A lung ground glass nodule segmentation method based on 3D residual U-Net multi-task learning

Abstract

The application belongs to the field of medical image analysis, and provides a lung CT ground glass nodule segmentation and extraction method based on 3D residual U-Net and multi-task collaborative learning. A low-dose spiral CT sequence is collected, lung region pre-segmentation, resampling standardization, gray scale truncation mapping and enhancement are performed according to the lung window, and reference data is constructed by layer-by-layer labeling by two doctors; a 3D residual network containing three-dimensional encoding, decoding and skip connection is built by cutting the body block centered on the nodule; a multi-task collaborative module for voxel segmentation, benign and malignant classification and boundary distance regression is designed, and the same variance uncertainty automatic weighted Dice, cross entropy and distance regression loss is trained to obtain the MT-3DResUNet fusion model. The inference result is analyzed by three-dimensional connected domain analysis and morphological post-processing, the nodule image is output, the volume, diameter, CT value and follow-up data are counted, the segmentation and extraction and quantitative representation of various ground glass nodules are realized, and the early screening and follow-up management of lung cancer are assisted.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, specifically to a method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning. Background Technology

[0002] Ground-glass opacities (GGOs) in the lungs are one of the main manifestations of early-stage lung adenocarcinoma and its precancerous lesions on CT imaging. The morphology, size, borders, and internal solid component of the nodules directly determine the risk assessment of benign or malignant transformation and the formulation of clinical follow-up strategies. GGOs appear as slightly increased density, cloud-like shadows on CT images, characterized by low contrast, blurred borders, irregular shapes, and easy adhesion to blood vessels and pleural structures, posing a significant challenge to accurate segmentation and quantitative characterization of the nodules. Accurate segmentation and quantitative characterization of pulmonary GGOs are of significant clinical importance and application value for early lung cancer screening, differentiation of benign and malignant nodules, and follow-up management. In traditional methods of lung nodule analysis, radiologists typically use manual image reading combined with semi-automatic delineation tools for nodule marking and measurement, which suffers from low efficiency, high subjectivity, and poor inter-physician consistency. In recent years, researchers have successively introduced traditional image processing techniques such as thresholding, region growing, level set, and morphological analysis for nodule extraction. However, these methods lack robustness against interferences such as low contrast, blurred boundaries, and adhesion to blood vessels in ground-glass nodules, making it difficult to achieve automated and high-precision nodule segmentation. With the development of deep learning technology, semantic segmentation networks such as 2D U-Net, V-Net, and nnU-Net have achieved significant results in the field of medical image segmentation and have been gradually introduced into lung nodule analysis. However, ground-glass nodules have characteristics such as strong three-dimensional spatial continuity, incomplete axial single-layer information, and blurred boundaries. Segmentation networks based on two-dimensional slices still have the following shortcomings: (1) Layer-by-layer two-dimensional segmentation lacks inter-layer three-dimensional context information, resulting in discontinuity of nodule contours between layers and omission of small nodules; (2) Single segmentation tasks do not utilize high-level semantic information such as benign or malignant nodules, limiting feature expression capabilities; (3) The loss function does not adequately consider the distinction between blurred pixels at ground-glass boundaries and pixels in normal lung tissue, easily leading to overly smooth boundaries or over-segmentation. Therefore, there is an urgent need for an automatic segmentation and extraction method that is tailored to the characteristics of ground-glass nodules and has the ability to perform three-dimensional context modeling and multi-task collaborative expression. Summary of the Invention

[0003] To address the technical challenges of automated and high-precision segmentation and extraction of pulmonary ground-glass nodules in CT images, such as low contrast, blurred boundaries, adhesion to blood vessels and pleura, and discontinuity of inter-slice information, existing technologies cannot overcome these problems. This invention aims to provide a method for segmenting and extracting pulmonary ground-glass nodules based on 3D residual U-Net and multi-task collaborative learning. The specific technical solution is as follows: A method for segmenting and extracting pulmonary ground-glass nodules based on 3D residual U-Net and multi-task collaborative learning, comprising: acquiring low-dose spiral CT chest scan sequences and performing pre-segmentation and resampling preprocessing of the lung region to obtain standardized CT volume data; performing grayscale normalization and data enhancement on the CT volume data according to the window width and level of the lung windows; manually annotating the enhanced volume data layer by layer along the nodule boundary to obtain a reference volume dataset; and designing a three-dimensional feature encoding module based on 3D residual convolution to convert the input volume data into tensors to extract the nodules. Three-dimensional contextual features; a three-dimensional feature decoding module based on 3D deconvolution is designed to restore volumetric data resolution; a multi-task collaborative learning module is designed, including voxel segmentation subtask, nodule benign / malignant classification subtask, and boundary distance regression subtask; skip connections are added between corresponding layers of the encoding and decoding modules to form the MT-3DResUNet fusion model; a multi-task uncertainty weighted loss function is designed based on the characteristics of blurred boundaries and sparse foreground voxels of ground-glass nodules to obtain the loss function weights for each subtask; multi-task collaborative supervision calculation is performed based on the loss function weights to obtain model feedback; the MT-3DResUNet fusion model is trained to obtain a trained model; the trained model is used to segment and extract ground-glass nodules from lung CT volumetric data; three-dimensional connected component analysis and morphological post-processing are performed on the segmentation and extraction results to obtain nodule image results; nodule volume, diameter, CT value statistics and follow-up data analysis are performed on the nodule image results.

[0004] Furthermore, the CT data acquisition and preprocessing method includes: acquiring a low-dose spiral CT chest scan sequence with a slice thickness of no more than 1.25 mm, a scanning tube voltage of 120 kV, a tube current of 30-50 mAs, and a reconstruction matrix of 512×512; pre-segmenting the lung region of the CT sequence, extracting the left and right lung masks and removing the trachea and main bronchus using a thresholding method combined with morphological operations; and resampling the CT sequence to an isotropic voxel spacing of 1×1×1 mm to obtain standardized CT volume data.

[0005] Furthermore, the data augmentation and annotation method includes: truncating and linearly mapping the CT volume data to a lung window width of 1500 HU and a window level of -600 HU, with the normalization formula as follows: I_norm(x,y,z) = (I_in(x,y,z) -W_l) / W_w, where I_norm(x,y,z) represents the normalized voxel gray value; I_in(x,y,z) represents the CT value of the volume data at (x,y,z); W_l represents the lower limit of the window level, with a value of -1350HU; W_w represents the window width, with a value of 1500HU; the truncated and mapped voxel values ​​are linearly scaled to the [0,1] interval; the volume data is manually annotated layer by layer along the nodule boundary, independently annotated by two thoracic radiology physicians and reviewed and confirmed by senior physicians to generate reference volume data; a 64×128×128 volume data block is cropped with the nodule as the center and combined with random flipping and elastic deformation to expand the data to obtain the standard dataset.

[0006] Furthermore, the method for obtaining the MT-3DResUNet fusion model includes: reshaping the input volume data block into a 64×128×128×1 tensor; the three-dimensional feature encoding module consists of four encoding stages, each stage containing two three-dimensional residual convolutional blocks and one three-dimensional downsampling convolution with a stride of 2×2×2, extracting three-dimensional contextual features of 16, 32, 64, and 128 channels stage by stage; the three-dimensional feature decoding module consists of four decoding stages, each stage containing three-dimensional deconvolution upsampling and three-dimensional residual convolutional blocks, gradually restoring the volume data resolution. Skip connections are added between corresponding layers of the encoding and decoding modules to concatenate the encoded and decoded features along the channel dimension. A multi-task collaborative learning module is designed, in which the voxel segmentation subtask outputs a voxel probability map of nodules by performing 1×1×1 convolution and sigmoid activation on the decoded features; the nodule benign / malignant classification subtask outputs benign / malignant probabilities by performing 3D global average pooling and fully connected layers on the deepest encoded features; and the boundary distance regression subtask outputs a distance map from voxels to nodule boundaries by performing 1×1×1 convolution on the decoded features. The MT-3DResUNet fusion model is obtained by fusing these features.

[0007] Furthermore, the method for obtaining the multi-task uncertainty weighted loss function includes: addressing the problems of blurred boundaries of ground-glass nodules, sparse foreground voxels, and large differences in the magnitude of multi-task events, introducing an automatic multi-task weighting mechanism based on homoscedastic uncertainty to obtain the multi-task uncertainty weighted loss function, calculated as follows: L_seg = 1 - 2TP / (2TP + FP + FN), L_cls = -[y·log(p) + (1-y)·log(1-p)], L_dist = MAE(D_pred,D_gt), L_total = Σ_k (1 / (2σ_k^2))·L_k + log(σ_k), k∈{seg, cls, In the formula, L_seg is the Dice loss for nodule voxels; TP is the number of true positive voxels in the nodule region; FP is the number of false positive voxels in lung tissue misclassified as nodules; FN is the number of false negative voxels in nodules misclassified as lung tissue; L_cls is the cross-entropy loss for benign / malignant classification; y is the ground truth label for benign / malignant; p is the predicted probability of malignancy; L_dist is the mean absolute error loss for boundary distance regression; D_pred is the predicted distance map; D_gt is the ground truth distance map generated by the ground truth mask distance transformation; σ_k is the learnable homoscedasticity uncertainty parameter for each subtask; the loss function is used to perform multi-task collaborative supervision calculation on the MT-3DResUNet fusion model to obtain model feedback.

[0008] Furthermore, the model training method includes: using voxels of the nodule region as foreground targets and voxels of the normal lung tissue region as background; initializing the three-dimensional feature encoding module with three-dimensional convolutional weights pre-trained on a public lung nodule dataset and performing transfer learning; during training, first freezing the shallow stage of the encoding module for warm-up training, and then unfreezing all parameters for fine-tuning; setting the initial learning rate to 3×10^-4 and dynamically adjusting it according to the cosine annealing strategy; setting the batch size to 2 and iterating for no less than 200 epochs; stopping training when the Dice coefficient on the validation set no longer increases for 20 consecutive epochs and saving the optimal model weights.

[0009] Furthermore, the three-dimensional connected component analysis and morphological post-processing method includes: binarizing the nodule voxel probability map output by the model according to a preset threshold; filling the small cavities inside the nodule and smoothing the nodule surface using three-dimensional morphological closing operations; marking each nodule region according to the three-dimensional connected component analysis and removing pseudo-nodule regions with a volume smaller than a preset threshold; calculating the volume, equivalent diameter, maximum diameter, average CT value, and solid component ratio of each nodule; mapping the centroid of the nodule to the anatomical region of the lung lobe, and calculating the nodule volume doubling time in conjunction with follow-up CT scans to obtain the final nodule image results and statistical parameters.

[0010] This invention offers the following superior advantages: By acquiring thin-slice, low-dose spiral CT sequences and performing lung region pre-segmentation and resampling, isotropic standardized CT volumetric data is obtained. This volumetric data possesses rich nodule morphology and scale variations, providing ample samples for deep learning models. Due to the large grayscale dynamic range of CT volumetric data, lung window width and window level truncation mapping is used to augment the volumetric data, improving model generalization ability while enriching data diversity. The volumetric data is manually annotated layer by layer along nodule boundaries and reviewed by two physicians to obtain high-quality reference volumetric data. The volumetric data undergoes nodule-centered cropping and normalization to reduce computational load and improve training efficiency, resulting in a standard dataset. Because ground-glass nodules exhibit strong three-dimensional spatial continuity and blurred boundaries, a three-dimensional residual convolution is used to construct a three-dimensional feature encoding module, fully utilizing inter-slice three-dimensional contextual information and avoiding inter-slice discontinuities inherent in two-dimensional layer-by-layer segmentation. The problem is as follows: A 3D feature decoding module is constructed using 3D deconvolution and skip connections to gradually restore nodule details; a multi-task collaborative learning module is designed, with voxel segmentation as the main task and nodule benign / malignant classification and boundary distance regression as collaborative sub-tasks, utilizing high-level semantics of benign / malignant and boundary geometric constraints to enhance feature representation capabilities; due to the large differences in the scale and difficulty of the multiple tasks, an automatic weighting mechanism based on homoscedastic uncertainty is introduced to design a multi-task uncertainty weighted loss function to avoid task imbalance caused by manual parameter tuning, and to improve boundary localization capabilities without sacrificing segmentation accuracy; the trained model is used to segment lung CT body data; since the model segmentation results may contain small noise and holes, 3D morphological closing operations and 3D connected component analysis are used for post-processing to obtain complete nodule image results; finally, statistical data analysis is performed on the nodule images for volume, diameter, CT value, proportion of solid components, and volume doubling time. This invention fully considers the complexity of ground-glass nodules in terms of three-dimensional morphology, boundaries, and grayscale. It designs unique three-dimensional residual coding, three-dimensional deconvolution decoding, and multi-task collaborative learning modules, which enhance the three-dimensional contextual expression ability and multi-task collaborative representation ability of deep learning segmentation models. This improves the completeness, inter-layer continuity, and boundary accuracy of nodule segmentation, providing a reliable technical means for early lung cancer screening and nodule follow-up management. Attached Figure Description

[0011] To more clearly illustrate the solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a flowchart of a lung CT ground-glass nodule segmentation and extraction method based on 3D residual U-Net and multi-task collaborative learning, provided by an embodiment of the present invention. Figure 2 This is a network structure diagram of the MT-3DResUNet fusion model provided in one embodiment of the present invention. Detailed Implementation

[0012] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a lung CT ground-glass nodule segmentation and extraction method based on 3D residual U-Net and multi-task collaborative learning proposed according to the present invention. In the following description, different embodiments or one embodiment do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The specific scheme of the lung CT ground-glass nodule segmentation and extraction method based on 3D residual U-Net and multi-task collaborative learning provided by the present invention is described in detail below with reference to the accompanying drawings.

[0013] Please see Figure 1 This invention illustrates a method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning, which includes the following steps.

[0014] Step S1: Acquire low-dose spiral CT chest scan sequences and perform lung region pre-segmentation and resampling preprocessing to obtain standardized CT volume data. This embodiment of the invention is mainly applied to the segmentation of ground-glass nodules in low-dose chest CT screening. To obtain ground-glass nodule volume data samples, thin-slice CT scan sequences are first acquired, and standardized volume data is obtained after lung region pre-segmentation and resampling. Since this embodiment of the invention uses a deep learning model to segment the nodules, the model network requires a large number of volume data samples containing nodule morphology to ensure that the model has sufficient generalization ability. Preferably, in one embodiment of the present invention, the CT data acquisition and preprocessing method includes: acquiring a low-dose spiral CT chest scan sequence with a slice thickness of 1.0 mm, a scanning tube voltage of 120 kV, a tube current of 40 mAs, a reconstruction matrix of 512×512, and 300-400 reconstruction layers; extracting the left and right lung masks from the CT sequence using the -600 HU threshold method combined with three-dimensional morphological opening and closing operations, removing the trachea and main bronchus using region connectivity analysis, and repairing the convex bulge of the lung mask boundary; resampling the CT sequence to a 1×1×1 mm isotropic voxel spacing using trilinear interpolation to obtain standardized CT volume data; collecting no less than 500 CT cases containing ground-glass nodules to ensure the diversity and representativeness of the dataset.

[0015] Step S2: Perform grayscale normalization on the CT volume data based on the lung window width and level; manually annotate the enhanced volume data layer by layer along the nodule boundary to obtain a reference volume dataset. Because the acquired CT volume data has a large grayscale dynamic range and low contrast between nodules and lung tissue, lung window truncation mapping is used to perform grayscale normalization enhancement on the volume data, improving the model's generalization ability while enriching data diversity. Since the CT value distribution between the nodule region and the normal lung tissue region differs significantly, voxels are manually annotated layer by layer along the nodule boundary. Since nodule segmentation and extraction only requires focusing on the relative grayscale differences within the lung window, the volume data is normalized to reduce computation and improve training efficiency, obtaining a standard dataset. Preferably, in one embodiment of the present invention, the data augmentation and annotation method includes: performing grayscale truncation and linear mapping on the CT volume data according to the lung window width of 1500 HU and the window level of -600 HU, and linearly scaling the voxel values ​​in the [-1350, 150] HU interval to the [0,1] interval; performing layer-by-layer manual annotation on the volume data along the nodule boundary, which is independently annotated by two thoracic imaging diagnostic physicians with more than 5 years of experience, and the divergence points are reviewed and confirmed by senior physicians, marking the nodule voxels as 1 and the background voxels as 0 to generate reference volume data; cropping a 64×128×128 volume data block centered on the nodule, and combining random three-axis flipping, ±15° random rotation and elastic deformation to expand the dataset to 8-10 times the original size; and randomly dividing it into training set, validation set and test set in a ratio of 8:1:1.

[0016] Step S3: Design a 3D feature encoding module based on 3D residual convolution; design a 3D feature decoding module based on 3D deconvolution; design a multi-task collaborative learning module; and obtain the MT-3DResUNet fusion model through skip connections. In one embodiment of the present invention, the method for obtaining the MT-3DResUNet fusion model includes: extracting 3D contextual features of nodules using a 3D feature encoding module; restoring the volumetric data resolution using a 3D feature decoding module; and jointly completing voxel segmentation, benign / malignant classification, and boundary distance regression through a multi-task collaborative learning module. Preferably, in one embodiment of the present invention, the design method of the three-dimensional feature encoding module includes: the input layer reshaping the input volume data block into a tensor of 64×128×128×1; the encoding module consists of 4 encoding stages, each stage containing 2 three-dimensional residual convolutional blocks, the residual blocks being composed of two 3×3×3 convolutions, batch normalization and ReLU activation concatenated and added to the input identity, the stages using 2×2×2 stride three-dimensional convolution downsampling, extracting 16, 32, 64 and 128 channels of three-dimensional features stage by stage, the deepest stage outputting 4×8×8×128 encoded features. Preferably, in one embodiment of the present invention, the design method of the three-dimensional feature decoding module and the multi-task collaborative learning module includes: the decoding module consists of four decoding stages, each stage containing 2×2×2 three-dimensional deconvolution upsampling and two three-dimensional residual convolution blocks, and a skip connection is added between the corresponding layers of the encoding module and the decoding module to concatenate the encoded features and the decoded features in the channel dimension; the voxel segmentation subtask outputs a 64×128×128 nodule voxel probability map by performing a 1×1×1 convolution and sigmoid activation on the final decoded features; the nodule benign and malignant classification subtask outputs benign and malignant probabilities by performing three-dimensional global average pooling, a 128-dimensional fully connected layer and Softmax activation on the deepest encoded features; the boundary distance regression subtask outputs a distance map from the voxel to the nodule boundary by performing a 1×1×1 convolution on the decoded features, and the MT-3DResUNet fusion model is obtained by fusing the data.

[0017] Step S4: Based on the characteristics of ground-glass nodules—blurred boundaries and sparse foreground voxels—a multi-task uncertainty weighted loss function is designed to obtain the loss function weights for each sub-task. Multi-task collaborative supervision calculations are then performed based on these weights to obtain model feedback. The grayscale and structural information of the nodule region and normal lung tissue region are obtained through the above process. Because ground-glass nodules have blurred boundaries, low foreground voxel proportions, and large differences in magnitude across multiple tasks, a multi-task uncertainty weighted loss function is designed and calculated based on the aforementioned characteristics to obtain model feedback. Preferably, in one embodiment of the present invention, the model feedback acquisition method includes: obtaining the loss weights of each subtask according to the multi-task uncertainty weighted loss function calculation formula, and calculating the model feedback. The calculation formula is as follows: L_seg = 1 - 2TP / (2TP + FP + FN), L_cls = -[y·log(p) + (1-y)·log(1-p)], L_dist = MAE(D_pred, D_gt), L_total = Σ_k(1 / (2σ_k^2))·L_k + log(σ_k), k∈{seg, cls, dist}, where σ_k is the learnable homoscedasticity uncertainty parameter of each subtask, which is initialized to 1.0 and automatically adjusted during training; D_gt is generated by transforming the ground truth mask through three-dimensional Euclidean distance; and the model feedback is obtained by performing multi-task collaborative supervision calculation on the model using the total loss.

[0018] Step S5: Train the MT-3DResUNet fusion model to obtain a trained model; use the trained model to segment and extract ground-glass nodules from lung CT data; perform three-dimensional connected component analysis and morphological post-processing on the segmentation and extraction results to obtain nodule image results; perform statistical analysis on nodule volume, diameter, CT value, and follow-up data analysis on the nodule image results. Preferably, in one embodiment of the present invention, the model obtained by training the MT-3DResUNet fusion model includes: employing a transfer learning strategy, using pre-trained 3D convolutional weights on the LUNA16 public dataset to initialize the 3D feature encoding module; during training, freezing the first two stages of the encoding module for 30 epochs of warm-up training, and then unfreezing all parameters for fine-tuning; setting the initial learning rate to 3×10^-4, and smoothly decaying it from 3×10^-4 to 1×10^-6 using a cosine annealing strategy; setting the batch size to 2, using the AdamW optimizer, and a weight decay coefficient of 1×10^-5; iteratively training for 200 epochs, stopping training when the Dice coefficient on the validation set no longer increases for 20 consecutive epochs, and saving the optimal model weights. The trained model is then used to segment lung CT data, minimizing incomplete segmentation results and class confusion caused by the similarity between nodules and blood vessels or pleural structures, thereby improving segmentation accuracy and completing the segmentation of ground-glass nodules. Preferably, in one embodiment of the present invention, the three-dimensional connected component analysis and morphological post-processing method includes: binarizing the nodule voxel probability map output by the model with a threshold of 0.5; performing three-dimensional morphological closing operations using spherical structural elements with a radius of 2 voxels to fill small cavities inside the nodule and smooth the nodule surface; marking each nodule region according to the 26-neighbor three-dimensional connected component analysis and removing pseudo-nodule regions with a volume less than 30 mm^3; calculating the volume, equivalent diameter, maximum diameter, average CT value, and proportion of solid component of each nodule; mapping the centroid of the nodule to the anatomical region of the lung lobe, and calculating the nodule volume doubling time by combining baseline and follow-up CT scans to obtain the final nodule image results and statistical parameters. Thus, the segmentation and extraction of pulmonary ground-glass nodules are completed.

[0019] In summary, low-dose spiral CT chest scan sequences were acquired, and lung regions were pre-segmented and resampled to obtain standardized CT volume data. Gray-level normalization and data enhancement were performed on the CT volume data based on lung window width and level. Manual layer-by-layer annotation was performed along the nodule boundaries to obtain a reference volume dataset. A 3D feature encoding module was designed based on 3D residual convolution. A 3D feature decoding module was designed based on 3D deconvolution. A multi-task collaborative learning module was designed. The MT-3DResUNet fusion model was obtained through skip connection fusion. A multi-task uncertainty weighted loss function was designed based on the characteristics of blurred ground-glass nodule boundaries and sparse foreground voxels. Multi-task collaborative supervision calculation was performed based on the loss function weights to obtain model feedback. The MT-3DResUNet fusion model was trained to obtain a trained model. The trained model was used to segment and extract ground-glass nodules from the lung CT volume data. 3D connected component analysis and morphological post-processing were performed on the segmentation and extraction results to obtain nodule image results. Statistics on nodule volume, diameter, and CT value, as well as follow-up data analysis, were performed on the nodule image results. This invention can segment, extract, and quantitatively characterize different types of lung nodules, such as pure ground-glass nodules and mixed ground-glass nodules, providing technical support for early lung cancer screening and nodule follow-up management.

[0020] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning, characterized in that, The method includes: acquiring low-dose spiral CT chest scan sequences and performing lung region pre-segmentation and resampling preprocessing to obtain standardized CT volume data; performing grayscale normalization data enhancement on the CT volume data according to the lung window width and window level; manually annotating the enhanced volume data layer by layer along the nodule boundary to obtain a reference volume dataset; designing a 3D feature encoding module based on 3D residual convolution to convert the input volume data into tensors to extract the 3D context features of the nodules; designing a 3D feature decoding module based on 3D deconvolution to restore the volume data resolution; designing a multi-task collaborative learning module, including a voxel segmentation subtask, a nodule benign / malignant classification subtask, and a boundary distance regression subtask; and integrating the encoding and decoding modules... Skip connections are added between corresponding layers to form the MT-3DResUNet fusion model. A multi-task uncertainty weighted loss function is designed based on the characteristics of blurred boundaries and sparse foreground voxels in ground-glass nodules to obtain the loss function weights for each sub-task. Multi-task collaborative supervision calculations are performed based on the loss function weights to obtain model feedback. The MT-3DResUNet fusion model is then trained to obtain a trained model. The trained model is used to segment and extract ground-glass nodules from lung CT scan data. Three-dimensional connected component analysis and morphological post-processing are performed on the segmentation and extraction results to obtain nodule image results. The nodule image results are then statistically analyzed for nodule volume, diameter, and CT value, as well as follow-up data.

2. The method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning according to claim 1, characterized in that, The CT data acquisition and preprocessing method includes: acquiring a low-dose spiral CT chest scan sequence with a slice thickness of no more than 1.25 mm, a scanning tube voltage of 120 kV, a tube current of 30-50 mAs, and a reconstruction matrix of 512×512; pre-segmenting the lung region of the CT sequence, extracting the left and right lung masks and removing the trachea and main bronchus using a thresholding method combined with morphological operations; and resampling the CT sequence to an isotropic voxel spacing of 1×1×1 mm to obtain standardized CT volume data.

3. The method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning according to claim 1, characterized in that, The data augmentation and annotation methods include: truncating and linearly mapping the CT volume data to a lung window width of 1500 HU and a window level of -600 HU, with the normalization formula as follows: I_norm(x,y,z) =(I_in(x,y,z) - W_l) / W_w, where I_norm(x,y,z) represents the normalized voxel gray value; I_in(x,y,z) represents the CT value of the volume data at (x,y,z); W_l represents the lower limit of the window level, with a value of -1350HU; W_w represents the window width, with a value of 1500HU; the truncated and mapped voxel values ​​are linearly scaled to the [0,1] interval; the volume data is manually annotated layer by layer along the nodule boundary, independently annotated by two thoracic radiology physicians and reviewed and confirmed by senior physicians to generate reference volume data; a 64×128×128 volume data block is cropped with the nodule as the center and combined with random flipping and elastic deformation to expand the data to obtain the standard dataset.

4. The method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning according to claim 1, characterized in that, The method for obtaining the MT-3DResUNet fusion model includes: reshaping the input volume data block into a 64×128×128×1 tensor; the 3D feature encoding module consists of 4 encoding stages, each stage containing 2 3D residual convolutional blocks and 1 3D downsampling convolution with a stride of 2×2×2, extracting 16, 32, 64, and 128 channels of 3D contextual features stage by stage; the 3D feature decoding module consists of 4 decoding stages, each stage containing 3D deconvolution upsampling and 3D residual convolutional blocks, gradually restoring the volume data resolution; in the encoding... Skip connections are added between corresponding layers of the module and the decoding module to concatenate the encoded and decoded features in the channel dimension. A multi-task collaborative learning module is designed, in which the voxel segmentation subtask outputs a voxel probability map of nodules by performing 1×1×1 convolution and sigmoid activation on the decoded features; the nodule benign and malignant classification subtask outputs benign and malignant probabilities by performing 3D global average pooling and fully connected layers on the deepest encoded features; and the boundary distance regression subtask outputs a distance map from voxels to nodule boundaries by performing 1×1×1 convolution on the decoded features. The MT-3DResUNet fusion model is obtained by fusing the features.

5. The method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning according to claim 1, characterized in that, The method for obtaining the multi-task uncertainty weighted loss function includes: addressing the issues of blurred boundaries of ground-glass nodules, sparse foreground voxels, and large differences in the magnitude of multi-task data, introducing an automatic multi-task weighting mechanism based on homoscedastic uncertainty to obtain the multi-task uncertainty weighted loss function, calculated as follows: L_seg = 1- 2TP / (2TP + FP + FN), L_cls = -[y·log(p) + (1-y)·log(1-p)], L_dist = MAE(D_pred, D_gt), L_total = Σ_k (1 / (2σ_k^2))·L_k + log(σ_k), k∈{seg, In the formula, L_seg is the Dice loss for nodule voxels; TP is the number of true positive voxels in the nodule region; FP is the number of false positive voxels in the lung tissue misclassified as nodules; FN is the number of false negative voxels in the nodule misclassified as lung tissue; L_cls is the cross-entropy loss for benign / malignant classification; y is the ground truth label for benign / malignant; p is the predicted probability of malignancy; L_dist is the mean absolute error loss for boundary distance regression; D_pred is the predicted distance map; D_gt is the ground truth distance map generated by the ground truth mask distance transformation; σ_k is the learnable homoscedasticity uncertainty parameter for each subtask; the loss function is used to perform multi-task collaborative supervision calculation on the MT-3DResUNet fusion model to obtain model feedback.

6. The method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning according to claim 1, characterized in that, The model training method includes: using voxels of the nodule region as foreground targets and voxels of the normal lung tissue region as backgrounds; initializing the three-dimensional feature encoding module with three-dimensional convolutional weights pre-trained on a public lung nodule dataset and performing transfer learning; freezing the shallow stage of the encoding module for pre-training during training, and then unfreezing all parameters for fine-tuning; setting the initial learning rate to 3×10^-4 and dynamically adjusting it according to the cosine annealing strategy; setting the batch size to 2 and iterating for no less than 200 epochs; stopping training when the Dice coefficient on the validation set no longer increases for 20 consecutive epochs and saving the optimal model weights.

7. The method for segmenting and extracting ground-glass nodules in lung CT based on 3D residual U-Net and multi-task collaborative learning according to claim 1, characterized in that, The three-dimensional connected component analysis and morphological post-processing method includes: binarizing the nodule voxel probability map output by the model according to a preset threshold; filling the small cavities inside the nodule and smoothing the nodule surface using three-dimensional morphological closing operations; marking each nodule region according to the three-dimensional connected component analysis and removing pseudo-nodule regions with a volume smaller than a preset threshold; calculating the volume, equivalent diameter, maximum diameter, average CT value and solid component ratio of each nodule; mapping the centroid of the nodule to the anatomical region of the lung lobe, and calculating the nodule volume doubling time in conjunction with follow-up CT scans to obtain the final nodule image results and statistical parameters.

Images