An automated segmentation and scoring method and system for FDG PET-CT lesions in lymphoma

Abstract

This invention discloses an automatic segmentation and scoring method and system for FDG PET-CT lesions in lymphoma, belonging to the field of medical image analysis technology. It aims to improve the segmentation accuracy of lymphoma lesions and the objectivity of the Deauville score. First, standardized uptake values ​​are calculated for FDG PET images, and lymph node morphological features are extracted from CT images based on multi-scale Hessian enhancement filtering to construct a dual-modality PET-CT image pair. Then, a dual-channel depth network is used to extract anatomical structural features and metabolic distribution features respectively. Cross-modal gating fusion is used to suppress physiological uptake interference, outputting preliminary lesion segmentation results. Next, three-dimensional connected component analysis is performed on the segmentation mask, and metabolic heterogeneity index is extracted by combining kurtosis and Haar wavelet multi-scale energy. A graph attention network is used to identify key lesions. Finally, the ratio of key lesions to standardized liver uptake values ​​is combined with the metabolic heterogeneity index to correct the Deauville score, achieving automation from lesion detection to treatment efficacy evaluation.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, specifically to an automatic segmentation and scoring method and system for FDG PET-CT lesions of lymphoma. Background Technology

[0002] With the increasing incidence of malignant lymphoma, the clinical demand for accurate assessment of lesion distribution, metabolic activity, and treatment response is constantly growing. Positron emission tomography (FDG PET) combined with computed tomography (CT) has become a core imaging method for lymphoma staging, efficacy evaluation, and recurrence monitoring. This technology reveals the metabolic activity level of lesions by detecting areas of abnormal glucose metabolism, and combined with the anatomical information provided by CT, it provides doctors with multi-dimensional diagnostic evidence. However, lesions in PET-CT images often exhibit multifocal distribution, significant differences in metabolic intensity, and blurred boundaries. Furthermore, the presence of physiological uptake areas in the images makes manual segmentation and subjective scoring susceptible to the influence of observer experience, fatigue, and bias, reducing diagnostic consistency and quantitative reliability. In the context of the rapid development of precision medicine, there is an urgent need for automated and objective lesion identification and scoring technologies to assist doctors in improving diagnostic efficiency and the accuracy of efficacy evaluation.

[0003] Existing automated or semi-automated methods have several shortcomings. Fixed-threshold segmentation cannot adapt to metabolic heterogeneity and ambiguous boundaries, easily leading to undersegmentation or oversegmentation. While some deep learning models integrate PET and CT information, they often use shallow stitching methods, failing to effectively utilize the structural and metabolic connections between the two modalities, resulting in insufficient accuracy in distinguishing pathological lymph nodes from physiologically hyper-uptake areas. Most methods treat lesions independently, ignoring the spatial and metabolic connections between lesions throughout the body, making it difficult to reflect the overall condition. The segmentation and scoring processes are disconnected, still requiring manual intervention, resulting in low automation levels and limiting clinical application. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an automatic segmentation and scoring method and system for FDG PET-CT lesions in lymphoma, solving the problems mentioned above.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma, comprising the following steps: S1. Calculating standardized uptake values ​​based on body weight and injection dose for the input FDG PET images; extracting lymph node morphological features from CT images using a multi-scale Hessian matrix enhancement filter; establishing a correspondence between landmark points in the dual-modal images based on the extracted morphological features; constructing PET-CT image pairs; S2. Inputting the PET-CT image pairs into a dual-channel deep network, wherein the CT channel uses deformable convolution to extract lymph node anatomical features, and the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features; establishing cross-modal gating fusion between feature maps at each decoding level of the two channels; dynamically adjusting the weights of metabolic features in the PET channel according to the tissue density features of the CT channel; suppressing missegmentation of physiological uptake regions; and outputting a preliminary segmentation mask for lymphoma lesions. S3. Perform three-dimensional connected component analysis on the segmentation mask to extract independent lesion instances. Calculate the statistical kurtosis of the voxel-normalized uptake values ​​within the lesion and the multi-scale energy distribution based on Haar wavelet transform to obtain the metabolic heterogeneity index for each lesion. Construct a topology graph with lesions as nodes and edge weights determined by metabolic feature similarity and anatomical proximity. Identify metabolically active key lesions located along the lymphatic drainage path using a graph attention network. S4. Calculate the ratio of the maximum normalized uptake value of the key lesion to the average normalized uptake value of normal liver tissue. Combine this with the metabolic heterogeneity index to correct the Deauville score of the key lesion. When the metabolic heterogeneity index exceeds a preset threshold, adjust the Deauville score of the key lesion accordingly.

[0006] Further, step S1 includes the following steps: calculating the standardized uptake value of each voxel in the FDG PET image; setting the parameter range of the multi-scale Hessian matrix enhancement filter based on the Heinz unit value of the CT image; enhancing the morphological features of tubular and globular lymph node structures by calculating the response of the eigenvalues ​​of the Hessian matrix in different scale spaces; extracting points with local maximum Hessian eigenvalues ​​as morphological feature markers in the enhanced CT image, and locating the center point of the metabolically active region in the corresponding PET image as a metabolic feature marker; optimizing the spatial correspondence between the two sets of markers through the iterative nearest point algorithm, and establishing a bimodal marker correspondence based on morphological and metabolic features.

[0007] Furthermore, the specific process of constructing a PET-CT image pair is as follows: Based on the established correspondence between bimodal marker points, the spatial transformation parameters between the PET image and the CT image are calculated; a non-rigid spatial transformation is performed on the PET image using a thin-plate spline interpolation algorithm to align the metabolic feature marker points of the PET image with the morphological feature marker points of the CT image in three-dimensional space; the registration error index is calculated using the aligned bimodal images, and the spatial transformation parameters are re-optimized when the registration error exceeds the preset tolerance until a PET-CT image pair that meets the registration accuracy requirements is obtained.

[0008] Furthermore, the CT channel uses deformable convolution to extract lymph node anatomical features, while the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features. The specific process of establishing cross-modal gating fusion between the feature maps of each decoding level of the two channels is as follows: In the CT channel, the sampling position of the deformable convolution kernel is adaptively adjusted according to the expected morphological features of the lymph node to extract the anatomical features of the lymph node boundary and internal structure; In the PET channel, the metabolic correlation between voxels within the lesion is first calculated through a local self-attention mechanism, and then the metabolic correlation between different lesions is established through a global self-attention mechanism; At each decoding level, the anatomical features extracted from the CT channel are used as a gating signal to control the weight allocation of the metabolic features of the PET channel in cross-modal fusion, and the dynamic fusion of dual-modal features is achieved through a gating recurrent unit.

[0009] Furthermore, the specific process of dynamically adjusting the weights of metabolic features in the PET channel based on the tissue density characteristics of the CT channel to suppress missegmentation of physiological uptake areas and output a preliminary segmentation mask for lymphoma lesions is as follows: Based on the tissue density features extracted from the CT channel, the anatomical locations of physiologically high-uptake organs, including the liver, kidneys, and bladder, are identified; a spatial mask is generated based on the identified physiological organ locations, and the weights of PET metabolic features in the corresponding regions are reduced during cross-modal gating fusion. By iteratively optimizing the threshold settings of the gating signals, the model can distinguish between pathological lymph node uptake and physiological organ uptake; based on the optimized dual-modal feature fusion results, a probability map of each voxel belonging to a lymphoma lesion is generated using the softmax function, and the preliminary segmentation mask is output through probability thresholding.

[0010] Furthermore, a three-dimensional connected component analysis is performed on the segmentation mask to extract independent lesion instances. The specific process for obtaining the metabolic heterogeneity index of each lesion by calculating the statistical kurtosis of the voxel-normalized uptake values ​​within the lesion and the multi-scale energy distribution based on Haar wavelet transform is as follows: A three-dimensional morphological opening operation is performed on the initial segmentation mask to eliminate noise interference. Independent lesion instances in three-dimensional space are identified and labeled using a connected component labeling algorithm based on 26 neighborhoods. For each independent lesion instance, the statistical distribution characteristics of the voxel-normalized uptake values ​​within it are calculated. The kurtosis is calculated using the fourth-order central moment to quantify the concentration of metabolic activity. Simultaneously, multi-level Haar wavelet decomposition is performed on the lesion region to extract wavelet energy coefficients at different scales, constructing an energy distribution vector characterizing the multi-scale characteristics of metabolic distribution. The statistical kurtosis characteristics and the multi-scale energy distribution vector are nonlinearly weighted and fused to generate a heterogeneity index that comprehensively reflects the metabolic inhomogeneity within the lesion.

[0011] Furthermore, a topological graph is constructed with lesions as nodes and edge weights determined by both metabolic feature similarity and anatomical proximity. The specific process of identifying metabolically active key lesions located along the lymphatic drainage path using a graph attention network is as follows: Each independent lesion instance is treated as a graph node, with node features including metabolic parameters and spatial coordinates. The metabolic feature similarity distance and three-dimensional Euclidean distance between any two nodes are calculated, and the edge weights are determined by the weighted harmonic mean of the two distances. Based on the anatomical features of medical images, key anatomical regions along the main lymphatic drainage path are identified. The shortest spatial distance from each lesion to the key anatomical region is calculated. During the message passing process of the graph attention network, higher attention weights are assigned to lesion nodes located along the lymphatic drainage path. After multiple rounds of iterative updates, the node with the highest feature significance is selected as the key lesion.

[0012] Furthermore, the specific process of correcting the Deauville score of key lesions by calculating the ratio of the maximum standardized uptake value of the key lesion to the average standardized uptake value of normal liver tissue, combined with the metabolic heterogeneity index, is as follows: Multiple regions of interest are selected in the right lobe of normal liver tissue, and their average standardized uptake values ​​are calculated as baseline reference values. The ratio of the maximum standardized uptake value of the key lesion to the baseline reference value is calculated, and the baseline Deauville score is determined according to the standard. A score correction function is constructed based on the metabolic heterogeneity index, which maps the metabolic heterogeneity index to a correction factor. The correction factor is then weighted and combined with the Deauville score. When the metabolic heterogeneity index shows a highly uneven metabolic distribution within the lesion, the score level is correspondingly increased.

[0013] Furthermore, when the metabolic heterogeneity index exceeds a preset threshold, the specific process for adjusting the Deauville score of key lesions is as follows: Based on the statistical analysis results of a large-scale clinical dataset, the grading threshold range of the metabolic heterogeneity index is determined, and the metabolic heterogeneity index is divided into different risk levels according to the preset threshold, establishing a correspondence between metabolic heterogeneity and score correction; when the metabolic heterogeneity index exceeds the first threshold but does not exceed the second threshold, the Deauville score is increased by one sub-level; when the metabolic heterogeneity index exceeds the second threshold, the Deauville score is increased by a full level; a score adjustment decision tree is established, recording the metabolic heterogeneity index, risk level, and corresponding score adjustment amount for each key lesion, forming a complete score correction traceability chain.

[0014] An automatic segmentation and scoring system for FDG PET-CT lesions in lymphoma includes the following modules: an image preprocessing module, used to calculate standardized uptake values ​​based on body weight and injection dose from the input FDG PET images, extract lymph node morphological features from the CT images using a multi-scale Hessian matrix enhancement filter, establish a correspondence between landmark points in the dual-modal images based on the extracted morphological features, and construct PET-CT image pairs; a lesion segmentation module, used to input the PET-CT image pairs into a dual-channel deep network, where the CT channel uses deformable convolution to extract lymph node anatomical features, and the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features, establish cross-modal gating fusion between feature maps at each decoding level of the two channels, dynamically adjust the weights of metabolic features in the PET channel according to the tissue density features of the CT channel, suppress missegmentation of physiological uptake regions, and output a preliminary segmentation mask for lymphoma lesions; The lesion analysis module performs 3D connected component analysis on the segmentation mask to extract independent lesion instances. By calculating the statistical kurtosis of the voxel-normalized uptake values ​​within the lesion and the multi-scale energy distribution based on Haar wavelet transform, it obtains the metabolic heterogeneity index for each lesion. It constructs a topology graph with lesions as nodes and edge weights determined by the similarity of metabolic features and the proximity of anatomical locations. A graph attention network is used to identify key lesions that are metabolically active and located along the lymphatic drainage path. The scoring output module calculates the ratio of the maximum normalized uptake value of the key lesion to the average normalized uptake value of normal liver tissue. Combined with the metabolic heterogeneity index, it corrects the Deauville score of the key lesion. When the metabolic heterogeneity index exceeds a preset threshold, the Deauville score of the key lesion is adjusted accordingly.

[0015] The present invention has the following beneficial effects:

[0016] (1) An automatic segmentation and scoring method for FDG PET-CT lesions in lymphoma is proposed. By standardizing the uptake values ​​of PET images, metabolic bias caused by individual weight and injection dose differences is eliminated. Multi-scale Hessian filters are used to enhance the morphological edge features of lymph nodes in CT images, and a correspondence between landmark points in dual-modal images is established to achieve accurate alignment of metabolic and structural information. Based on this, a dual-channel deep network is constructed to adapt to different morphological lymph node structural features through deformable convolution. Multi-scale metabolic distribution information is extracted by combining local-global self-attention mechanism, and cross-modal gating fusion is introduced at the decoding level to effectively suppress interference from physiological uptake areas such as the liver and intestines, and improve the accuracy of lesion boundary recognition and segmentation of multiple lesions.

[0017] (2) An automatic segmentation and scoring system for FDG PET-CT lesions in lymphoma extracts independent lesion instances through three-dimensional connected component analysis. A metabolic heterogeneity index is established by combining the statistical kurtosis of standardized uptake values ​​and the energy distribution of Haar wavelets, reflecting the differences in activity within the lesions from a metabolic perspective. Furthermore, a topological map is constructed based on the metabolic similarity and anatomical proximity between lesions, and a graph attention network is used to identify key lesions with active metabolism and drainage characteristics. Finally, the metabolic intensity of key lesions is compared with that of normal liver tissue, and the Deauville score is dynamically corrected based on the metabolic heterogeneity index. This system can comprehensively assess the metabolic complexity and systematic distribution characteristics of lesions while preserving clinical interpretability, thereby improving the objectivity and consistency of efficacy assessment.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] Figure 1 This is a flowchart of an automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to the present invention.

[0020] Figure 2 This is a flowchart of an automatic segmentation and scoring system for FDG PET-CT lesions of lymphoma according to the present invention. Detailed Implementation

[0021] This application provides an automatic segmentation and scoring method and system for FDG PET-CT lesions in lymphoma, which solves the problems of inaccurate lesion boundary identification, insufficient utilization of metabolic features, and reliance on manual judgment and lack of unified standards in the scoring process in existing lymphoma image analysis.

[0022] The overall concept of the solution in this application embodiment is as follows:

[0023] First, in the image preprocessing stage, the uptake values ​​of PET images are standardized to eliminate metabolic bias caused by individual differences; multi-scale Hessian filtering is used to enhance CT images to highlight the morphological and structural features of lymph nodes, and a correspondence between bimodal image markers is established to ensure accurate spatial registration between PET and CT images.

[0024] Secondly, in the segmentation stage, a dual-channel depth network structure is designed. The CT channel uses deformable convolution to extract anatomical features, while the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features. The dynamic weighting of the two is achieved through cross-modal gating fusion to suppress interference from physiological uptake and output a high-precision lesion mask.

[0025] Furthermore, in the lesion analysis stage, the segmentation mask is processed to extract independent lesions through three-dimensional connected domain processing. The metabolic heterogeneity index is calculated by statistical kurtosis and Haar wavelet energy features. Based on metabolic similarity and anatomical proximity, a lesion topology map is constructed, and key lesions are identified by graph attention network.

[0026] Finally, in the scoring stage, the ratio of the metabolic intensity of key lesions to the reference value of liver tissue is calculated, and the Deauville score is corrected in combination with the metabolic heterogeneity index, realizing an integrated and automated process from lesion detection to clinical quantitative assessment.

[0027] Please see Figure 1This invention provides a technical solution: an automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma, comprising the following steps: S1. Calculating standardized uptake values ​​based on body weight and injection dose for the input FDG PET images; extracting lymph node morphological features from CT images using a multi-scale Hessian matrix enhancement filter; establishing a correspondence between landmark points in the dual-modality images based on the extracted morphological features; constructing PET-CT image pairs; S2. Inputting the PET-CT image pairs into a dual-channel deep network, wherein the CT channel uses deformable convolution to extract lymph node anatomical features, and the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features; establishing cross-modal gating fusion between feature maps at each decoding level of the two channels; dynamically adjusting the weights of metabolic features in the PET channel according to the tissue density features of the CT channel to suppress missegmentation of physiological uptake regions; and outputting a preliminary segmentation mask for lymphoma lesions. S3. Perform three-dimensional connected component analysis on the segmentation mask to extract independent lesion instances. Calculate the statistical kurtosis of the voxel-normalized uptake values ​​within the lesion and the multi-scale energy distribution based on Haar wavelet transform to obtain the metabolic heterogeneity index for each lesion. Construct a topology graph with lesions as nodes and edge weights determined by metabolic feature similarity and anatomical proximity. Identify metabolically active key lesions located along the lymphatic drainage path using a graph attention network. S4. Calculate the ratio of the maximum normalized uptake value of the key lesion to the average normalized uptake value of normal liver tissue. Combine this with the metabolic heterogeneity index to correct the Deauville score of the key lesion. When the metabolic heterogeneity index exceeds a preset threshold, adjust the Deauville score of the key lesion accordingly.

[0028] In this implementation scheme, S1: This step first calculates standardized uptake values ​​for PET images (a dimensionless index obtained by correcting the radioactive uptake of voxels according to the patient's weight and injection dose, often used to compare metabolic intensity in different individuals or at different time phases) to eliminate metabolic biases caused by individual differences, facilitating subsequent quantitative comparisons and threshold determination. A multi-scale Hessian matrix enhancement filter is applied to CT images to highlight the morphological features of small structures such as lymph nodes (a Hessian matrix is ​​a second-derivative matrix that can be used to detect tubular or point-like structures in images; multi-scale processing can simultaneously enhance targets of different sizes). Based on the extracted morphological features, a correspondence between marker points in the dual-modal images is established, i.e., a pairing of structures or feature points is found between PET and CT to complete precise registration, ensuring that metabolic signals correspond one-to-one with the corresponding anatomical locations, reducing misjudgments or positioning errors caused by misalignment. S2: This step uses the registered PET-CT images as dual-channel input. The CT channel employs deformable convolution to flexibly adapt to local deformations in lymph node morphology (deformable convolution introduces learnable sampling displacement on top of conventional convolution, enhancing the modeling ability for irregular shapes and positional shifts) to extract anatomical structural information. The PET channel uses a local-global self-attention mechanism (self-attention is a mechanism that dynamically measures the interaction between different locations; this implementation combines local and global scales, focusing on both neighborhood details and capturing systemic metabolic patterns) to extract multi-scale metabolic distribution features. Cross-modal gating fusion is introduced between each decoding layer, dynamically adjusting the weights of PET features based on the tissue density or structural confidence provided by CT. This increases trust in CT when there is clear anatomical evidence and enhances PET response when metabolic indicators are significant, thereby suppressing missegmentation of physiologically high-uptake areas such as the liver and intestines and outputting a high-quality preliminary segmentation mask. S3: Three-dimensional connected component analysis is performed on the initial mask to distinguish independent lesion instances (connected component analysis can identify spatially connected voxel groups as individual lesions). Within each lesion, the statistical kurtosis of standardized uptake values ​​is calculated (kurtosis reflects the thickness of the distribution tail and the tendency for extreme values, indicating high-metabolic foci or necrotic heterogeneity within the lesion), and the multi-scale energy distribution based on Haar wavelet transform is calculated (Haar wavelets can characterize signal energy at different scales, used to quantify the spatial frequency features of metabolic texture). This yields a metabolic heterogeneity index to characterize the complexity of metabolic distribution within the lesion. Subsequently, a topology map is constructed using each lesion as a node, along with metabolic feature similarity and anatomical proximity. A graph attention network (a graph-based neural network that learns importance weights between nodes through an attention mechanism) is used to identify key lesions that are both metabolically active and located along lymphatic drainage pathways, facilitating the highlighting of clinically representative lesions for subsequent scoring and follow-up.S4: This step first compares the maximum standardized uptake value of the identified key lesion with the average standardized uptake value of a reference normal liver to obtain the relative metabolic intensity (the liver is often used as a homeostatic reference tissue). Then, it dynamically corrects the traditional score of the key lesion by combining it with the aforementioned metabolic heterogeneity index. The score refers to the clinically commonly used Deauville scoring system (a standardized assessment system based on PET-CT imaging, mainly used to assess the efficacy of treatment in lymphoma patients (especially after chemotherapy or radiotherapy). The score ranges from 1 to 5 points, helping doctors judge the treatment response by comparing the metabolic activity of tumor tissue with that of the liver and mediastinal blood pool (through the uptake of radioactive tracers). The specific criteria for a score of 5 are: the radioactive uptake of the lesion is significantly higher than the liver background (e.g., more than 2-3 times the maximum standardized uptake value SUVmax of the liver), or the appearance of new lesions. When the metabolic heterogeneity index exceeds a preset threshold, the original score based on single-point intensity is adjusted up or down to avoid misjudgment caused by a single extreme value, so that the final score maintains clinical interpretability and reflects the actual metabolic complexity of the lesion and its distribution throughout the body.

[0029] Specifically, step S1 includes the following steps: calculating the standardized uptake value of each voxel in the FDG PET image; setting the parameter range of the multi-scale Hessian matrix enhancement filter based on the Heinz unit value of the CT image; enhancing the morphological features of tubular and globular lymph node structures by calculating the response of the eigenvalues ​​of the Hessian matrix in different scale spaces; extracting points with local maximum Hessian eigenvalues ​​as morphological feature markers in the enhanced CT image, and locating the center point of the metabolically active region in the corresponding PET image as a metabolic feature marker; optimizing the spatial correspondence between the two sets of markers using the iterative nearest point algorithm, and establishing a bimodal marker correspondence based on morphological and metabolic features.

[0030] In this implementation scheme, the structural features of lymph nodes are first enhanced using the grayscale information of CT images. CT images use Henle units (HU) to characterize tissue density; lymph nodes typically appear as medium-density, tubular or spherical structures with relatively clear edges. To highlight these morphological features, a multi-scale Hessian matrix enhancement filter is set, whose response function is defined as: ;in, , , In scale The eigenvalues ​​of the lower Hessian matrix are used to describe the principal curvature directions of local gray-level changes; This is the morphology sensitivity adjustment coefficient, used to control the balance between noise suppression intensity and structure preservation capability. When , Approaching zero A large negative value indicates that the region has obvious tubular structure characteristics. This can be determined by calculating different... Below The response value is calculated, and the maximum response is taken as the filter output, which can effectively enhance the contrast of target areas such as lymph nodes and suppress background tissue interference. To further adapt to the density distribution of CT images from different individuals, the Hessian response is selected from the enhanced CT image. Local maxima locations serve as a set of morphological feature markers. These points represent the spatial distribution centers of lymph nodes and similar structures. Simultaneously, metabolic activity is quantified in the corresponding PET images by calculating the Standardized Uptake Value (SUV) for each voxel. The formula for calculating SUV is: ;in, The radioactivity concentration at the voxel; This refers to the dose of the tracer injected; The subject's weight is used as the indicator. This indicator reflects the glucose metabolism level of tissues, thus distinguishing pathological hypermetabolic areas from normal tissues. Based on the local maxima of the SUV (Radius of Volatility), the centers of metabolically active regions are selected as the set of metabolic feature markers. Subsequently, to achieve spatial alignment of PET and CT dual-modal images, the Iterative Closest Point (ICP) algorithm was used to spatially register the two sets of marker points. The optimization objective function is: ;in, For the PET image of the first The three-dimensional coordinates of a metabolic marker; These are the coordinates of the corresponding morphological landmarks in the CT image; This is a spatial transformation matrix, consisting of a rotation matrix and a translation vector. The objective function is minimized... This allows for a gradual approach to the optimal correspondence between the two sets of markers, enabling a high degree of alignment between the metabolic features of PET and the structural features of CT in three-dimensional space, thereby establishing a dual-modal marker correspondence based on morphological and metabolic features.

[0031] Specifically, the process of constructing a PET-CT image pair is as follows: Based on the established correspondence between bimodal markers, the spatial transformation parameters between the PET image and the CT image are calculated; a non-rigid spatial transformation is performed on the PET image using a thin-plate spline interpolation algorithm to align the metabolic feature markers of the PET image with the morphological feature markers of the CT image in three-dimensional space; the registration error index is calculated using the aligned bimodal images, and the spatial transformation parameters are re-optimized when the registration error exceeds the preset tolerance until a PET-CT image pair that meets the registration accuracy requirements is obtained.

[0032] In this implementation scheme, to further eliminate non-rigid differences caused by patient posture or respiratory motion, a thin-plate spline (TPS) algorithm is used to achieve non-rigid fine registration of bimodal images. Its spatial transformation function is expressed as: ;in, The coordinates of the point to be transformed; It is a linear affine matrix used to describe the overall rotation and scaling; It is a translation vector; The first in the CT image One morphological feature point; For non-rigid deformation weighting coefficients; The radial basis functions represent the smooth diffusion characteristics of deformation in space. To obtain the optimal weighting coefficients... The energy function is defined as follows: The first term reflects the registration error; the second term is the smoothing term. To smooth the constraint coefficients and achieve a balance between deformation accuracy and spatial continuity, this scheme dynamically adjusts the error ratio to make the algorithm adaptive. : ;in, and These are the registration errors of the previous round and the current iteration, respectively; This is an empirical coefficient. If the current error decreases slowly, then... It automatically decreases to enhance deformation flexibility; conversely, it increases. To prevent overfitting, after registration is completed, the average spatial error index is calculated: ;when Greater than the preset error threshold Then, the spatial transformation parameters are re-optimized until the error meets the accuracy requirements. Finally, highly aligned PET-CT image pairs in three-dimensional space are obtained, providing an accurate dual-modal input basis for subsequent feature extraction and lesion segmentation.

[0033] Specifically, the CT channel uses deformable convolution to extract lymph node anatomical features, while the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features. The specific process of establishing cross-modal gating fusion between the feature maps of each decoding level of the two channels is as follows: In the CT channel, the sampling position of the deformable convolution kernel is adaptively adjusted according to the expected morphological features of the lymph node to extract the anatomical features of the lymph node boundary and internal structure; In the PET channel, the metabolic correlation between voxels within the lesion is first calculated through a local self-attention mechanism, and then the metabolic correlation between different lesions is established through a global self-attention mechanism; At each decoding level, the anatomical features extracted from the CT channel are used as a gating signal to control the weight allocation of the metabolic features of the PET channel in cross-modal fusion, and the dynamic fusion of dual-modal features is achieved through a gating recurrent unit.

[0034] In this implementation scheme, the CT channel uses a deformable convolution module to extract the anatomical features of lymph nodes. Unlike traditional fixed convolution kernels, deformable convolution can automatically adjust the position of the convolution kernel sampling points according to the input features, thereby better adapting to the morphological changes of lymph nodes. Let the input feature map be... The convolution kernel is Its sampling position offset is The output of deformable convolution can then be expressed as: ;in, This is the current output position; This refers to the basic sampling location of the convolution kernel; The offset predicted by the shape guidance module; This represents the number of sampling points for the convolution kernel. The offset is obtained through learning. The network can adaptively focus on key structures such as lymph node margins, cortex, and medulla, thereby enhancing the expression of anatomical features. In the PET channel, metabolic features are extracted through a local-global self-attention mechanism. First, within a local area, attention weights are calculated based on voxel metabolic similarity to capture the consistency of metabolic distribution within the lesion. The local self-attention calculation can be expressed as: ;in, and These are the query and key vectors for PET features, respectively. and The coordinates of the voxels within the local neighborhood; The feature dimension scaling factor; voxels The local neighborhood set. Local attention enhances the coherence of metabolism within the lesion and helps distinguish high-uptake areas with blurred boundaries. Subsequently, a global self-attention mechanism calculates the metabolic correlations between different lesions, enabling the network to understand the co-metabolic features of distant lesions, thereby improving its sensitivity to multiple lesions. Its global feature aggregation expression is as follows: ;in, This is a value vector of metabolic features; This represents the global attention matrix. During the decoding phase, dual-channel features are fused through a cross-modal gating mechanism. The structural features of the CT channels serve as gating signals, guiding the weight allocation of PET metabolic features to achieve dynamic modal balance. The gating fusion unit can be formally represented as: ;in, Features of fusion; and The weights and biases of the gating unit; For the Sigmoid function; Represents element-wise multiplication; This represents the dynamic control signal output by the gated recurrent unit. Under this design, the network can adaptively adjust the contribution of metabolic features based on structural information, enhancing the model's ability to identify low-contrast lesions.

[0035] Specifically, the process of dynamically adjusting the weights of metabolic features in the PET channel based on the tissue density characteristics of the CT channel to suppress missegmentation of physiological uptake areas and output a preliminary segmentation mask for lymphoma lesions is as follows: Based on the tissue density features extracted from the CT channel, the anatomical locations of physiologically high-uptake organs, including the liver, kidneys, and bladder, are identified; a spatial mask is generated based on the identified physiological organ locations, and the weights of PET metabolic features in the corresponding regions are reduced during cross-modal gating fusion. By iteratively optimizing the threshold settings of the gating signals, the model can distinguish between pathological lymph node uptake and physiological organ uptake; based on the optimized dual-modal feature fusion results, a probability map of each voxel belonging to a lymphoma lesion is generated using the softmax function, and the preliminary segmentation mask is output after probability thresholding.

[0036] In this implementation scheme, the weights of metabolic features in the PET channel are dynamically adjusted based on the tissue density features extracted from the CT channel to suppress missegmentation of physiological uptake regions. CT images can clearly distinguish high-density structures such as the liver, kidneys, and bladder. These regions often exhibit high uptake physiological signals in PET images, and failure to differentiate them can lead to false positives. Therefore, the system first automatically identifies these physiological organ regions based on CT density thresholds and generates a spatial mask. In the cross-modal fusion process, the metabolic weight coefficient of PET features Dynamic adjustment is performed based on this mask: ;in, voxels Tissue density characteristic response in CT channels; The adaptive threshold is determined by the organ recognition module; This is a smoothing adjustment coefficient used to control the steepness of the weight transition. If the voxel is in a physiologically high-density region, then... , thus When the threshold is close to 0, the influence of the PET channel in this region is significantly suppressed. To optimize the threshold... An iterative update strategy based on validation set error is adopted: ;in, For the segmentation loss function; The learning rate. It is automatically adjusted via backpropagation. This allows the model to achieve optimal differentiation between pathological and physiological uptake. On the fused feature map, the probability of each voxel belonging to a lymphoma lesion is calculated using the Softmax function. ;in, and These are the activation values ​​corresponding to the "lesion" and "background" in the fusion features, respectively. A probability threshold is set. (Based on the ROC curve, a value that balances sensitivity and specificity is automatically selected), generating a preliminary segmentation mask: This yields preliminary segmentation results of lymphoma lesions, enabling accurate identification of pathological hypermetabolic regions while effectively avoiding interference from physiological hyperuptake regions, thus providing reliable input for subsequent scoring and quantitative analysis.

[0037] Specifically, the process of performing three-dimensional connected component analysis on the segmentation mask to extract independent lesion instances, and obtaining the metabolic heterogeneity index for each lesion by calculating the statistical kurtosis of the voxel-normalized uptake values ​​within the lesion and the multi-scale energy distribution based on Haar wavelet transform, is as follows: Three-dimensional morphological opening operations are performed on the initial segmentation mask to eliminate noise interference; independent lesion instances in three-dimensional space are identified and labeled using a connected component labeling algorithm based on 26 neighborhoods; for each independent lesion instance, the statistical distribution characteristics of all voxel-normalized uptake values ​​within it are calculated; the kurtosis is calculated using the fourth-order central moment to quantify the concentration of metabolic activity; simultaneously, multi-level Haar wavelet decomposition is performed on the lesion region to extract wavelet energy coefficients at different scales, constructing an energy distribution vector characterizing the multi-scale features of metabolic distribution; the statistical kurtosis features and the multi-scale energy distribution vector are nonlinearly weighted and fused to generate a heterogeneity index that comprehensively reflects the metabolic inhomogeneity within the lesion.

[0038] In this implementation scheme, to quantify the metabolic heterogeneity within each lesion, a three-dimensional morphological opening operation is first performed on the initial segmentation mask to remove isolated noise points and disconnected shadows, ensuring the continuity of lesion boundaries. Subsequently, a three-dimensional connected component labeling algorithm based on 26 neighborhoods is used to identify and number independent lesion instances. Let the... The voxel set of each lesion instance is ,in The number of voxels contained in the lesion, and the standardized uptake value of the voxels is denoted as . To describe the concentration of metabolic activity within the lesion, the statistical kurtosis of its uptake distribution is calculated: ;in, lesion The mean of internal voxel uptake values; the numerator is the fourth central moment, reflecting the extreme deviation of the uptake values; the denominator is the square of the variance, used for normalization. If A larger value indicates a significant metabolic peak in the lesion, with concentrated metabolic distribution; a smaller value indicates a more uniform metabolic distribution. To further characterize the complexity of metabolic distribution at different spatial scales, Haar wavelet decomposition is performed on the lesion region. Let the metabolic image of the lesion region be... ,exist The wavelet coefficient matrix at the level decomposition scale is The energy coefficient at this scale is defined as: ;in, Indicates the first The energy intensity at a given scale reflects the degree of spatial variation in lesion metabolism at that scale. Multi-scale energy vector. It can comprehensively characterize the hierarchical structure of metabolic distribution in lesions. To avoid a single feature dominating the results, a nonlinear weighted fusion method is used to combine statistical kurtosis with multi-scale energy features to obtain the metabolic heterogeneity index: ;in, lesion The metabolic heterogeneity index; and The feature fusion coefficient; For the first The weights of scale energy are automatically optimized by minimizing the segmentation error using the validation set; For offset items; The sigmoid function is used to constrain the output range. This fusion strategy prevents lesions with high peak density but simple structure from being misclassified as highly heterogeneous, while preserving the metabolic complexity of regions with significant multi-scale energy changes.

[0039] Specifically, the process of constructing a topological graph with lesions as nodes and edge weights determined by both metabolic feature similarity and anatomical proximity, and identifying key lesions that are metabolically active and located along the lymphatic drainage path using a graph attention network is as follows: Each independent lesion instance is treated as a graph node, with node features including metabolic parameters and spatial coordinates. The metabolic feature similarity distance and the three-dimensional Euclidean distance between any two nodes are calculated, and the edge weights are determined by the weighted harmonic mean of the two distances. Based on the anatomical features of medical images, key anatomical regions along the main lymphatic drainage path are identified, and the shortest spatial distance from each lesion to the key anatomical region is calculated. During the message passing process of the graph attention network, higher attention weights are assigned to lesion nodes located along the lymphatic drainage path. After multiple rounds of iterative updates, the node with the highest feature significance is selected as the key lesion.

[0040] In this implementation scheme, to identify key lesions with high metabolic activity along the lymphatic drainage pathway, a graph structure model is constructed based on the aforementioned lesion instances. Each lesion instance is considered as a graph node, and the node characteristics are derived from its metabolic feature vector. and spatial coordinates Composition. Any two nodes and The metabolic feature similarity distance between them is represented as: ;in, Let be the Euclidean norm. Spatial proximity is defined by the three-dimensional Euclidean distance: To combine the effects of the two distances, edge weights based on a weighted harmonic average are constructed: ;in, For nodes and Edge weights between them; and These are adjustment coefficients for metabolic features and spatial distance, respectively, and their values ​​are automatically determined based on the joint objective of graph structure sparsity and classification performance in the training set. This weighting method can maintain a high correlation even when metabolic features are similar but spatial distances are far, thereby capturing cross-regional lesion collaboration relationships. A Graph Attention Network (GAT) is introduced into the constructed graph to screen key lesions. The network aggregates neighbor node information through a message passing mechanism, and its attention weights are defined as follows: ;in, For nodes For nodes Attention weights; The characteristic linear transformation matrix of the graph; This is the attention parameter vector; For nodes The set of neighbors; This represents a vector concatenation operation. To strengthen anatomical constraints, key regions along the main lymphatic drainage pathways, such as the neck, mediastinum, and groin, are first identified using CT anatomical segmentation results. The shortest spatial distances from lesion nodes to these regions are then calculated. During the attention update phase, a position correction coefficient is set for nodes close to the traffic path. The final attention weights are adjusted to .in, This is a decay factor used to control the strength of the distance's influence on attention. After multiple rounds of message passing and feature updates, the node feature saliency vector... The process gradually converges, selecting several nodes with the highest significance as key lesions. These lesions simultaneously possess high metabolic activity, complex metabolic distribution, and typical spatial distribution characteristics, providing core input for subsequent grading and scoring, and disease progression assessment.

[0041] Specifically, the process of adjusting the Deauville score of key lesions by calculating the ratio of the maximum standardized uptake value of the key lesion to the average standardized uptake value of normal liver tissue, combined with the metabolic heterogeneity index, is as follows: Multiple regions of interest are selected in the right lobe of normal liver tissue, and their average standardized uptake values ​​are calculated as baseline reference values. The ratio of the maximum standardized uptake value of the key lesion to the baseline reference value is calculated, and the baseline Deauville score is determined according to the standard. A score correction function is constructed based on the metabolic heterogeneity index, which maps the metabolic heterogeneity index to a correction factor. The correction factor is then weighted and combined with the Deauville score. When the metabolic heterogeneity index shows a highly uneven metabolic distribution within the lesion, the score level is correspondingly increased.

[0042] In this implementation scheme, firstly, several regions of interest are selected in the right lobe of a normal liver, and the mean of their standardized uptake values ​​is calculated, denoted as […]. As a reference value for metabolic background; the maximum standardized uptake value of the key lesion area is recorded as Calculate the ratio ;in, The ratio of increased metabolism in the lesion relative to the liver background is used to initially determine the metabolic activity of the lesion. A higher ratio indicates stronger metabolic activity in the lesion, corresponding to a higher Deauville baseline score. After determining the baseline score, the metabolic heterogeneity index is introduced. This index is used to quantitatively characterize the spatial dispersion of metabolic values ​​within a lesion, thus reflecting the uniformity of metabolic distribution within the lesion. This ratio measures the degree of heterogeneity in metabolism at both the local and systemic levels; a higher value indicates a more unstable metabolic distribution. Based on A scoring correction function is constructed to enable the Deauville score to reflect the internal metabolic complexity of lesions. The scoring correction function can be written as: ;in, This is the revised final score; To be based on the ratio A defined baseline score; To map the metabolic heterogeneity index to a function with adjusted weights, reflecting the impact of heterogeneity on the score; This refers to the scoring weighting coefficient, used to control the correction range. In this scheme, the weighting coefficient... The value of is determined through multi-case fitting analysis, selecting the optimal solution that maximizes the scoring accuracy and minimizes the error rate to ensure the statistical reliability of the correction. Through the above correction process, the system can comprehensively reflect the metabolic intensity and distribution complexity of lesions, thereby avoiding misjudgments caused by relying on a single ratio and making the scoring results more consistent with actual clinical manifestations.

[0043] Specifically, when the metabolic heterogeneity index exceeds a preset threshold, the process for adjusting the Deauville score of key lesions is as follows: Based on the statistical analysis results of a large-scale clinical dataset, the grading threshold range of the metabolic heterogeneity index is determined, and the metabolic heterogeneity index is divided into different risk levels according to the preset threshold, establishing a correspondence between metabolic heterogeneity and score correction; when the metabolic heterogeneity index exceeds the first threshold but does not exceed the second threshold, the Deauville score is increased by one sub-level; when the metabolic heterogeneity index exceeds the second threshold, the Deauville score is increased by a full level; a score adjustment decision tree is established, recording the metabolic heterogeneity index, risk level, and corresponding score adjustment amount for each key lesion, forming a complete score correction traceability chain.

[0044] In this implementation plan, to ensure the objectivity and quantifiability of the scoring adjustment process, a statistical distribution analysis of the metabolic heterogeneity index for different lesion types is first performed based on a large-scale clinical database to obtain its correlation curve with the actual diagnostic conclusion. Based on this curve, two threshold levels are set. and And establish a risk classification model. When When, it is considered a low heterogeneous lesion, and the score remains unchanged; when At that time, it belonged to moderate heterogeneous lesions, and the Deauville score improved by one sub-level; when At that time, the lesion was classified as highly heterogeneous, and the score was raised by a full level. To improve the continuity and interpretability of the score adjustment, a score adjustment function was further constructed: ;in, Adjust the increment for the score; A correction intensity coefficient was used to control the smoothness of score improvement within different threshold intervals. The correction intensity coefficient was determined through cross-validation to maintain a stable correction trend across different data distributions. Finally, the system adjusts the score based on the intensity of each key lesion. The value automatically locates its corresponding threshold range and outputs the corresponding value. The score is recorded in the score adjustment decision tree, along with the lesion's heterogeneity index, grading interval, correction coefficient, and final score, forming a traceable scoring traceability chain. This mechanism makes the scoring process transparent, facilitating clinical review and algorithm validation.

[0045] Please see Figure 2 An automatic segmentation and scoring system for FDG PET-CT lesions in lymphoma includes the following modules: an image preprocessing module, used to calculate standardized uptake values ​​based on body weight and injection dose from the input FDG PET images, extract lymph node morphological features from the CT images using a multi-scale Hessian matrix enhancement filter, establish a correspondence between landmark points in the dual-modal images based on the extracted morphological features, and construct PET-CT image pairs; a lesion segmentation module, used to input the PET-CT image pairs into a dual-channel deep network, where the CT channel uses deformable convolution to extract lymph node anatomical features, and the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features, establish cross-modal gating fusion between feature maps at each decoding level of the two channels, dynamically adjust the weights of metabolic features in the PET channel according to the tissue density features of the CT channel, suppress missegmentation of physiological uptake areas, and output a preliminary segmentation mask for lymphoma lesions; The lesion analysis module performs 3D connected component analysis on the segmentation mask to extract independent lesion instances. By calculating the statistical kurtosis of the voxel-normalized uptake values ​​within the lesion and the multi-scale energy distribution based on Haar wavelet transform, it obtains the metabolic heterogeneity index for each lesion. It constructs a topology graph with lesions as nodes and edge weights determined by the similarity of metabolic features and the proximity of anatomical locations. A graph attention network is used to identify key lesions that are metabolically active and located along the lymphatic drainage path. The scoring output module calculates the ratio of the maximum normalized uptake value of the key lesion to the average normalized uptake value of normal liver tissue. Combined with the metabolic heterogeneity index, it corrects the Deauville score of the key lesion. When the metabolic heterogeneity index exceeds a preset threshold, the Deauville score of the key lesion is adjusted accordingly.

[0046] In this implementation, the image preprocessing module calculates standardized uptake values ​​from FDG PET images and performs multi-scale Hessian enhancement on CT images, aiming to achieve a precise correspondence between metabolic and structural information at the image level. By establishing spatial correspondences of bimodal markers based on morphological features, spatial deviations between PET and CT during acquisition can be effectively corrected, ensuring that the bimodal images input in the subsequent segmentation stage are aligned in the same anatomical coordinate system, thereby improving the spatial accuracy and feature consistency of lesion identification. The core function of this module is to provide a structurally clear and metabolically stable input foundation for subsequent deep segmentation, reducing identification errors caused by image distortion or modal misalignment. The lesion segmentation module is based on a dual-channel deep network structure, performing collaborative feature extraction and fusion to address the differences between PET and CT modal data. The CT channel adapts to different lymph node boundary variations through deformable convolution, extracting anatomical structural features; the PET channel captures the spatial distribution patterns of metabolic activity through a local-global self-attention mechanism. Through a cross-modal gating fusion mechanism, the weights of PET features are dynamically adjusted at each decoding level, enabling the model to utilize CT density features to suppress interference from physiologically high uptake regions. The final output preliminary segmentation mask more accurately reflects the true boundaries and metabolically active range of lesions, providing a high-confidence spatial region for subsequent lesion feature extraction. The lesion analysis module performs three-dimensional connected component analysis on the preliminary segmentation results to accurately distinguish independent lesion instances and calculates the metabolic heterogeneity index for each lesion based on metabolic distribution characteristics. This index reflects the non-uniformity of metabolism within the lesion, distinguishing lesion types with different levels of activity. Building upon this, the module further constructs a topological relationship graph between lesions, using metabolic similarity and anatomical proximity as dual constraints to capture lesion association structures with spatial lymphatic drainage relationships. By learning this topological structure through a graph attention network, key metabolically active lesions located along lymphatic drainage paths can be automatically identified, achieving an intelligent transition from segmentation results to lesion importance ranking, providing a target priority basis for the scoring stage. The scoring output module is based on the clinically commonly used Deauville scoring system, introducing the metabolic heterogeneity index to achieve adaptive score correction. The module first uses the ratio of the maximum standardized uptake value of key lesions to the average standardized uptake value of normal liver tissue as the basic scoring criterion, ensuring consistency with clinical standards. Then, it differentiates the scoring results based on the range of metabolic heterogeneity index values. When the heterogeneity index indicates high metabolic heterogeneity within the lesion, the system automatically increases the score level to reflect the greater metabolic complexity of the lesion. By establishing a threshold-based scoring adjustment decision logic, the scoring output module can achieve an objective scoring process driven by quantitative imaging features, avoiding subjective interpretation differences, and ultimately generating traceable comprehensive scoring results, providing accurate and reliable quantitative evidence for clinical diagnosis and efficacy evaluation.

[0047] In summary, this application has at least the following effects:

[0048] An automatic segmentation and scoring method and system for FDG PET-CT lesions in lymphoma is proposed. By introducing multi-scale Hessian enhancement and bimodal landmark registration, precise correspondence between PET and CT images at the structural and metabolic levels is achieved, significantly improving image fusion accuracy. Furthermore, by constructing a dual-channel deep network and introducing a cross-modal gating fusion mechanism in the decoding layer, the model can suppress physiological high uptake interference while enhancing the extraction of pathological metabolic features, thereby improving the accuracy and robustness of lesion segmentation. Finally, a lesion metabolic heterogeneity index is constructed through three-dimensional connected component analysis and Haar wavelet multi-scale feature extraction. This approach enables quantitative assessment of the metabolic complexity within lesions; further, it models the metabolic and spatial relationships between lesions using graph attention networks, automatically identifying key lesions located along lymphatic drainage pathways and improving the clinical relevance of lesion identification; finally, it modifies the Deauville score by combining standardized uptake values ​​and metabolic heterogeneity indices of key lesions, allowing the scoring results to more accurately reflect the metabolic activity and spatial distribution characteristics of lesions, reducing the influence of subjective human factors, and improving the objectivity and consistency of efficacy assessment, thereby achieving full-process precision in lymphoma image analysis from automatic segmentation to intelligent scoring.

[0049] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0051] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0052] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0053] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0054] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An automatic segmentation and scoring method for FDG PET-CT lesions in lymphoma, characterized in that, Includes the following steps: S1. Calculate the standardized uptake value based on body weight and injection dose for the input FDG PET image, extract lymph node morphological features from the CT image using a multi-scale Hessian matrix enhancement filter, establish the correspondence between landmark points of the dual-modal image based on the extracted morphological features, and construct PET-CT image pairs. S2. Input PET-CT images into a dual-channel deep network. The CT channel uses deformable convolution to extract lymph node anatomical features, while the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features. Establish cross-modal gating fusion between feature maps at each decoding level of the two channels. Dynamically adjust the weight of metabolic features in the PET channel according to the tissue density features of the CT channel to suppress missegmentation of physiological uptake areas and output a preliminary segmentation mask for lymphoma lesions. S3. Perform three-dimensional connected domain analysis on the segmentation mask to extract independent lesion instances. By calculating the statistical kurtosis of the voxel-normalized uptake values ​​inside the lesion and the multi-scale energy distribution based on Haar wavelet transform, obtain the metabolic heterogeneity index of each lesion. Construct a topology graph with lesions as nodes and edge weights determined by the similarity of metabolic features and the proximity of anatomical location. Identify key lesions that are metabolically active and located in the lymphatic drainage path through a graph attention network. S4. Calculate the ratio of the maximum standardized uptake value of the key lesion to the average standardized uptake value of normal liver tissue, and adjust the Deauville score of the key lesion in combination with the metabolic heterogeneity index. When the metabolic heterogeneity index exceeds the preset threshold, adjust the Deauville score of the key lesion accordingly.

2. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 1, characterized in that: Step S1 includes the following steps: The normalized uptake value of each voxel in the FDG PET image is calculated. The parameter range of the multi-scale Hessian matrix enhancement filter is set based on the Henle unit value of the CT image. The morphological features of tubular and globular lymph node structures are enhanced by calculating the response of the eigenvalues ​​of the Hessian matrix in different scale spaces. Points with local maximum Hessian eigenvalues ​​are extracted from the enhanced CT images as morphological feature markers, and the center points of metabolically active regions are located in the corresponding PET images as metabolic feature markers. By optimizing the spatial correspondence between two sets of marker points using the iterative nearest point algorithm, a bimodal marker point correspondence based on morphological and metabolic features is established.

3. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 2, characterized in that: The specific process of constructing PET-CT image pairs is as follows: Based on the established correspondence between bimodal marker points, the spatial transformation parameters between PET images and CT images are calculated; The thin-plate spline interpolation algorithm is used to perform non-rigid spatial transformation on PET images, so that the metabolic feature markers of PET images are aligned with the morphological feature markers of CT images in three-dimensional space. The registration error index is calculated using the aligned dual-modal images. When the registration error exceeds the preset tolerance, the spatial transformation parameters are re-optimized until a PET-CT image pair that meets the registration accuracy requirements is obtained.

4. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 1, characterized in that: The CT channel uses deformable convolution to extract lymph node anatomical features, while the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features. The specific process of establishing cross-modal gating fusion between the feature maps of each decoding level of the two channels is as follows: In the CT channel, the sampling position of the deformable convolution kernel is adaptively adjusted according to the expected morphological characteristics of the lymph node to extract the anatomical features of the lymph node boundary and internal structure. In the PET channel, the metabolic correlation between voxels within the lesion is first calculated through a local self-attention mechanism, and then the metabolic correlation between different lesions is established through a global self-attention mechanism. At each decoding level, the anatomical features extracted from the CT channel are used as gating signals to control the weight allocation of metabolic features from the PET channel in cross-modal fusion, and dynamic fusion of dual-modal features is achieved through a gated recurrent unit.

5. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 4, characterized in that: The specific process of dynamically adjusting the weights of metabolic features in the PET channel based on the tissue density characteristics of the CT channel, suppressing missegmentation of physiological uptake areas, and outputting a preliminary segmentation mask for lymphoma lesions is as follows: Based on tissue density features extracted from CT channels, the anatomical locations of physiologically high-uptake organs, including the liver, kidneys, and bladder, are identified. A spatial mask is generated based on the identified physiological organ locations. During the cross-modal gating fusion process, the weight of the corresponding region's PET metabolic features is reduced. By iteratively optimizing the threshold setting of the gating signal, the model can distinguish between pathological lymph node uptake and physiological organ uptake. Based on the optimized dual-modal feature fusion results, a probability map of each voxel belonging to a lymphoma lesion is generated by the softmax function, and a preliminary segmentation mask is output by processing the probability threshold.

6. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 1, characterized in that: The process of performing three-dimensional connected component analysis on the segmentation mask to extract independent lesion instances, and obtaining the metabolic heterogeneity index of each lesion by calculating the statistical kurtosis of the voxel-normalized uptake values ​​within the lesion and the multi-scale energy distribution based on Haar wavelet transform is as follows: Three-dimensional morphological opening operations are performed on the initial segmentation mask to eliminate noise interference. Independent lesion instances in three-dimensional space are identified and labeled by a connected component labeling algorithm based on 26 neighborhoods. For each independent lesion instance, the statistical distribution characteristics of the normalized uptake values ​​of all voxels within it are calculated. The distribution kurtosis is calculated by the fourth-order central moment to quantify the concentration of metabolic activity. At the same time, the lesion region is subjected to multi-level Haar wavelet decomposition to extract wavelet energy coefficients at different scales and construct an energy distribution vector that characterizes the multi-scale features of metabolic distribution. By nonlinearly weighting and fusing statistical kurtosis features with multi-scale energy distribution vectors, a heterogeneity index that comprehensively reflects the metabolic inhomogeneity within lesions is generated.

7. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 6, characterized in that: The specific process of constructing a topological graph with lesions as nodes and edge weights determined by both metabolic feature similarity and anatomical location proximity, and then using a graph attention network to identify key lesions that are metabolically active and located along the lymphatic drainage path, is as follows: Each independent lesion instance is treated as a graph node. The node features include metabolic parameters and spatial coordinates. The metabolic feature similarity distance and the three-dimensional Euclidean distance between any two nodes are calculated. The edge weights are determined by the weighted harmonic mean of the two distances. Based on the anatomical features of medical images, key anatomical regions of the main lymphatic drainage pathways are identified. The shortest spatial distance from each lesion to the key anatomical region is calculated. During the message passing process of the graph attention network, higher attention weights are assigned to lesion nodes located on the lymphatic drainage pathways. After multiple rounds of iterative updates, the node with the highest feature salience is selected as the key lesion.

8. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 1, characterized in that: The specific process of calculating the ratio of the maximum standardized uptake value of the key lesion to the mean standardized uptake value of normal liver tissue, and then adjusting the Deauville score of the key lesion in conjunction with the metabolic heterogeneity index, is as follows: Multiple regions of interest were selected in the right lobe of normal liver tissue, and their average standardized uptake values ​​were calculated as the baseline reference values. The ratio of the maximum standardized uptake value of the key lesion to the baseline reference value was calculated, and the Deauville score was determined according to the standard. A scoring correction function is constructed based on the metabolic heterogeneity index. This function maps the metabolic heterogeneity index to a correction factor and then weights the correction factor with the Deauville score. When the metabolic heterogeneity index shows a highly uneven distribution of metabolism within the lesion, the score level is increased accordingly.

9. The automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma according to claim 8, characterized in that: When the metabolic heterogeneity index exceeds a preset threshold, the specific process for adjusting the Deauville score of the key lesions is as follows: Based on the statistical analysis results of large-scale clinical datasets, the grading threshold range of the metabolic heterogeneity index was determined, and the metabolic heterogeneity index was divided into different risk levels according to the preset threshold, establishing the correspondence between metabolic heterogeneity and score correction. When the metabolic heterogeneity index exceeds the first threshold but does not exceed the second threshold, the Deauville score is raised by one sub-level; when the metabolic heterogeneity index exceeds the second threshold, the Deauville score is raised by one full level. Establish a score adjustment decision tree, record the metabolic heterogeneity index, risk level, and corresponding score adjustment amount for each key lesion, forming a complete score correction traceability chain.

10. An automatic segmentation and scoring system for FDG PET-CT lesions of lymphoma, applied to the automatic segmentation and scoring method for FDG PET-CT lesions of lymphoma as described in any one of claims 1-9, characterized in that, Includes the following modules: The image preprocessing module is used to calculate the standardized uptake value based on body weight and injection dose for the input FDG PET image, extract the lymph node morphological features of the CT image using a multi-scale Hessian matrix enhancement filter, establish the correspondence between the marker points of the dual-modal image based on the extracted morphological features, and construct PET-CT image pairs. The lesion segmentation module is used to input PET-CT images into a dual-channel depth network. The CT channel uses deformable convolution to extract lymph node anatomical features, while the PET channel uses a local-global self-attention mechanism to extract metabolic distribution features. Cross-modal gating fusion is established between the feature maps of each decoding level of the two channels. The weight of metabolic features in the PET channel is dynamically adjusted according to the tissue density features of the CT channel to suppress missegmentation of physiological uptake areas and output a preliminary segmentation mask for lymphoma lesions. The lesion analysis module is used to perform three-dimensional connected domain analysis on the segmentation mask to extract independent lesion instances. By calculating the statistical kurtosis of the voxel-normalized uptake values ​​inside the lesion and the multi-scale energy distribution based on Haar wavelet transform, the metabolic heterogeneity index of each lesion is obtained. A topology graph with lesions as nodes and edge weights determined by the similarity of metabolic features and the proximity of anatomical location is constructed. Key lesions that are metabolically active and located in the lymphatic drainage path are identified through graph attention network. The scoring output module is used to calculate the ratio of the maximum standardized uptake value of the key lesion to the average standardized uptake value of normal liver tissue, and to correct the Deauville score of the key lesion in combination with the metabolic heterogeneity index. When the metabolic heterogeneity index exceeds the preset threshold, the Deauville score of the key lesion is adjusted accordingly.

Images