A lung CT image analysis method based on deep learning and electronic equipment

Abstract

The application discloses a lung CT image analysis method and electronic equipment based on deep learning, which comprises the following steps: constructing a data set; establishing an expert pool, including vectorizing CT image data and constructing an expert pool expert segmentation submodel containing multiple expert segmentation submodels for processing data queues with specific image distribution characteristics; constructing a feature extractor for gating, mapping the CT image into a high-dimensional semantic vector reflecting the confidence and distribution label of the abnormal area; using a gating network to adaptively segment the CT image, and selecting the optimal expert segmentation submodel from the expert pool according to the input mixed image feature vector. The application improves the segmentation accuracy of heterogeneous lesions, realizes adaptive selection of different image phenotypes, and outputs three-dimensional segmentation masks and attribute prediction vectors, supporting subsequent quantitative evaluation and clinical decision-making; and the application alleviates the sample imbalance problem and can accurately route rare lesion forms.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, specifically a lung CT image analysis method and electronic device based on deep learning. Background Technology

[0002] Automatic lesion segmentation on lung CT images is a crucial step in quantitative and reproducible disease assessment. However, lesions such as tuberculosis exhibit significant variations in morphology and histological features on CT images, with substantial differences in lesion shape, density, and anatomical distribution among patients. Therefore, general models trained on heterogeneous patient populations may perform poorly on atypical or complex lesion patterns, highlighting the need for adaptive segmentation methods that explicitly address disease heterogeneity.

[0003] In recent years, deep learning methods have become the mainstream approach for automated medical image analysis. From basic architectures such as U-Net and V-Net to the self-configurable nnU-Net, 3D convolution and end-to-end segmentation techniques have significantly improved the ability to locate and estimate the volume of lesions. Work on tuberculosis has gradually expanded from 2D CXR to 3D CT. Studies have shown that deep networks can extract high-dimensional representations such as texture, shape, and space to improve segmentation and classification tasks. However, due to limitations in training data scale, scarcity of pixel-level annotations, and insufficient coverage of lesion categories, many methods still have insufficient generalization ability on atypical or complex lesions. In addition, single general-purpose models often exhibit performance fluctuations in actual tuberculosis CT segmentation applications, requiring more adaptive strategies to alleviate the performance limitations imposed by heterogeneity.

[0004] The Mixture-of-Experts (MoE) paradigm has gained increasing attention in medical image segmentation and multimodal fusion in recent years, but its application in lung CT image analysis remains limited. Summary of the Invention

[0005] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based lung CT image analysis method and electronic device capable of end-to-end adaptive lesion segmentation. This invention employs a gating mechanism to dynamically route each case to the most suitable expert segmentation sub-model from a dedicated network. Each expert segmentation sub-model is trained for patient subgroups with different imaging characteristics. This method solves the problem of heterogeneity in abnormal regions during tuberculosis imaging.

[0006] The lung CT image analysis method based on deep learning provided by this invention includes,

[0007] Construct a dataset, which is a pixel-level labeled dataset of lung CT lesions, with each type of lesion corresponding to a unique integer label value;

[0008] Establishing an expert pool includes vectorizing CT image data and constructing an expert pool containing multiple expert segmentation sub-models. The vectorized data includes semantic vectors and visual vectors. The expert segmentation sub-models are used to process data queues with specific image distribution characteristics. The data queues are generated during the clustering process of vectorized data and include semantic queues, visual queues, and baseline queues.

[0009] Construct a feature extractor for gating to map CT images into high-dimensional semantic vectors that reflect the confidence and distribution labels in abnormal regions;

[0010] Adaptive segmentation of CT images is performed using a gating network, and the optimal expert segmentation sub-model is selected from the expert pool based on the input mixed image feature vector.

[0011] Furthermore, the process of constructing the expert pool, which includes multiple expert segmentation sub-models, includes semantic vector-based clustering, visual vector-based clustering, and training of the expert segmentation sub-models.

[0012] Furthermore, the visual vector-based clustering specifically includes: constructing an unsupervised feature extraction module to capture potential texture, density distribution, and global morphological features in CT images that are not described by attribute labels; mapping the high-dimensional visual vectors into compact feature vectors; and then dividing the images into different visual queues using a clustering algorithm.

[0013] Furthermore, the training of the expert segmentation sub-model specifically includes the following: each expert segmentation sub-model is a fully convolutional neural network, using a single sample of 3D CT image voxel data as input. The encoder extracts high-dimensional semantic features of the CT image data and compresses spatial resolution through continuous 3D convolution and pooling operations to capture global contextual information of abnormal regions. Then, the decoder gradually restores the spatial size of the feature map through 3D deconvolution or interpolation operations, and uses skip connections to fuse high-resolution shallow features of the corresponding level of the encoder to compensate for the loss of spatial information during downsampling. Finally, an abnormal region segmentation probability map is output, which is the probability prediction of each voxel data belonging to different abnormal region types.

[0014] The feature extractor is preferably a residual network with 1 input channel to adapt to grayscale CT images and outputs a 512-dimensional attribute vector for use by the gating network.

[0015] Furthermore, the feature extractor is optimized using a binary cross-entropy loss function with Sigmoid activation.

[0016]

[0017] Where C represents the total number of abnormal region categories, This represents the network's prediction confidence value for the input CT image in the i-th type of abnormal region. This refers to the manually annotated abnormal region distribution labels corresponding to this CT image. This represents the Sigmoid activation function.

[0018] Furthermore, the adaptive segmentation of CT images using a gating network specifically includes concatenating the compact feature vector with the 512-dimensional attribute vector to form a 640-dimensional hybrid image feature vector, which serves as the input to the gating network; the hybrid image feature vector contains a complete digital fingerprint of the CT image in terms of pixel distribution and geometric structure.

[0019] Furthermore, the gating network is trained, including:

[0020] Obtain training labels, calculate the segmentation performance of each expert segmentation sub-model on each image in the training set, and use the Dice similarity coefficient as the evaluation metric.

[0021] For each CT image, the model ID that achieves the highest Dice score among all expert segmentation sub-models is used as the true class label for that image;

[0022] The gating network parameters are optimized by using a category-weighted cross-entropy loss function. By minimizing this loss function, the gating network learns the mapping relationship between image features and the optimal processing path.

[0023] During the inference phase, for the input CT image, the gating network outputs the expert segmentation sub-model corresponding to the ID with the highest confidence.

[0024] Furthermore, it also includes the output of a three-dimensional abnormal region segmentation mask and an abnormal region attribute prediction vector after the CT image data is analyzed and processed by an expert segmentation sub-model selected by a gating network. The abnormal region attribute prediction vector includes spatial location and prediction value.

[0025] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the steps of any of the lung CT image analysis methods described above.

[0026] Beneficial technical effects of the present invention:

[0027] 1. Improved segmentation accuracy for heterogeneous lesions, with Dice coefficient significantly superior to single general models and multiple mainstream segmentation networks;

[0028] 2. By combining features (visual + semantic) and intelligent routing, adaptive selection of different image phenotypes is achieved, avoiding the performance bottleneck of a one-size-fits-all model;

[0029] 3. The system outputs a 3D segmentation mask and attribute prediction vectors to support subsequent quantitative assessment and clinical decision-making;

[0030] 4. The gating network uses class-weighted cross-entropy, which alleviates the sample imbalance problem and can accurately route to rare lesion morphologies.

[0031] 5. It has end-to-end characteristics, from raw CT to segmentation results and attribute output, with a high degree of automation. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the lung CT image analysis method based on deep learning according to the present invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0036] This embodiment describes a deep learning-based method for analyzing lung CT images, such as... Figure 1 As shown, it includes the following steps:

[0037] S1. Construct a dataset, which is a pixel-level labeled dataset of lung CT lesions, with each type of lesion corresponding to a unique integer label value.

[0038] It should be noted that the data in this embodiment is exemplary and comes from a self-built large-scale dataset, covering 17 typical tuberculosis-related lesion types. This dataset encompasses common and diagnostically significant imaging features, such as different morphologies of consolidation, nodules, cavitation, and fibrosis. Each lesion type corresponds to a unique integer label value for subsequent segmentation mask calculation, while also annotating its specific pulmonary tuberculosis lesion subtype for multi-class classification tasks. To ensure consistency and repeatability of annotation, a unified annotation specification was established, clarifying the criteria for determining different lesions in CT images, boundary definition principles, and processing rules for multiple coexisting lesions.

[0039] In addition to self-built datasets, publicly available CT lesion data can be integrated to enhance the model's generalization and comparability. Both self-built and publicly available datasets undergo resampling, unified spatial resolution, and format conversion to ensure compatibility with the main CT image dataset in terms of input dimensions and labeling system.

[0040] S2. Establish an expert pool, including vectorizing CT image data and constructing an expert pool containing multiple expert segmentation sub-models. The vectorized data includes semantic vectors and visual vectors. The expert segmentation sub-models are used to process data queues with specific image distribution characteristics. The data queues are generated during the clustering process of vectorized data and include semantic queues, visual queues, and baseline queues.

[0041] The raw input of CT images is a 3D CT sequence, usually in DICOM or NIfTI format. Essentially, it is a 3D matrix, such as a voxel grid of size 128×128×128. The value of each point (voxel) in the matrix represents the density value (Henness units, HU) of the human tissue at that location. The arrangement and combination of these values ​​constitute the visible "ground-glass opacity", "cavitation", and other shapes.

[0042] Because directly processing millions of voxel data points involves excessive computation and makes it difficult to capture global patterns, this embodiment uses an encoder to compress the aforementioned three-dimensional matrix into a 128-dimensional numerical array, i.e., a visual vector. This vector can be understood as the "digital fingerprint" of the CT image. Each number in the vector no longer represents a specific pixel brightness, but rather represents some abstract image texture feature, such as "whether there is a diffuse distribution" or "whether the edges are sharp".

[0043] Meanwhile, this embodiment transforms the knowledge containing 17 abnormal region types into a 17-dimensional "switch" array, i.e., a semantic vector. For example, if a patient only has "solidification" and "calcification", the vector is marked as 1 at the corresponding position, and 0 for the rest.

[0044] To address the high heterogeneity in texture, morphology, and spatial distribution of abnormal regions in CT images, this invention proposes a segmentation framework based on a hybrid expert strategy. The construction process of the expert pool, comprising multiple expert segmentation sub-models, includes semantic vector-based clustering, visual vector-based clustering, and training of the expert segmentation sub-models.

[0045] S21. The semantic vector-based clustering process includes: To capture the combination patterns of different types of abnormal regions in CT images, such as the co-occurrence patterns of "cavitation + nodules" or "solidification + fibrosis," the CT image samples in the training set are first attribute-encoded based on pre-defined labels. Specifically, for each CT image sample, a 17-dimensional multi-hot encoded vector (v∈{0,1}17) is generated based on its corresponding 17 types of abnormal region annotation information. Each dimension of this vector corresponds to the presence or absence of a specific abnormal texture or structure in the CT image. These vectors are then clustered to discover image groups with similar abnormal region combination patterns. The HDBSCAN algorithm can be optionally applied as the clustering algorithm.

[0046] Considering the discrete binary nature of this feature vector, this embodiment uses Hamming distance as the similarity metric between image samples. The distance calculation formula for two vectors u and v representing different CT image sample features is as follows:

[0047]

[0048] Where u and v represent the feature vectors of abnormal regions in two different CT image samples; N is the vector dimension, and in this embodiment N=17, that is, 17 types of abnormal regions. This is an indicator function; its value is 1 when the condition is met, and 0 otherwise. This formula is used to quantify the degree of difference between two CT images in the composition of abnormal regions, thereby grouping images with similar pathological texture combinations into the same data queue, i.e., the semantic queue.

[0049] S22. Clustering based on visual vectors specifically includes: constructing an unsupervised feature extraction module to capture potential texture, density distribution, and global morphological features in CT images that are not described by attribute labels; mapping the high-dimensional visual vectors into compact feature vectors; and then dividing the images into different visual queues using a clustering algorithm.

[0050] In this embodiment, the unsupervised feature extraction module is based on a 3D variational autoencoder (VAE). The training objective of the VAE is to minimize the following loss function to ensure that the extracted feature vectors can restore the spatial information of the original CT image to the greatest extent:

[0051]

[0052] Where x represents the input three-dimensional CT image voxel matrix, and z represents the latent spatial visual feature vector of the CT image obtained by the encoder, representing the global texture and morphological fingerprint of the image. The neural network parameters representing the encoder, The neural network parameters representing the decoder, This represents the overall objective function determined by the parameters of two neural networks, given an input x. This represents the distribution of the latent variable z calculated given the input image x; This represents the probability distribution of the decoder reconstructing the original CT image based on the feature vector; This represents the average expectation of the values ​​in the expected symbol obtained by continuously sampling z from the encoder's distribution; Let be the prior distribution of the latent variables. This represents the KL divergence, used to constrain the distribution of eigenvectors and prevent overfitting.

[0053] Specifically, the KL divergence term is calculated as follows:

[0054]

[0055] in, and These are the mean and standard deviation of the CT image feature vector z in the j-th dimension, respectively.

[0056] After training, the encoder is used to extract the 128-dimensional visual feature vector of each CT image, and the images are divided into different visual queues by UMAP dimensionality reduction and HDBSCAN clustering.

[0057] Based on the multiple data queues generated by the above clustering, this embodiment trains independent nnU-Net models as expert segmentation sub-models. Each model is a fully convolutional neural network, taking a single sample of 3D CT image voxel data as input. After downsampling feature extraction and upsampling feature fusion, the output is an anomaly region segmentation probability map with the same size as the input, i.e., a probability prediction of each voxel belonging to 17 anomaly region types. The expert segmentation sub-model preferably uses a deep convolutional neural network based on the 3DU-Net architecture (nnU-Net scheduling framework) as the base model. This network structure is specifically designed for the voxel characteristics of 3D CT image data, and has an overall "U"-shaped symmetrical topology, including an encoder path (downsampling) for feature extraction and a decoder path (upsampling) for spatial localization.

[0058] S23. The training of the expert segmentation sub-model specifically includes the following: each expert segmentation sub-model is a fully convolutional neural network. It uses the voxel data of a single sample of 3D CT image as input. The encoder extracts the high-dimensional semantic features of the CT image data and compresses the spatial resolution through continuous 3D convolution and pooling operations to capture the global context information of abnormal regions. Then, the decoder gradually restores the spatial size of the feature map through 3D deconvolution or interpolation operations, and uses skip connections to fuse the high-resolution shallow features of the corresponding level of the encoder to compensate for the loss of spatial information during the downsampling process.

[0059] The final output layer of the network is processed by the Softmax activation function to generate a 3D voxel-level probability distribution map with the exact same size as the original input image. By calculating the maximum probability index for each voxel location, the system finally outputs a 3D segmentation mask containing labels for various anomaly regions, achieving an end-to-end mapping from continuous grayscale data to discrete semantic labels.

[0060] S3. Construct a feature extractor for gating, mapping CT images into high-dimensional semantic vectors that reflect the confidence and distribution labels in abnormal regions.

[0061] Specifically, in order to intelligently route newly input CT images to the most suitable expert model, this step also requires pre-training an image attribute feature extractor. This embodiment uses an adjusted 3DResNet-18 network as its backbone, with the input channel set to 1 to adapt to grayscale CT images. The network training objective is to predict the types of abnormal regions contained in the voxel data of the CT images.

[0062] The feature extractor is optimized using a binary cross-entropy loss function with Sigmoid activation.

[0063]

[0064] Where C represents the total number of abnormal region categories (e.g., 17 in this embodiment), This represents the network's prediction confidence value for the input CT image in the i-th type of abnormal region. This refers to the manually annotated abnormal region distribution labels corresponding to this CT image. This represents the Sigmoid activation function.

[0065] This formula ensures that the feature vectors learned by the network accurately reflect the category distribution of abnormal regions in CT images. Finally, the 512-dimensional vector output from the penultimate layer of the network is extracted as the high-level semantic feature representation of the CT image, which is then used by subsequent gating networks.

[0066] S4. Adaptive segmentation of CT images is performed using a gating network. Based on the input hybrid image feature vector, the optimal expert segmentation sub-model is selected from the expert pool.

[0067] To achieve adaptive segmentation of CT images with different morphologies, this invention constructs an intelligent gating network. This network functions similarly to a "classifier," selecting an optimal expert segmentation sub-model from the aforementioned expert pool based on the input image feature vector.

[0068] The adaptive segmentation of CT images using a gating network specifically includes concatenating a compact feature vector from a VAE with a 512-dimensional attribute vector from a feature extractor to form a 640-dimensional hybrid image feature vector, which serves as the input to the gating network; the hybrid image feature vector contains a complete digital fingerprint of the CT image in terms of pixel distribution and geometric structure.

[0069] In specific implementations, the gating network also needs to be trained. The training objective of the gating network is to predict an "optimal model index (ID)". The training process includes:

[0070] Training labels are obtained, and the segmentation performance of each expert segmentation sub-model on each image in the training set is calculated. The Dice similarity coefficient is used as the evaluation metric.

[0071] Dice = (2*TP) / (2*TP+FP+FN)

[0072] This formula is used to quantify the degree of overlap between the segmentation results and the ground truth annotations in 3D space. TP (TruePositive) represents the number of voxels that the model correctly predicts as anomalous regions; FP (FalsePositive) represents the number of voxels that the model incorrectly predicts as anomalous regions; and FN (FalseNegative) represents the number of anomalous voxels that the model misses.

[0073] For each CT image, the model ID that achieves the highest Dice score among all expert segmentation sub-models is used as the true class label for that image.

[0074] In this embodiment, the gated network architecture and loss function adopt a multilayer perceptron (MLP) as the core architecture of the gated network, containing three hidden layers with 1024, 512, and 256 neurons respectively. Considering the imbalance in the number of images of different types, which leads to some expert models being selected as "optimal" less frequently, this embodiment introduces a class-weighted cross-entropy loss function to optimize the network parameters:

[0075]

[0076] Where x represents the 640-dimensional hybrid image feature vector input to the gating network; x j represents the j-th value output by the gating network, indicating the recommendation confidence of selecting the j-th expert model for segmentation of the image; where C represents the total number of expert segmentation sub-models; e is the natural constant. class represents the actual optimal expert model ID corresponding to the image, i.e., the true label; w class This represents the weighting coefficient for the expert model of class number, used to balance the differences in the number of samples for different image types, ensuring that rare CT images can also be accurately routed.

[0077] Minimize this loss function so that the gating network learns the mapping relationship between image features and the optimal processing path.

[0078] During the inference phase, for the input CT image, the gating network outputs the expert segmentation sub-model corresponding to the ID with the highest confidence.

[0079] In some embodiments, the method further includes S5, where the CT image data is analyzed and processed by an expert segmentation sub-model selected by a gating network, and a three-dimensional abnormal region segmentation mask and an abnormal region attribute prediction vector are output, wherein the abnormal region attribute prediction vector includes spatial location and prediction value.

[0080] The core output of the system is a 3D abnormal region segmentation mask, which is a 3D matrix with the exact same dimensions as the original CT image. In this embodiment, each voxel in the matrix is ​​assigned an integer value between 0 and 17, where 0 represents background tissue and 1-17 represent the precise location and boundaries of different types of abnormal regions, such as solid areas, cavitation areas, and calcifications, in 3D space. This mask can be directly imported into a medical imaging workstation for 3D reconstruction or volume measurement.

[0081] The anomaly region attribute prediction vector utilizes the aforementioned feature extractor to output a 17-dimensional probability vector. Each value (0.0-1.0) in the vector represents the confidence level that the anomaly region belongs to a specific morphological category, such as "tree bud sign" or "ground glass shadow." This provides a structured quantitative indicator for subsequent image retrieval or data statistics.

[0082] This invention provides a holistic end-to-end framework designed to address the significant challenges posed by the substantial radiographic heterogeneity in lung image analysis during CT imaging. It establishes a large-scale, expert-annotated CT dataset systematically covering 17 different types of tuberculosis-related lesions. Based on this unique dataset, an intelligent routing analysis framework based on the nnU-Net baseline model is trained, demonstrating superior performance compared to several contrasting models. This invention achieves robust 3D anomaly region segmentation mask output and attribute predictions for these regions, providing a novel end-to-end framework for addressing disease heterogeneity in medical imaging.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A lung CT image analysis method based on deep learning, characterized in that, Includes the following steps, Construct a dataset, which is a pixel-level labeled dataset of lung CT lesions, with each type of lesion corresponding to a unique integer label value; Establishing an expert pool involves vectorizing CT image data and constructing an expert pool containing multiple expert segmentation sub-models. The vectorized data includes semantic vectors and visual vectors. The expert segmentation sub-models are used to process data queues with specific image distribution characteristics. These data queues are generated during the clustering process of vectorized data and include a semantic queue, a visual queue, and a baseline queue. The construction process of the expert pool containing multiple expert segmentation sub-models includes semantic vector-based clustering, visual vector-based clustering, and training of the expert segmentation sub-models. The clustering based on visual vectors specifically includes: constructing an unsupervised feature extraction module to capture the potential texture, density distribution and global morphological features in CT images that are not described by attribute labels; mapping the high-dimensional visual vectors into compact feature vectors; and then dividing the images into different visual queues through a clustering algorithm. The semantic vector-based clustering process includes first encoding the attributes of the CT image samples in the training set based on preset labels. For each CT image sample, a multi-dimensional multi-hot encoded vector is generated based on its corresponding abnormal region annotation information. Each dimension of the vector corresponds to whether a specific abnormal texture or structure exists in the CT image. Then, these vectors are clustered to discover image groups with similar abnormal region combination patterns. The training of the expert segmentation sub-model includes the process of encoding and decoding three-dimensional CT image voxel data; Construct a feature extractor for gating to map CT images into high-dimensional semantic vectors that reflect the confidence and distribution labels in abnormal regions; Adaptive segmentation of CT images is performed using a gating network, and the optimal expert segmentation sub-model is selected from the expert pool based on the input mixed image feature vector.

2. The lung CT image analysis method according to claim 1, characterized in that, The training of the expert segmentation sub-model specifically includes the following steps: each expert segmentation sub-model is a fully convolutional neural network. It uses a single sample of 3D CT image voxel data as input. The encoder extracts high-dimensional semantic features from the CT image data and compresses the spatial resolution through continuous 3D convolution and pooling operations to capture the global contextual information of the abnormal region. Then, the decoder gradually recovers the spatial size of the feature map through 3D deconvolution or interpolation operations, and uses skip connections to fuse high-resolution shallow features from the corresponding encoder layer to compensate for spatial information loss during downsampling. Finally, it outputs an abnormal region segmentation probability map, which is a probability prediction of each voxel data belonging to different abnormal region types.

3. The lung CT image analysis method according to claim 1, characterized in that, The feature extractor is a residual network with 1 input channel to adapt to grayscale CT images and outputs a 512-dimensional attribute vector for use by the gating network.

4. The lung CT image analysis method according to claim 3, characterized in that, The feature extractor is optimized using a binary cross-entropy loss function with Sigmoid activation. Where C represents the total number of abnormal region categories, This represents the network's prediction confidence value for the input CT image in the i-th type of abnormal region. This refers to the labeling of abnormal regions corresponding to the actual manually annotated areas in the CT image. This represents the Sigmoid activation function.

5. The lung CT image analysis method according to claim 3, characterized in that, The adaptive segmentation of CT images using a gating network specifically includes concatenating the compact feature vector with the 512-dimensional attribute vector to form a 640-dimensional hybrid image feature vector, which serves as the input to the gating network; the hybrid image feature vector contains a complete digital fingerprint of the CT image in terms of pixel distribution and geometric structure.

6. The lung CT image analysis method according to claim 5, characterized in that, Training the gating network includes: Obtain training labels, calculate the segmentation performance of each expert segmentation sub-model on each image in the training set, and use the Dice similarity coefficient as the evaluation metric. For each CT image, the model ID that achieves the highest Dice score among all expert segmentation sub-models is used as the true class label for that image; The gating network parameters are optimized by using a category-weighted cross-entropy loss function. By minimizing this loss function, the gating network learns the mapping relationship between image features and the optimal processing path. For the input CT image, the gating network outputs the expert segmentation sub-model corresponding to the ID with the highest confidence.

7. The lung CT image analysis method according to any one of claims 1 to 6, characterized in that, It also includes the output of a three-dimensional abnormal region segmentation mask and an abnormal region attribute prediction vector after the CT image data is analyzed and processed by an expert segmentation sub-model selected by a gating network. The abnormal region attribute prediction vector includes spatial location and prediction value.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the steps of the lung CT image analysis method as described in any one of claims 1 to 7.

Images