detection module of Example 1. During the inference phase, the detection network processes the input lung field volume data and outputs classification confidence and bounding box regression values ​​for each anchor point position. After non-maximum suppression (NMS) processing, the final detection result is output. Since the detection step is completed before resampling, the nodule in-plane size information required in the subsequent step 300 can be directly obtained from the 3D bounding box output in this step, thus ensuring the fully automated nature of the entire process.

[0091] Step 300: Adaptive dynamic resampling. Based on the 3D bounding box output in step 200, the region of interest is cropped from the lung field volume data generated in step 100, and the maximum pixel size within the nodule plane determined by the 3D bounding box is obtained.

[0092] Step 310: Determine the downsampling factor. Based on the ratio of the maximum pixel size within the nodule plane to the preset clipping size of the model, determine the downsampling factor according to the determination rules described in the adaptive resampling module of Example 1.

[0093] Step 320: Perform resampling. Adaptive dynamic resampling is performed on the region of interest (ROI), maintaining the physical slice thickness along the Z-axis while downsampling along the X and Y axes by a predetermined downsampling factor. Cubic spline interpolation is used for the CT image, and nearest-neighbor interpolation is used for the segmentation mask (if applicable). The resampled ROI is then obtained.

[0094] Step 400: Two-stage 3D fine segmentation. The resampled region of interest obtained in step 300 is input into a 3D fine segmentation network based on a dual-cascaded feature pyramid network architecture for voxel-level segmentation.

[0095] Step 410, First Segmentation Stage (Coarse Segmentation). Multi-scale features of the resampled region of interest are extracted using the first feature pyramid module FPN1, and a preliminary 3D probability map is output by the first segmentation module Segmentation1. The probability value of each voxel belonging to a nodule is calculated using the sigmoid activation function shown in Equation 2, and a preliminary segmentation mask is generated after thresholding. The specific structure of the first feature pyramid module is consistent with that of the 3D fine segmentation module in Example 1.

[0096] Step 420, Second Segmentation Stage (Fine Segmentation). The preliminary segmentation mask generated in Step 410 is expanded in 3D space and then concatenated with the resampled region of interest obtained in Step 300 along the feature channel dimension to form a multi-channel input. Feature extraction is further optimized by the second feature pyramid module FPN2, and the final 3D probability map is output by the second segmentation module Segmentation2.

[0097] Step 500: Post-processing and quantitative feature calculation. The final 3D probability map output from Step 400 is post-processed to generate the final binarized 3D segmentation mask, and quantitative features are calculated.

[0098] Step 510, Thresholding and Morphological Optimization. The probability map is converted into a binary mask using a threshold of 0.5. Hole-filling is performed on the binary mask, and small connected components with volumes smaller than the preset threshold are removed. The final binary 3D segmentation mask is obtained.

[0099] Step 520: Component Separation and Feature Calculation. Based on the final binarized 3D segmentation mask and the original CT value data obtained in Step 100, the ground-glass component and solid component are separated using a preset physical CT value threshold of -300 HU. Quantitative features are calculated, including at least the total nodule volume, solid component volume, average CT value, and CT value histogram distribution.

[0100] This method executes steps 100 to 500 sequentially and collaboratively, forming an end-to-end fully automated processing flow from the input of the original CT image to the output of the three-dimensional segmentation mask and quantitative features.

[0101] Example 3: Electronic Devices and Storage Media This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a computing platform equipped with a graphics processing unit (GPU). The memory may be a combination of volatile memory and / or non-volatile memory. When the processor executes the program, it implements the steps of the method described in Embodiment 2.

[0102] This embodiment also provides a computer-readable storage medium on which a computer program is stored. The computer-readable storage medium includes, but is not limited to, a read-only memory, a random access memory, a magnetic disk, an optical disk, and flash memory. When the computer program is executed by a processor, it implements the steps of the method described in Embodiment 2.

[0103] Example 4: Clinical Validation To verify the effectiveness of the system proposed in this application, performance evaluation was conducted on an independent external test set. The validation data came from clinical centers different from the training data, and all cases were confirmed by surgical pathology, covering pulmonary subsolid nodules with different anatomical locations (including solitary nodules, nodules closely attached to the pleura, and nodules adhered to blood vessels) and different pathological subtypes.

[0104] The Dice similarity coefficient (DSC) is used as the primary evaluation metric to measure the spatial overlap between the automatic segmentation result and the gold standard segmentation. The Dice coefficient is calculated as follows: DSC = 2 × |A ∩ B| / (|A| + |B|) (Formula 7) Where A is the set of voxels from the automatic segmentation result, B is the set of voxels from the gold standard segmentation, |·| represents the number of voxels in the set, and ∩ represents the intersection of the sets. The DSC value ranges from [0, 1], and the larger the value, the higher the overlap between the two segmentation results. The 95% Hausdorff distance (HD95) is also calculated to evaluate the boundary error.

[0105] During the validation process, radiology experts were invited to make minimal corrections (correcting only obviously erroneous areas) based on the automatic segmentation results. The Dice coefficient between the automatic segmentation results and the expert-corrected segmentation was calculated to assess the gap between the automatic segmentation results and the clinically acceptable level. Furthermore, in some cases with both conventional resolution CT (512×512 matrix) and ultra-high resolution target scan CT (1024×1024 matrix), the same segmentation network was used to process the CT data at both resolutions, comparing the differences in segmentation performance at different resolutions. It should be noted that the above validation results are experimental data under specific test conditions. The technical effects of this application should be understood based on the overall technical solution of the system, and quantitative indicators obtained under specific conditions should not be taken as general commitments.

[0106] The internal structure of the receptive field enhancement module (RfbBlock3d) and the enhanced receptive field module (RfbeBlock3d) is illustrated below as an example.

[0107] The receptive field enhancement module (RfbBlock3d) employs a multi-branch parallel convolutional structure, comprising three parallel branches: the first branch uses a 1×1×1 convolution to capture point-level features; the second branch uses a 1×1×1 convolution followed by a 3×3×3 convolution (with a dilation rate of 1) to capture small-scale contextual features; and the third branch uses a 1×1×1 convolution followed by two convolutions, one with a dilation rate of 1 and the other with a dilation rate of 3, to capture large-scale contextual features. The outputs of the three branches are concatenated along the channel dimension, then fused using a 1×1×1 convolution, and residually connected to the input before being output via a ReLU activation function. The enhanced receptive field module (RfbBlock3d) is an enhancement of RfbBlock3d. The main differences are: a batch normalization layer and a ReLU activation function are added after the convolution of each branch, and the dilatation rate of the third branch is increased from 3 to 5 to obtain a wider range of contextual information; additionally, an extra 3×3×3 convolutional layer is added after channel fusion for feature refinement, before residual connection with the input. The term "enhanced" signifies that this module achieves a wider effective receptive field than the standard RfbBlock3d through a larger dilatation rate and deeper branching structure.

[0108] The following example illustrates the specific calculation method of the attention gating mechanism.

[0109] The specific implementation of the attention gating mechanism is as follows: Let the encoder features passed through skip connections be xl, and the deep features upsampled by the decoder be g. First, perform 1×1×1 convolutions on xl and g respectively to unify the number of channels. Then, add them element-wise and pass them through the ReLU activation function, followed by another 1×1×1 convolution and the Sigmoid activation function to generate an attention weight map α with values ​​ranging from [0, 1]. The calculation formula is as follows: Where Wx and Wg are the weight parameters of a 1×1×1 convolution, ψ is the 1×1×1 convolution, σ is the sigmoid function, and b is the bias term. The final output is... , where ⊙ denotes element-wise multiplication. This mechanism enables the decoder to adaptively select spatial regions in the encoder features that are relevant to the segmentation target based on semantic information, automatically amplify the feature values ​​of nodules and their boundary regions, and suppress irrelevant lung parenchyma background noise.

[0110] The channel compression ratio of the SE attention mechanism in this application is illustrated below by way of example.

[0111] In the SE attention mechanism, the channel compression ratio is set to 4. Specifically, the SE module first compresses each channel of the feature map into a scalar value using global average pooling, forming a channel description vector. Then, it passes through two fully connected layers. The first fully connected layer compresses the number of channels to 1 / 4 of the original number (i.e., a compression ratio of 4). After ReLU activation, the second fully connected layer restores the original number of channels, and the weight coefficients of each channel are output through the Sigmoid activation function. Finally, these weight coefficients are multiplied channel by channel of the original feature map to achieve adaptive weighting of the channel dimension.

[0112] The following example illustrates the specific operations and parameters for "expanding the initial segmentation mask in three-dimensional space".

[0113] The "expansion in three-dimensional space" operation specifically involves performing a three-dimensional morphological dilation operation on the positive regions marked as nodules (value 1) in the initial segmentation mask, using a spherical structuring element with a radius of 5 voxels. This dilation operation expands the boundaries of the initial segmentation mask outwards, incorporating the contextual information near the nodule boundaries into the processing scope of the second stage, thus avoiding the loss of edge information in the second stage due to potential undersegmentation in the first stage. The expanded mask, as an independent channel, is concatenated with the original resampled region of interest in the feature channel dimension, forming a 2-channel input (the first channel is CT image data, and the second channel is the expanded initial segmentation mask), which is then fed into the second feature pyramid module for fine segmentation feature extraction.

[0114] The following example illustrates the criteria for selecting the final result in double downsampling verification.

[0115] In the dual downsampling verification, the system performs a complete resampling and segmentation process on the same region of interest using two different downsampling factors, resulting in two sets of segmentation results. The selection strategy is as follows: after transforming the two sets of segmentation results back to the original resolution space, the connectivity and compactness indices of the positive regions in their respective output 3D probability maps are compared. Specifically, this includes the proportion of the largest connected component to the total positive voxels and the sphericity of the positive regions. The segmentation result with better connectivity and a sphericity closer to a reasonable range is selected as the final output. If the two sets of results do not differ significantly in the above indicators, the result corresponding to the smaller downsampling factor is selected first to retain more spatial details. This optimization process does not require the participation of a gold standard and can be automatically executed during the inference stage.

[0116] The following example illustrates the specific values ​​of the model's preset cutting dimensions and their applicable stages.

[0117] In this application, the correspondence between the preset model clipping size and different stages is as follows: In the adaptive resampling stage, the "preset model clipping size" used to determine the XY-axis downsampling factor refers to the number of voxels in the XY direction of the standard input of the segmentation network, set to 256 voxels. The actual input volume size in the segmentation network training stage can be adjusted to configurations such as [48, 256, 256] or [64, 320, 320] depending on the downsampling factor. The Z-axis dimension is dynamically adapted according to the number of nodal layers, and the XY-axis dimension matches the ROI size after adaptive resampling. The inference stage maintains the same size setting as the training stage, aligning to the same size through resampling and padding.

[0118] The following example illustrates the segmentation effect in complex scenarios.

[0119] In cases of pleural adhesion nodules, the first segmentation stage provides an initial segmentation of the main nodule region. The second segmentation stage, combining the expanded initial mask and original grayscale information, further distinguishes the boundary closest to the pleura. Attention gating helps suppress interference from linear features on the pleural surface, thereby reducing the possibility of misclassifying pleural tissue as part of the nodule. In cases of vascular adhesion nodules, the enhanced receptive field module and multi-scale feature fusion help distinguish nodule tissue from slender vascular structures. Attention gating further suppresses feature responses in the vascular region at skip connections, making the segmentation boundary more closely match the true contour of the nodule rather than excessively extending along the vascular direction.

[0120] Technical effect The fully automated lung nodule segmentation system and method based on ultra-high resolution target scanning CT provided in the above embodiments, through a series of interconnected technical links consisting of a data preprocessing module, a 3D target detection module, an adaptive resampling module, a 3D fine segmentation module, and a post-processing and feature extraction module, can effectively improve the accuracy and stability of automatic detection and three-dimensional segmentation of subsolid lung nodules from multiple levels.

[0121] At the data preprocessing level, CT value intensity standardization based on the window width and window level range of [-1500 HU, 400 HU] covers the complete range from ground-glass opacity to soft tissue density, avoiding the loss of image information that is crucial for distinguishing subsolid nodule components during grayscale mapping. In morphological lung field extraction, the sequential combination of closing operation, connected component analysis and dilation operation removes irrelevant tissue outside the pleural cavity while ensuring that nodules located under the pleura are not clipped due to overly tight mask boundaries, providing a complete and accurate input volume for subsequent detection and segmentation.

[0122] At the collaborative level of detection and resampling, the 3D target detection module employs a feature pyramid network structure containing 8 downsampling levels and 7 upsampling levels. Combined with the receptive field enhancement module (RfbBlock3d) embedded in the detection head and multi-anchor point settings, this enhances the detection capability for subsolid nodules of different sizes. More importantly, the 3D bounding box output by the detection network simultaneously provides the adaptive resampling module with prior information on the maximum in-plane pixel size of the nodule. This allows the determination of the resampling factor to be completed automatically without manual intervention, ensuring the fully automated nature of the entire system's processing flow. The adaptive resampling module maintains a constant physical slice thickness along the Z-axis and employs an asymmetric strategy of downsampling only along the XY axes at a predetermined factor. This effectively alleviates the inherent anisotropy problem of ultra-high resolution target scanning CT data, preserving significantly higher in-plane spatial details than conventional CT while keeping the data size within a computable range.

[0123] At the segmentation level, the 3D fine segmentation network based on a dual-cascaded feature pyramid network architecture outputs a preliminary segmentation mask in the first segmentation stage. This mask is then expanded in 3D space and concatenated with the resampled region of interest along the channel dimension before being fed into the second segmentation stage for fine segmentation. This allows the feature extraction in the second stage to focus on the nodule region and its surrounding edge gradient zone, which has been preliminarily identified in the first stage. Combined with the enhanced receptive field module (RfbeBlock3d) in the downsampling level, the residual basic blocks with SE attention mechanism, and the attention gating mechanism at skip connections, this helps suppress feature interference from irrelevant background tissues in complex adjacent scenarios such as those closely adhering to the pleura or adhering to blood vessels, improving the segmentation accuracy of the nodule boundary region. After thresholding and morphological optimization to generate the final binarized 3D segmentation mask, the post-processing and feature extraction module further automatically separates the ground-glass component from the solid component using physical CT value thresholds and calculates quantitative features including the total nodule volume, solid component volume, average CT value, and CT value histogram, providing structured data support for subsequent clinical quantitative analysis. In summary, the various technical aspects of the above embodiments work together organically, enabling the system to achieve high-precision three-dimensional segmentation of subsolid lung nodules in ultra-high resolution target scan CT images in a fully automated end-to-end process.

[0124] The physical meaning of the positioning anchor point and classification anchor point parameters are illustrated below by example.

[0125] The positioning anchor point is set to [3, 3], meaning that the bounding box parameter dimension predicted by the positioning branch of the detection head at each spatial location is 3×3, corresponding to the number of regression prediction channels for the 3D bounding box center coordinate offset (Δx, Δy, Δz) and bounding box size (dx, dy, dz). The classification anchor point is set to [1, 1, 1, 1], meaning that the classification branch outputs 4 classification channels at each anchor point position, each corresponding to a different prediction dimension.

[0126] Furthermore, in this application, the mid-to-low-level features refer to the features output by several layers with higher spatial resolution (i.e., layers closer to the network output) in the upsampling path of the feature pyramid network. Taking a feature pyramid module containing 6 upsampling layers as an example, the mid-to-low-level features correspond to the feature outputs of the 1st to 4th upsampling layers counting from the output. These layers have relatively high spatial resolution and can preserve the spatial details of the nodule edge region.

[0127] Furthermore, for example, regarding the "sphericity" index in dual downsampling verification, the sphericity is calculated as follows: Where V is the volume (voxels) of the three-dimensional connected domain, and A is the surface area of ​​the three-dimensional connected domain (estimated by the marchingcubes algorithm or voxel count method). The sphericity value ranges from (0, 1], with values ​​closer to 1 indicating a shape closer to a sphere. In the selection strategy for double downsampling validation, "sphericity closer to a reasonable range" means that the sphericity value is within the interval [0.3, 0.9], which covers the morphological characteristics of the vast majority of subsolid pulmonary nodules in clinical practice.

[0128] It should be noted that in this patent application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this patent application, if it refers to performing an action according to an element, it means performing the action at least according to that element, including two cases: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "repeatedly," and "various" include two, two times, two kinds, and more than two, more than two times, and more than two kinds.

[0129] All documents mentioned in this application are considered to be incorporated in their entirety into the disclosure of this application so that they can serve as a basis for modifications if necessary. Furthermore, it should be understood that after reading the foregoing disclosure of this application, those skilled in the art can make various alterations or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A fully automated lung nodule segmentation system based on ultra-high resolution target scanning CT, characterized in that, include: The data preprocessing module is configured to acquire ultra-high resolution target scan CT image data of the lungs with physical spacing attributes, perform CT value intensity standardization based on set window width and window level and morphological lung field extraction, and generate complete lung field volume data. The 3D target detection module is configured to input the lung field volume data into an anchor-based 3D target detection network, locate the spatial position of the lung nodules, and output a three-dimensional bounding box containing the lung nodules. An adaptive resampling module is configured to crop the region of interest from the lung field volume data based on the three-dimensional bounding box, and obtain the maximum pixel size within the nodule plane determined by the three-dimensional bounding box; determine the downsampling factor based on the ratio of the maximum pixel size within the nodule plane to the preset cropping size of the model, and perform adaptive dynamic resampling on the region of interest. The adaptive dynamic resampling maintains the physical layer thickness along the Z-axis unchanged, and performs adaptive downsampling along the XY-axis according to the downsampling factor to obtain the resampled region of interest. The 3D fine segmentation module is configured to input the resampled region of interest into a 3D fine segmentation network based on a dual-cascaded feature pyramid network architecture for voxel-level segmentation; the first segmentation stage of the 3D fine segmentation network extracts coarse segmentation features and outputs a preliminary 3D probability map and a preliminary segmentation mask; the second segmentation stage of the 3D fine segmentation network expands the preliminary segmentation mask in 3D space, concatenates it with the resampled region of interest in the feature channel dimension as input, extracts fine segmentation features, and outputs the final 3D probability map. The post-processing and feature extraction module is configured to perform thresholding and morphological optimization on the final three-dimensional probability map, generate the final binary three-dimensional segmentation mask, and separate nodule components and calculate quantitative features based on the three-dimensional segmentation mask and a preset physical CT value threshold.

2. The system according to claim 1, characterized in that, In the adaptive resampling module, the rule for determining the downsampling factor based on the ratio of the maximum pixel size within the nodule surface to the preset clipping size of the model is as follows: When the maximum pixel size within the nodule surface is less than or equal to the preset clipping size of the model, a 1x downsampling is used; When the maximum pixel size within the nodule surface is greater than the preset clipping size of the model but less than or equal to 1.5 times the preset clipping size of the model, a 2x downsampling is used; When the maximum pixel size within the nodule surface is greater than 1.5 times the preset clipping size of the model and less than or equal to 2 times the preset clipping size of the model, verification is performed using 2x and 3x double downsampling respectively; When the maximum pixel size within the nodule plane is greater than twice the preset clipping size of the model, verification is performed using 3x and 4x double downsampling, respectively. The adaptive dynamic resampling uses cubic spline interpolation when interpolating CT images and nearest neighbor interpolation when resampling segmentation masks.

3. The system according to claim 1, characterized in that, The anchor-based 3D target detection network has 1 input channel, 16 basic filters, and a detection range of subsolid nodules with a diameter of 4-30 mm. The 3D target detection network includes a feature pyramid network structure, which contains 8 downsampling levels and 7 upsampling levels. The downsampling levels are constructed using convolutional blocks, residual basic blocks, and max pooling layers, while the upsampling levels are constructed using upsampling blocks. The 3D target detection network also includes a detection head module, which embeds a receptive field enhancement module for fusing and optimizing multi-scale features, predicting the coordinates of the three-dimensional bounding box through a localization head, and determining the nodule category through a classification head; the localization anchor point corresponding to the localization head is set to [3, 3], and the classification anchor point corresponding to the classification head is set to [1, 1, 1, 1].

4. The system according to claim 1, characterized in that, The 3D fine segmentation network based on the dual-cascaded feature pyramid network architecture has 1 input channel and 8 basic filters. The 3D fine segmentation network comprises two cascaded feature pyramid modules and two corresponding segmentation modules constructed using 1×1 convolutional layers, wherein: The first feature pyramid module corresponding to the first segmentation stage and the second feature pyramid module corresponding to the second segmentation stage both contain 7 downsampling levels and 6 upsampling levels; the downsampling levels are constructed using an enhanced receptive field module and residual basic blocks with a squeezing and excitation SE attention mechanism; the upsampling levels are constructed using upsampling blocks, and a receptive field enhancement module is embedded at the mid-to-low-level feature output end to enhance the feature extraction capability of the nodule edge region; The dual-cascaded feature pyramid network architecture introduces an attention gating mechanism at skip connections to adaptively suppress feature responses from irrelevant background regions in order to highlight nodule boundary information.

5. The system according to claim 1, characterized in that, The system also includes a model training module, configured as follows: For training the anchor-based 3D object detection network, a weighted sum of Focal Loss and Smooth L1 Loss is used as the loss function, with a weight ratio of 1:

1. Focal Loss is used as the classification loss, with a focusing parameter γ set to 2 and a balance factor α set to 0.

25. Smooth L1 Loss is used as the regression loss. Positive samples are defined as IoU ≥ 0.5 between the anchor point and the gold standard bounding box, and negative samples are defined as IoU < 0.3, with the rest ignored. The ratio of positive to negative samples is controlled to 1:3 through an online hard example mining strategy. The parameters of the anchor-based 3D object detection network are frozen, and the region of interest cropped and scaled by the gold standard bounding box is used as input sample to train the 3D fine segmentation network to avoid the propagation of detection errors. The combined loss function L used for training the 3D fine segmentation network is: in, The standard Dice similarity coefficient loss is used to measure the spatial morphological overlap between the segmentation mask predicted by the model and the ground truth labeled mask. This is the standard binary cross-entropy loss, used to measure the difference in classification distribution at the voxel level.

6. A fully automated lung nodule segmentation method based on ultra-high resolution target scanning CT, characterized in that, The steps are executed sequentially and in a coordinated manner, including the following steps: Step S1: Acquire ultra-high resolution target scan CT image data of the lungs with physical spacing attributes, perform CT value intensity normalization based on set window width and window level and morphological lung field extraction to generate complete lung field volume data; Step S2: Input the lung field volume data generated in step S1 into an anchor-based 3D target detection network to locate the spatial position of the lung nodules and output a three-dimensional bounding box containing the lung nodules. Step S3: Based on the 3D bounding box output in step S2, crop the region of interest from the lung field volume data generated in step S1, and obtain the maximum pixel size within the nodule plane determined by the 3D bounding box; determine the downsampling factor based on the ratio of the maximum pixel size within the nodule plane to the preset cropping size of the model, and perform adaptive dynamic resampling on the region of interest. The adaptive dynamic resampling maintains the physical layer thickness on the Z-axis unchanged, and performs adaptive downsampling on the XY-axis according to the downsampling factor to obtain the resampled region of interest. Step S4: Input the resampled region of interest obtained in Step S3 into a 3D fine segmentation network based on a dual-cascaded feature pyramid network architecture for voxel-level segmentation; the first segmentation stage of the 3D fine segmentation network extracts coarse segmentation features and outputs a preliminary 3D probability map and a preliminary segmentation mask; the second segmentation stage of the 3D fine segmentation network expands the preliminary segmentation mask in 3D space and concatenates it with the resampled region of interest obtained in Step S3 in the feature channel dimension as input to extract fine segmentation features and output the final 3D probability map; Step S5: Threshold and morphologically optimize the final three-dimensional probability map output in step S4 to generate the final binary three-dimensional segmentation mask. Based on the three-dimensional segmentation mask and the original CT value data obtained in step S1, separate the nodule components and calculate quantitative features using a preset physical CT value threshold.

7. An electronic 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 program, it implements the steps of the method as described in claim 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in claim 6.

Images