Lung nodule pet-ct image segmentation method based on gan and metabolic reliability guidance

Abstract

The present application relates to the technical field of medical image processing, deep learning, computer-aided diagnosis and multi-modal medical image analysis, in particular to a lung nodule PET-CT image segmentation method based on GAN and metabolic reliability guidance. The present application generates a synthetic PET image corresponding to the anatomical structure according to the CT image through a conditional generative adversarial network, so that the CT single-mode input can be converted into a dual-mode input composed of CT and synthetic PET, thereby expanding the application range of the PET-CT lung nodule segmentation method in the conventional CT screening, follow-up and PET examination resource limited scene. The present application adopts a generator and a discriminator with a convolution block attention module, which is conducive to improving the feature expression capability of the lung nodule region and its surrounding tissue in the synthetic PET image, so that the synthetic PET image is more reasonable in local texture, metabolic distribution and anatomical structure correspondence.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, deep learning, computer-aided diagnosis, and multimodal medical image analysis, specifically to a PET-CT image segmentation method for lung nodules based on GAN and metabolic reliability guidance. Background Technology

[0002] Pulmonary nodules are an important imaging finding in the early screening, diagnosis, and efficacy evaluation of lung cancer. Accurate segmentation of pulmonary nodule regions provides fundamental data for nodule size measurement, volume calculation, boundary morphology analysis, metabolic activity assessment, follow-up comparison, and subsequent auxiliary diagnosis.

[0003] Computed tomography (CT) images can provide high-resolution anatomical information on lung tissue, blood vessels, bronchi, and the margins of pulmonary nodules; positron emission tomography (PET) images can reflect the glucose metabolism level and functional activity of tissues. Combined PET and CT imaging can simultaneously provide structural and metabolic information, which is of great value for the identification, localization, and boundary analysis of pulmonary nodules.

[0004] However, PET-CT imaging still faces several limitations in practical applications. First, PET examinations are costly and involve lengthy procedures. In some routine physical examinations, initial screenings, and follow-up scenarios, only CT images are available, making it difficult to obtain PET images paired with CT images. Second, PET images have low spatial resolution and are easily affected by factors such as noise, partial volume effects, inflammatory hyperuptake, and physiological uptake. The reliability of metabolic information varies significantly across different regions of PET images. If PET images are directly stitched or simply fused with CT images, the low-reliability PET information may interfere with the extraction of nodule edges and structural features from CT images. Third, existing PET-CT-based lung nodule segmentation methods typically rely on actual PET-CT image pairs, making them difficult to directly apply to scenarios where only CT images are available.

[0005] In recent years, conditional generative adversarial networks (GANs) have shown promise in cross-modal medical image synthesis, enabling the generation of images from one modality to another. For example, CT images can be used as conditional input to generate PET images corresponding to their anatomical structures. However, most existing CT-to-PET image generation methods focus only on image generation quality and lack a unified design for subsequent PET-CT bimodal segmentation tasks of lung nodules. Furthermore, existing PET-CT segmentation methods often employ early stitching, feature stitching, or simple weighted fusion strategies, lacking a mechanism for dynamically evaluating the reliability of PET information based on PET standard uptake values ​​and adaptively adjusting the contributions of PET and CT features.

[0006] Therefore, it is necessary to provide a PET-CT segmentation method for lung nodules that can generate corresponding PET images when only CT images are available, and reliably guide the fusion of PET and CT features based on PET metabolic activity. Summary of the Invention

[0007] The purpose of this invention is to overcome the problems in the prior art where dual-modal lung nodule segmentation relies on real PET images, the reliability of PET information is not fully utilized, and it is difficult to apply PET-CT segmentation models when only CT images are available. This invention provides a PET-CT segmentation method for lung nodules based on CT-generated PET images and metabolic reliability-guided fusion.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A GAN- and metabolic reliability-guided PET-CT image segmentation method for lung nodules includes the following steps:

[0010] Step S1: Acquire lung CT images of the object to be analyzed, or acquire PET-CT images of the object to be analyzed; preprocess the acquired images to obtain preprocessed spatially corresponding PET-CT image pairs or CT images for generating PET images;

[0011] Step S2: Determine whether the object to be analyzed has a real PET image; when the object to be analyzed has a real PET image, combine the real PET image and the CT image to form a real PET-CT image pair; when the object to be analyzed does not have a real PET image, input the CT image into the pre-trained conditional generative adversarial network to generate a synthetic PET image corresponding to the CT image, and combine the synthetic PET image and the CT image to form a synthetic PET-CT image pair.

[0012] Step S3: Input the real PET-CT image pair or the synthetic PET-CT image pair into the pre-trained metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network, and extract CT structural features and PET metabolic features respectively;

[0013] Step S4: Generate a pixel-level metabolic reliability map based on the standard uptake values ​​in the PET image, and use the metabolic reliability map to dynamically weight the PET metabolic features;

[0014] Step S5: Perform multi-scale fusion of PET metabolic features weighted by metabolic reliability map and CT structural features to obtain lung nodule segmentation probability map, and obtain lung nodule segmentation mask based on the lung nodule segmentation probability map;

[0015] Step S6: Output the lung nodule segmentation mask and the corresponding auxiliary analysis results; when the PET image is a synthetic PET image, mark the result as an auxiliary segmentation result obtained based on the CT image and the synthetic PET image in the auxiliary analysis results.

[0016] Furthermore, in step S1, the preprocessing includes DICOM image reading, PET-CT spatial registration, CT grayscale adjustment, PET standard uptake value normalization, lung region extraction, region of interest cropping, and image size unification.

[0017] Furthermore, the CT images are adjusted for grayscale using a window width of 400 HU and a window level of 50 HU.

[0018] Furthermore, the conditional generative adversarial network in step S2 includes a generator and a discriminator; the generator takes CT images as conditional input and outputs synthesized PET images, and the discriminator is used to determine the authenticity of the image pair composed of CT images and PET images.

[0019] Furthermore, the generator employs a network structure with a self-designed encoder, decoder, and skip connections; the encoder performs multi-scale feature extraction on CT images, the decoder upsamples the features stepwise and outputs a synthetic PET image, and the skip connections are used to pass detailed features from the encoder to the decoder.

[0020] Furthermore, at least some skip connections of the generator are provided with convolutional block attention modules, which are used to enhance the expression of features related to the lung nodule region and its surrounding tissues.

[0021] Furthermore, the discriminator adopts a self-designed deep learning local discriminant structure, whose input is a real image pair composed of CT images and real PET images, or a pseudo image pair composed of CT images and synthetic PET images; the discriminator outputs the authenticity judgment results corresponding to multiple local regions.

[0022] Furthermore, the conditional generative adversarial network is trained under supervised conditions using paired CT images and real PET images during the training phase; the training loss of the conditional generative adversarial network includes adversarial loss and reconstruction loss, wherein the reconstruction loss is the L1 reconstruction loss between the synthetic PET image and the real PET image.

[0023] Furthermore, the total loss function of the conditional generative adversarial network is:

[0024]

[0025] Among them, L advTo counteract the loss, L1 is the L1 reconstruction loss, and λ is the weighting coefficient of the L1 reconstruction loss.

[0026] Furthermore, the conditional generative adversarial network is trained at a resolution of 256×256, and the generated PET images are upsampled to restore the same spatial resolution as the CT images.

[0027] Furthermore, the metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network in step S3 includes a CT backbone encoder, a PET auxiliary encoder, a reliability gating module, a multi-scale feature fusion module, and a decoder.

[0028] Furthermore, the CT backbone encoder is used to extract the structural, edge, texture, and morphological features of the lung nodules; the PET auxiliary encoder is used to extract the metabolic activity features of the lung nodule region.

[0029] Furthermore, the network parameter scale of the PET auxiliary encoder is smaller than that of the CT trunk encoder, so that the PET auxiliary encoder mainly provides metabolic guidance information, while the CT trunk encoder mainly provides fine structural information.

[0030] Further, step S4 is specifically operated as follows: the normalized PET image is input into the metabolic reliability gating module to obtain a metabolic reliability map with a value range of 0 to 1; in the metabolic reliability map, the position with the higher PET standard uptake value has a higher metabolic reliability value.

[0031] Furthermore, the metabolic reliability map is calculated using the following formula:

[0032]

[0033] Where R is the metabolic reliability map, σ is the Sigmoid function, P is the normalized PET image, T is the preset metabolic threshold, and γ is the threshold transition range adjustment parameter.

[0034] Furthermore, the preset metabolic threshold T corresponds to the standard intake value.

[0035] Further, in step S5, the PET metabolic features output by the PET auxiliary encoder are multiplied element-wise with the metabolic reliability map to obtain the reliability-weighted PET metabolic features; the reliability-weighted PET metabolic features are then fused with the corresponding scale CT structural features for multi-scale feature fusion.

[0036] Furthermore, the multi-scale feature fusion process is expressed as follows:

[0037]

[0038] Among them, F fuseFor fusion features, F CT For CT structural features, F PET denoted as PET metabolic features, R as metabolic reliability map, ⊙ as element-wise multiplication, and α as learnable fusion weight.

[0039] Furthermore, in step S5, when the standard uptake value of a certain location in the PET image is higher than the preset metabolic threshold, the contribution of the PET metabolic features at that location to the fusion features is increased; when the standard uptake value of a certain location in the PET image is lower than the preset metabolic threshold, the contribution of the PET metabolic features at that location to the fusion features is decreased, and CT structural features are used as the main segmentation basis.

[0040] Furthermore, the decoder performs stepwise upsampling and feature recovery on the fused features, and outputs a lung nodule segmentation probability map with the same size as the input image.

[0041] Furthermore, the metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network employs a joint loss function consisting of Dice loss and boundary-aware loss during the training phase:

[0042]

[0043] Among them, L Dice For Dice's loss, L edge For boundary-aware loss, λ edge The weighting coefficients for boundary-aware loss.

[0044] Furthermore, the boundary-aware loss is used to constrain the consistency between the predicted edges of the lung nodule segmentation probability map and the actual labeled edges of the lung nodules.

[0045] Compared with the prior art, the present invention has at least the following beneficial effects:

[0046] First, this invention targets application scenarios where only CT images are available but real PET images are lacking. By using a conditional generative adversarial network to generate synthetic PET images corresponding to the anatomical structures based on CT images, the single-modal input of CT can be converted into a dual-modal input consisting of CT and synthetic PET, thereby expanding the applicability of the PET-CT lung nodule segmentation method in scenarios where conventional CT screening, follow-up, and PET examination resources are limited.

[0047] Second, the present invention employs a generator and discriminator with a convolutional block attention module, which is beneficial to improving the feature expression ability of the lung nodule region and its surrounding tissue in the synthetic PET image, making the synthetic PET image more reasonable in terms of local texture, metabolic distribution and anatomical structure correspondence.

[0048] Third, the present invention uses a CT main encoder and a PET auxiliary encoder to form an asymmetric dual-modal segmentation structure. It uses CT images to extract high-resolution structural and edge information and PET images to provide metabolic activity guidance information, which can reduce the impact of low resolution and noise of PET on nodule boundary segmentation.

[0049] Fourth, this invention constructs a pixel-level metabolic reliability map based on PET standard uptake values, enhances the guiding role of PET metabolic features on segmentation results in high-metabolic regions, reduces the weight of PET features in low-metabolic regions while retaining the dominant role of CT structural features, thereby improving the segmentation stability of complex nodules, metabolically heterogeneous nodules, and PET low-reliability regions.

[0050] Fifth, this invention uses Dice loss and boundary-aware loss to jointly optimize the lung nodule segmentation network, which can simultaneously improve the overall overlap accuracy and edge localization accuracy of the lung nodule region, and is especially suitable for lung nodule segmentation with blurred edges, irregular shape, or adjacent to surrounding tissues.

[0051] Sixth, this invention adopts two input modes: real PET-CT images and CT and synthetic PET images, and distinguishes between real PET image mode and synthetic PET image mode in the output results, which helps to improve the applicability of the system under different clinical image acquisition conditions and the traceability of results. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall process of the PET-CT segmentation method for lung nodules described in this invention.

[0053] Figure 2 This is a schematic diagram of the structure of the conditional generative adversarial network described in this invention.

[0054] Figure 3 This is a schematic diagram of the structure of the metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network described in this invention. Detailed Implementation

[0055] The present invention will be further described below with reference to embodiments. It should be understood that the following embodiments are used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or adjustments made by those skilled in the art to the number of network layers, number of channels, training parameters, image size, loss function form, and threshold parameters without departing from the concept of the present invention should be considered to fall within the scope of protection of the present invention.

[0056] Example 1: Image Acquisition and Preprocessing

[0057] Acquire lung CT images of the subject to be analyzed, or acquire paired PET and CT images of the subject to be analyzed. For subjects with PET images, spatially register the PET and CT images so that they correspond to the same lung anatomical region in the same coordinate system.

[0058] The CT images are adjusted for window width and window level. In this embodiment, the window width is set to 400 HU and the window level is set to 50 HU. The PET images are processed using standard capture values, which are cropped to the range of 0 to 15 and linearly normalized to 0 to 1.

[0059] The lung region is extracted from the preprocessed image, or a region of interest is extracted centered on a candidate lung nodule region. The region of interest is then cropped into fixed-size two-dimensional image slices or three-dimensional image blocks. In this embodiment, the two-dimensional image slices are adjusted to 512×512 pixels.

[0060] Example 2: Construction and Training of CT-Conditioned PET Image Generation Network

[0061] A conditional generative adversarial network (GAN) is constructed, consisting of a generator G and a discriminator D. The generator G adopts a U-Net network structure with an encoder, a decoder, and skip connections. The encoder consists of multiple convolutional downsampling modules used to extract structural features at different scales from the input CT images; the decoder consists of multiple upsampling modules used to progressively restore the spatial resolution of the images and output synthetic PET images.

[0062] A convolutional block attention module is set in at least some skip connections of the generator. The convolutional block attention module includes a channel attention submodule and a spatial attention submodule, which are used to adaptively weight the features output by the encoder in terms of channel and spatial dimensions, so that the generator pays more attention to the structural features related to lung nodules and their surrounding tissues.

[0063] Discriminator D employs a PatchGAN local discriminant structure. The discriminator input is a stitched result of CT and PET images along the channel dimension. Real samples consist of real CT and PET images, while pseudo samples consist of real CT images and synthetic PET images output by the generator. The discriminator outputs realism judgments for multiple local regions to constrain the local texture and overall metabolic distribution of the synthetic PET images.

[0064] During the training phase, paired CT images and real PET images are input. The generator produces synthetic PET images based on the CT images, and a discriminator distinguishes between the real and synthetic PET images. The generator's total loss function includes adversarial loss and L1 reconstruction loss.

[0065] In this embodiment, the conditional generative adversarial network is trained at a resolution of 256×256. After training, the synthesized PET images output by the generator are upsampled to a resolution of 512×512, maintaining a spatial correspondence with the corresponding CT images.

[0066] Example 3: Construction and Training of Metabolic Reliability-Guided Asymmetric PET-CT Lung Nodule Segmentation Network

[0067] A metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network was constructed, comprising a CT backbone encoder, a PET auxiliary encoder, a reliability gating module, a multi-scale feature fusion module, and a decoder.

[0068] The CT trunk encoder is used to extract the edges, texture, morphology, and surrounding tissue structure features of lung nodules in CT images. The PET-assisted encoder is used to extract the distribution features of metabolic activity in PET images. The PET-assisted encoder has fewer basic channels than the CT trunk encoder, allowing the PET branch to primarily provide metabolic guidance information, while the CT branch primarily provides fine structural information.

[0069] The normalized PET image P is input into the reliability gating module to generate a metabolic reliability map R.

[0070] The PET metabolic features output from the PET auxiliary encoder are multiplied element-wise with the metabolic reliability map to obtain the reliability-weighted PET metabolic features. Subsequently, the reliability-weighted PET metabolic features are fused with the corresponding-scale CT structural features output from the CT backbone encoder.

[0071] In the high-metabolic regions of PET, the values ​​of the metabolic reliability map are higher, and PET metabolic features contribute significantly to the fusion results; in the low-metabolic regions of PET, the values ​​of the metabolic reliability map are lower, the contribution of PET metabolic features is suppressed, and the fusion results mainly depend on CT structural features.

[0072] The fused multi-scale features are input into the decoder. The decoder restores the spatial resolution through progressive upsampling, convolution, and skip connections, and outputs a lung nodule segmentation probability map. The lung nodule segmentation probability map is then binarized according to a preset threshold to obtain a lung nodule segmentation mask.

[0073] The segmentation network is jointly trained using Dice loss and boundary-aware loss.

[0074] Example 4: Lung nodule segmentation in real PET-CT input mode

[0075] For the object to be analyzed that has real PET images and CT images, after preprocessing, the CT images are input into the CT main encoder, and the real PET images are input into the PET auxiliary encoder and the reliability gating module.

[0076] The reliability gating module generates a metabolic reliability map based on standard uptake values ​​from real PET images. In regions with high standard uptake values, PET metabolic features are assigned higher weights; in regions with low standard uptake values, they are assigned lower weights. After multi-scale feature fusion and decoding, the module outputs a lung nodule segmentation mask and corresponding auxiliary analysis results.

[0077] Example 5: Lung nodule segmentation in CT input mode only

[0078] For an object to be analyzed that only has CT images, after completing the CT image preprocessing, the CT images are input into the trained generator G to obtain a synthetic PET image corresponding to the CT images.

[0079] Synthetic PET and CT images are combined to form a synthetic PET-CT image pair, which is then input into a trained metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network. The reliability gating module generates a metabolic reliability map based on the synthetic PET images and dynamically weights the corresponding PET metabolic features. After multi-scale feature fusion and decoding, a lung nodule segmentation mask is output.

[0080] The output results in this mode can be labeled as "Lung nodule auxiliary segmentation results obtained based on CT images and synthetic PET images" to distinguish them from lung nodule segmentation results obtained based on real PET-CT images.

[0081] Example 6: Output of Auxiliary Analysis Results

[0082] Lung nodule segmentation masks are overlaid on original CT images, real PET images, or synthetic PET images. Based on the lung nodule segmentation mask, the area, volume, maximum diameter, boundary morphology parameters, or PET metabolic characteristic parameters of the nodule region can be calculated, and corresponding auxiliary analysis results can be output.

[0083] When the input is a pair of real PET-CT images, the auxiliary analysis results are identified as obtained based on real PET-CT images; when the input is a pair of CT images and synthetic PET images, the auxiliary analysis results are identified as obtained based on both CT images and synthetic PET images.

[0084] It should be noted that this invention is not limited to the embodiments described above. Adjustments can be made under the concept of this invention for different image acquisition devices, different PET standard uptake ranges, different image resolutions, different network structure depths, different number of feature channels, different threshold parameters, and different training strategies. Any technical solution that uses CT-conditional PET image generation, dynamically weights PET features using PET metabolic reliability maps, and performs multi-scale fusion of reliability-weighted PET features with CT structural features to achieve lung nodule segmentation should fall within the protection scope of this invention.

Claims

1. A PET-CT image segmentation method for pulmonary nodules based on GAN and metabolic reliability guidance, characterized in that, Includes the following steps: Step S1: Acquire lung CT images of the object to be analyzed, or acquire PET-CT images of the object to be analyzed; preprocess the acquired images to obtain preprocessed spatially corresponding PET-CT image pairs or CT images for generating PET images; Step S2: Determine whether the object to be analyzed has a real PET image; when the object to be analyzed has a real PET image, combine the real PET image and the CT image to form a real PET-CT image pair; when the object to be analyzed does not have a real PET image, input the CT image into the pre-trained conditional generative adversarial network to generate a synthetic PET image corresponding to the CT image, and combine the synthetic PET image and the CT image to form a synthetic PET-CT image pair. Step S3: Input the real PET-CT image pair or the synthetic PET-CT image pair into the pre-trained metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network, and extract CT structural features and PET metabolic features respectively; Step S4: Generate a pixel-level metabolic reliability map based on the standard uptake values ​​in the PET images, and use the metabolic reliability map to dynamically weight the PET metabolic features; Step S5: Perform multi-scale fusion of PET metabolic features weighted by metabolic reliability map and CT structural features to obtain lung nodule segmentation probability map, and obtain lung nodule segmentation mask based on lung nodule segmentation probability map; Step S6: Output the lung nodule segmentation mask and the corresponding auxiliary analysis results; when the PET image is a synthetic PET image, mark the result as an auxiliary segmentation result obtained based on the CT image and the synthetic PET image in the auxiliary analysis results.

2. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, In step S1, the preprocessing includes DICOM image reading, PET-CT spatial registration, CT grayscale adjustment, PET standard uptake value normalization, lung region extraction, region of interest cropping, and image size unification. The CT images are adjusted for grayscale using a window width of 400 HU and a window level of 50 HU.

3. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, The conditional generative adversarial network in step S2 includes a generator and a discriminator; the generator takes CT images as conditional input and outputs synthesized PET images, and the discriminator is used to determine the authenticity of the image pair composed of CT images and PET images. The generator employs a network structure with a self-designed encoder, decoder, and skip connections. The encoder performs multi-scale feature extraction on CT images, the decoder upsamples the features stepwise and outputs a synthetic PET image, and the skip connections are used to pass detailed features from the encoder to the decoder. The generator includes a convolutional block attention module in at least some of its skip connections, the convolutional block attention module being used to enhance the expression of features related to the lung nodule region and its surrounding tissues; The discriminator adopts a self-designed deep learning local discriminant structure. Its input is a real image pair composed of CT images and real PET images, or a pseudo image pair composed of CT images and synthetic PET images. The discriminator outputs the authenticity judgment results corresponding to multiple local regions.

4. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, In step S2, the conditional generative adversarial network is trained under supervision using paired CT images and real PET images during the training phase; the training loss of the conditional generative adversarial network includes adversarial loss and reconstruction loss, wherein the reconstruction loss is the L1 reconstruction loss between the synthetic PET image and the real PET image. The total loss function of the conditional generative adversarial network is: Among them, L adv To counteract the loss, L1 is the L1 reconstruction loss, and λ is the weighting coefficient of the L1 reconstruction loss.

5. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, The conditional generative adversarial network is trained at a resolution of 256×256, and the generated PET images are upsampled to restore the same spatial resolution as the CT images.

6. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, The metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network in step S3 includes a CT backbone encoder, a PET auxiliary encoder, a reliability gating module, a multi-scale feature fusion module, and a decoder. The CT backbone encoder is used to extract the structural, edge, texture, and morphological features of lung nodules; the PET auxiliary encoder is used to extract the metabolic activity features of the lung nodule region. The network parameter scale of the PET auxiliary encoder is smaller than that of the CT trunk encoder, so that the PET auxiliary encoder mainly provides metabolic guidance information, while the CT trunk encoder mainly provides fine structural information.

7. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, The specific operation of step S4 is as follows: Normalized PET images are input into a metabolic reliability gating module to obtain a metabolic reliability map with values ​​ranging from 0 to 1; in the metabolic reliability map, the position with the higher PET standard uptake value has a higher metabolic reliability value. The metabolic reliability map is calculated using the following formula: Where R is the metabolic reliability map, σ is the Sigmoid function, P is the normalized PET image, T is the preset metabolic threshold, and γ is the threshold transition range adjustment parameter; the preset metabolic threshold T corresponds to the standard intake value.

8. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, In step S5, the PET metabolic features output by the PET auxiliary encoder are multiplied element-wise with the metabolic reliability map to obtain the reliability-weighted PET metabolic features; the reliability-weighted PET metabolic features are then fused with the corresponding scale CT structural features for multi-scale feature fusion. The multi-scale feature fusion process is represented as follows: Among them, F fuse For fusion features, F CT For CT structural features, F PET denoted as PET metabolic features, R as metabolic reliability map, ⊙ as element-wise multiplication, and α as learnable fusion weight.

9. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, In step S5, when the standard uptake value of a certain location in a PET image is higher than a preset metabolic threshold, the contribution of the PET metabolic features at that location to the fusion features is increased; when the standard uptake value of a certain location in a PET image is lower than the preset metabolic threshold, the contribution of the PET metabolic features at that location to the fusion features is decreased, and CT structural features are used as the main segmentation basis. The decoder performs stepwise upsampling and feature recovery on the fused features, and outputs a lung nodule segmentation probability map with the same size as the input image.

10. The lung nodule PET-CT image segmentation method based on GAN and metabolic reliability as described in claim 1, characterized in that, The metabolic reliability-guided asymmetric PET-CT lung nodule segmentation network described above uses a joint loss function consisting of Dice loss and boundary-aware loss during the training phase. Among them, L Dice For Dice's loss, L edge For boundary-aware loss, λ edge is the weighting coefficient of the boundary-aware loss; the boundary-aware loss is used to constrain the consistency between the predicted edges of the lung nodule segmentation probability map and the actual labeled edges of the lung nodules.

Images