An artificial intelligence-based gastrointestinal malignant tumor metastasis risk assessment method and system

Abstract

The application discloses a kind of gastrointestinal malignant tumor metastasis risk assessment method and system based on artificial intelligence, it is related to gastrointestinal malignant tumor metastasis risk assessment field;First, the PET-CT three-dimensional image data of colorectal cancer patient is acquired, and the three-dimensional voxel matrix of liver region is constructed and extracted according to PET-CT three-dimensional image data;The voxel matrix of liver region is input into pre-trained convolutional neural network model, and the probability distribution matrix of each voxel is output;The basic confidence threshold of malignant tumor metastatic focus is set, and the candidate lesion is extracted according to the probability distribution matrix and the confidence threshold of malignant tumor metastatic focus;The maximum diameter of the candidate lesion is acquired, and the confidence threshold of malignant tumor metastatic focus is dynamically self-adapted according to the maximum diameter, and binary mask matrix is generated;According to binary mask matrix, the comprehensive risk score of colorectal cancer liver metastasis is calculated and classified;The application can enhance false positive inhibition capacity and improve risk assessment accuracy.

Description

Technical Field

[0001] This invention belongs to the field of risk assessment for metastasis of gastrointestinal malignant tumors and involves artificial intelligence technology. Specifically, it is an artificial intelligence-based method and system for risk assessment of metastasis of gastrointestinal malignant tumors. Background Technology

[0002] Colorectal cancer is one of the most common and deadliest malignant tumors worldwide. More than 50% of its fatal cases are closely related to liver metastasis. Early and accurate assessment of liver metastasis risk is of great significance for developing personalized treatment plans and improving patient prognosis.

[0003] Currently, clinical diagnosis mainly relies on PET-CT imaging, but traditional methods have the following limitations:

[0004] Subjectivity and complexity: It relies on doctors to manually delineate the liver area and lesions, which can lead to inter-observer variability and is time-consuming. The false positive rate (such as benign hyperplasia) and the rate of missed diagnosis of micrometastases are relatively high.

[0005] Insufficient multimodal data fusion: Fixed HU threshold segmentation is susceptible to anatomical variations and noise interference, and single-modal analysis is difficult to capture metabolic-anatomical heterogeneity features;

[0006] Risk assessment parameters are too simplistic: Traditional indicators (such as maximum SUV and lesion volume) only reflect local characteristics and lack quantification of global characteristics such as metabolic heterogeneity and spatial distribution.

[0007] Insufficient model specificity: Existing deep learning models have a high misclassification rate for false positive areas (such as vascular shadows) and do not model the dynamic relationship between lesion size and diagnostic confidence.

[0008] To address the aforementioned problems, this invention proposes a solution. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for assessing the metastasis risk of gastrointestinal malignant tumors based on artificial intelligence.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] In a first aspect, the present invention discloses an artificial intelligence-based method for assessing the metastatic risk of gastrointestinal malignant tumors, comprising the following steps:

[0012] Acquire PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix based on the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is ​​a voxel, and the voxel contains position coordinates, HU value of CT channel and SUV value of PET channel;

[0013] Set a dynamically adjustable HU threshold, and extract the three-dimensional voxel matrix of the liver region from the three-dimensional voxel matrix based on the HU threshold;

[0014] The voxel matrix of the liver region is input into a pre-trained 3D convolutional neural network model, and the probability distribution matrix of each voxel is output. The probability distribution matrix includes the probability Pm that the voxel belongs to a malignant tumor metastasis and the probability Pf that the voxel belongs to a false positive region.

[0015] Set a baseline confidence threshold P0 for malignant tumor metastases, and extract candidate lesions from the three-dimensional voxel matrix of the liver region based on the probability distribution matrix and the confidence threshold for malignant tumor metastases.

[0016] The maximum diameter of the candidate lesion is obtained, and the confidence threshold P of the malignant tumor metastasis is dynamically and adaptively adjusted according to the maximum diameter, and a binary mask matrix is ​​generated; the comprehensive risk score of colorectal cancer liver metastasis is calculated and classified according to the binary mask matrix.

[0017] Secondly, this invention discloses an artificial intelligence-based system for assessing the metastasis risk of gastrointestinal malignant tumors, comprising:

[0018] Data acquisition module: used to acquire PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix based on the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is ​​a voxel, and the voxel contains position coordinates, HU value of CT channel and SUV value of PET channel;

[0019] Data extraction module: used to set a dynamically adjustable HU threshold, and extract the three-dimensional voxel matrix of the liver region from the three-dimensional voxel matrix based on the HU threshold;

[0020] Data analysis module: used to input the voxel matrix of the liver region into a pre-trained 3D convolutional neural network model and output the probability distribution matrix of each voxel, the probability distribution matrix including the probability Pm that the voxel belongs to the malignant tumor metastasis and the probability Pf that the voxel belongs to the false positive region;

[0021] Data judgment module: used to set the basic confidence threshold P0 for malignant tumor metastases, and extract candidate lesions from the three-dimensional voxel matrix of the liver region based on the probability distribution matrix and the confidence threshold for malignant tumor metastases;

[0022] Risk assessment module: used to obtain the maximum diameter of the candidate lesions, dynamically and adaptively adjust the confidence threshold P of the malignant tumor metastasis based on the maximum diameter, and generate a binary mask matrix; calculate and classify the comprehensive risk score of colorectal cancer liver metastasis based on the binary mask matrix.

[0023] The present invention has the following beneficial effects:

[0024] 1. By dynamically adjusting the HU threshold and combining it with U-Net network edge recognition, individual anatomical differences and image noise interference are eliminated, significantly reducing the segmentation error between the liver region and non-liver regions;

[0025] 2. By adopting an adversarial training strategy to optimize the 3D convolutional neural network, the model specificity is improved through training with false positive samples, effectively distinguishing metastatic lesions from interference areas such as vascular artifacts, and enhancing the ability to suppress false positives.

[0026] 3. By dynamically adjusting the confidence threshold based on the maximum diameter of candidate lesions, sensitivity and specificity can be balanced. For example, the threshold can be relaxed for large lesions to avoid false screening, while the threshold can be tightened for small lesions to reduce missed diagnoses and improve the detection rate of micrometastases.

[0027] 4. By integrating the metabolic heterogeneity index (H), volume (V), and SUVmax into a linear weighted scoring system, and optimizing the weights through logistic regression, the accuracy of risk stratification is improved, providing more comprehensive quantitative evidence for clinical practice. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is an overall method block diagram of Embodiment 1 of the present invention;

[0030] Figure 2 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0031] Figure 3 This is a flowchart illustrating the setting of an adjustable HU threshold in Embodiment 1 of the present invention;

[0032] Figure 4 This is an overall system block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Application Overview:

[0035] Colorectal cancer is one of the leading causes of cancer-related deaths worldwide, with over 50% of these deaths closely linked to liver metastasis. Early and accurate assessment of liver metastasis risk is crucial for developing personalized treatment plans and improving patient prognosis. Currently, clinical practice primarily relies on imaging techniques such as PET-CT for metastatic lesion detection and risk assessment. PET-CT achieves lesion localization and qualitative analysis by fusing functional metabolic information with anatomical structural information, but its practical application still faces many challenges:

[0036] Subjectivity and complexity in image interpretation: Traditional methods rely on doctors manually delineating liver regions and identifying metastases, which is time-consuming and subject to inter-observer variability. False positives, such as high rates of missed diagnoses of benign hyperplasia, inflammatory areas, and micrometastases, affect diagnostic accuracy. Limitations of multimodal data fusion: Existing algorithms often use fixed thresholds, such as CT thresholding, to segment liver regions, but variations in patient anatomical structure, image noise, and partial volume effects can easily lead to segmentation errors. Furthermore, single-modal analysis from CT or PET alone cannot fully characterize the metabolic-anatomical heterogeneity of lesions. Limited risk assessment parameters: Traditional indicators such as maximum SUV and lesion volume only reflect local characteristics, lacking quantification of global features such as metabolic heterogeneity and spatial distribution, making it difficult to comprehensively assess metastatic risk. In recent years, deep learning technology has provided new insights into medical image analysis. 3D convolutional neural networks, through end-to-end learning, can extract multi-scale features from 3D images and have been successfully applied to tumor detection and segmentation. However, the following bottlenecks still exist in the assessment of liver metastasis risk: (1) the imaging features of false positive areas and metastatic lesions are highly overlapping, resulting in insufficient model specificity; (2) the dynamic correlation between lesion size and diagnostic confidence has not been effectively modeled; and (3) the automated construction of a multi-parameter risk scoring system is not yet mature.

[0037] To address these issues, researchers noted population-specific HU value distributions in liver tissue, indicating that a fixed threshold could lead to over- or under-segmentation. To mitigate false positives, they discovered subtle differences in the dual-modal features of PET-CT between metastatic and benign areas, necessitating the use of deep networks to capture multidimensional features. Further investigation revealed a non-linear relationship between lesion size and diagnostic confidence, requiring stricter thresholds for smaller lesions to avoid misdiagnosis. This led to a technical approach combining dynamic threshold segmentation, dual-modal feature fusion, and adaptive confidence adjustment.

[0038] Therefore, this application proposes the following technical solution.

[0039] Example 1

[0040] like Figure 1-3As shown, an artificial intelligence-based method for assessing the risk of metastasis in gastrointestinal malignancies includes the following steps: acquiring PET-CT three-dimensional image data of colorectal cancer patients to construct a three-dimensional voxel matrix, where each voxel contains location coordinates, HU value, and SUV value. A dynamically adjustable HU threshold is set to extract a voxel matrix of the liver region, which is then input into a pre-trained 3D convolutional neural network to output the probability that each voxel belongs to a metastatic lesion or a false positive region. After extracting candidate lesions based on a baseline confidence threshold, the confidence threshold is dynamically adjusted according to the maximum diameter of the lesion to generate a binary mask matrix, calculate a comprehensive risk score, and classify the lesion.

[0041] The three-dimensional voxel matrix construction integrates PET-CT dual-modality data through spatial alignment and inter-slice interpolation. Each voxel contains anatomical and metabolic information, providing a multi-dimensional data foundation for subsequent analysis. The dynamically adjusted HU threshold uses a three-dimensional Otsu algorithm to optimize the segmentation range, automatically adapting to individual differences by statistically analyzing the HU distribution across the entire image. The 3D convolutional neural network employs an adversarial training strategy, using generative adversarial networks to distinguish between true and false positive features, improving model specificity. The dynamic adjustment mechanism for the confidence threshold automatically corrects the diagnostic criteria based on the lesion diameter, relaxing the threshold for large lesions to avoid over-exclusion and tightening the threshold for small lesions to reduce misjudgments.

[0042] Specifically, spatial alignment of the 3D voxel matrix eliminates inter-modal registration errors, ensuring that the same voxel location contains the corresponding HU and SUV values. During dynamic thresholding, the 3D Otsu algorithm adds spatial continuity constraints to the traditional 2D algorithm to avoid generating isolated noise regions. The 3D convolutional neural network captures both local details and global context simultaneously through a multi-scale feature fusion module, and the skip connection structure preserves edge information, improving the detection capability of small lesions. The confidence threshold adjustment formula introduces a decay coefficient, establishing an inverse correlation between the diameter and the threshold. When the diameter increases, the threshold is appropriately reduced but remains higher than the base value, balancing sensitivity and specificity.

[0043] Compared to existing technologies, the traditional HU threshold method, which uses a fixed range for segmentation, results in large volume errors. Our proposed 3D Otsu algorithm reduces these segmentation errors. Traditional risk assessment uses only the SUVmax parameter; our approach integrates multiple parameters, including metabolic heterogeneity and volume, thus improving the accuracy of risk classification.

[0044] Through the above technical solutions, this application achieves individualized adaptation for liver region segmentation, effectively eliminating the influence of image noise and some volume effects. Dual-modal feature fusion and adversarial training strategies significantly improve the model's ability to distinguish between micrometastases and false positive regions. A dynamic confidence adjustment mechanism optimizes diagnostic criteria based on lesion size, reducing false alarm rates while maintaining high sensitivity. A multi-parameter scoring system comprehensively assesses metabolic activity and spatial distribution characteristics, providing more accurate risk assessment criteria for clinical practice.

[0045] This application further proposes a technical solution that spatially aligns functional metabolic imaging data with anatomical structural imaging data when constructing a three-dimensional voxel matrix, and generates the three-dimensional voxel matrix through inter-slice interpolation.

[0046] Functional metabolic imaging data refers to standardized uptake distribution data acquired through PET scans, specifically using positron emission tomography (PET) to characterize the intensity of tissue metabolic activity. Anatomical imaging data refers to Henlein unit distribution data acquired through CT scans, specifically using X-ray computed tomography (CT) to characterize tissue density differences. Spatial alignment refers to the coordinate system unification of PET and CT images of the same patient, achieved using rigid registration algorithms to eliminate spatial offsets caused by respiratory motion or differences in equipment acquisition. Interslice interpolation refers to the interpolation calculation of raw data with different scan slice thicknesses to generate continuous three-dimensional voxel data, specifically using cubic spline interpolation algorithms to solve the step artifact problem caused by inconsistent slice spacing.

[0047] Specifically, PET images provide metabolic hotspot information, while CT images provide anatomical structure information. A rigid registration algorithm maps the coordinates of each voxel in the PET image to the coordinate system of the CT image, achieving pixel-level spatial matching of the dual-modal data. For slices with inconsistent slice thickness in the original data, a cubic spline interpolation algorithm is used to perform interpolation calculations along the axial direction, generating a continuous three-dimensional voxel matrix with uniform spatial resolution. Each voxel contains both HU and SUV values. The resulting three-dimensional data structure accurately reflects the correspondence between metabolic activity and anatomical location, providing spatially consistent multimodal features for subsequent dynamic thresholding.

[0048] Compared to existing technologies, traditional methods typically process PET and CT data separately or use fixed thresholds for single-modal segmentation, leading to spatial misalignment between metabolic hotspots and anatomical structures, which easily results in segmentation errors. This application eliminates image offset through dual-modal spatial alignment and ensures the continuity of three-dimensional data by combining inter-slice interpolation, enabling subsequent analysis to accurately correlate metabolically abnormal regions with anatomical locations and avoiding false positives or missed detections due to data inconsistencies.

[0049] Through the above technical solution, this application effectively solves the difficulty of feature fusion caused by spatial mismatch in multimodal images, and improves the accuracy of liver region segmentation. The continuity and multimodal properties of the three-dimensional voxel matrix provide a reliable data foundation for subsequent dynamic threshold adjustment and lesion identification, reducing the risk of misjudgment caused by data artifacts or spatial offset.

[0050] This application further proposes a process for setting a dynamically adjustable HU threshold, including performing Gaussian filtering denoising on the CT channels in the PET-CT three-dimensional voxel matrix, extracting the HU values ​​of all voxels, and calculating the standard deviation σH of all HU values; setting different T values, dividing voxels into foreground and background based on the comparison between HU values ​​and T values, and calculating:

[0051] σ 2 (T)=ω1(T)ω2(T)[μ1(T)-μ2(T)] 2 , where ω1 and ω2 are the voxel percentages of the foreground and background, respectively, and μ1 and μ2 are the mean HU values ​​of the foreground and background, respectively;

[0052] Choose to make σ 2 (T) The maximum T value is used as the initial value Tinitial; the initial HU threshold interval [Tlow, Thigh] is set, where Tlow is Tinitial-kσH and Thigh is Tinitial+kσH; the volume Vcurrent of the candidate liver region is calculated and it is determined whether it exceeds the anatomical range [Vmin, Vmax]. The volume Vcurrent is made to fall into a reasonable range by dynamically adjusting the k value.

[0053] Gaussian filtering denoising refers to convolving CT channel data using a two-dimensional Gaussian kernel, for example, using a Gaussian kernel with a standard deviation of 1.5 to eliminate random noise generated during image acquisition. Foreground and background are distinguished by comparing the T value and the HU value; when the T value is less than the HU value, it is classified as foreground, and vice versa. The volume Vcurrent of the candidate region is calculated by counting the number of voxels in the liver candidate region and multiplying it by the unit volume corresponding to each voxel. Dynamic adjustment of the k value refers to feedback adjustment based on the deviation between the volume Vcurrent of the liver candidate region and the preset anatomical range [Vmin, Vmax]. For example, when the volume exceeds the maximum threshold Vmax, the k value is increased by 0.2 to eliminate abnormal segmentation caused by some volume effects.

[0054] Specifically, after Gaussian filtering preprocessing, the HU value map of each voxel is statistically analyzed, and the inter-class variance corresponding to different HU thresholds is calculated. An initial threshold Tinitial that maximizes the variance is selected. Using this initial threshold Tinitial as the center, a dynamic threshold range is determined by combining the standard deviation σH of the HU values ​​across the entire image and an empirical coefficient k. The volume Vcurrent of the liver candidate region under the current threshold is calculated. When the volume Vcurrent is detected to exceed the reasonable anatomical range of liver tissue, the threshold range is adjusted by increasing or decreasing the k value. For example, when the volume exceeds 1500 cm³, the k value adjustment mechanism is triggered. This process is repeated until the volume Vcurrent stabilizes within a preset reasonable anatomical range, such as a liver volume range of 800-1200 cm³.

[0055] Compared to existing technologies, traditional fixed-threshold methods directly use a fixed HU range recommended in the literature for segmentation, such as a uniform threshold of 40-90 HU, which cannot adapt to the natural variations in liver density among individuals. Our proposed solution, however, dynamically sets the threshold range by incorporating standard deviation. For example, for patients with cirrhosis, the upper limit of the threshold can be automatically extended to 110 HU to cover fibrotic tissue. Existing technologies use a two-dimensional Otsu algorithm that only considers information from a single slice; our solution calculates inter-class variance in three-dimensional space, effectively distinguishing adjacent organs with similar density to the liver.

[0056] Through the above technical solution, this application solves the problem of missegmentation of the liver region caused by individual differences and image noise in traditional segmentation methods. It accurately distinguishes the liver from surrounding tissues through three-dimensional spatial analysis, and the dynamic threshold adjustment mechanism can adapt to the anatomical variations of different patients, avoiding abnormal candidate regions caused by partial volume effects, and providing an accurate liver region data basis for subsequent metastatic lesion detection.

[0057] This application further proposes a liver region segmentation method based on HU threshold screening and U-Net network edge recognition, specifically including: screening candidate liver regions according to HU threshold, using U-Net network to identify liver edges, accurately segmenting the liver region, and outputting a three-dimensional voxel matrix of the liver region.

[0058] The HU threshold screening process involves initially defining candidate liver regions using the range of Hounsfield unit values ​​from CT channels. This can be achieved using dynamically adjusted threshold intervals, such as determining the initial threshold range through 3D histogram analysis and inter-class variance maximization algorithms, and then dynamically adjusting the threshold boundaries based on the volume of the candidate liver regions. This step can quickly exclude regions with HU values ​​significantly deviating from liver tissue, such as bone and fat.

[0059] U-Net edge recognition refers to pixel-level segmentation using a convolutional neural network with an encoder-decoder structure. Specifically, it fuses shallow high-resolution features with deep semantic features through skip connections. For example, multi-scale features are extracted through 4-level downsampling in the encoding stage, and spatial details are restored by combining deconvolution and skip connections in the decoding stage. During network training, the Dice loss function can be used to optimize the continuity of segmentation boundaries, and pre-trained weights can be loaded to accelerate convergence.

[0060] Specifically, this method first generates candidate regions using HU thresholding, reducing the size of the voxels to be processed and minimizing computational redundancy. Subsequently, the liver candidate regions are input into a U-Net network for fine segmentation. The network uses 3×3×3 convolutional kernels to extract local texture features layer by layer and gradually expands the receptive field through max pooling. In the decoding stage, upsampling layers and skip connections work together to fuse low-level edge information with high-level semantic features, effectively identifying blurred boundaries caused by partial volumetric effects. The output layer uses a Sigmoid activation function to generate a probability map, which is then thresholded into a binary mask, ultimately outputting an accurate three-dimensional voxel matrix of the liver.

[0061] Compared to existing technologies, traditional methods rely on a fixed HU threshold for liver segmentation, which cannot adapt to individual density differences and image noise interference. For example, the liver density of patients with cirrhosis may be lower than the standard threshold range, leading to missing segmentation regions. Our proposed solution, however, combines dynamic threshold selection with deep learning edge detection, which can suppress noise effects while maintaining high recall. Experiments show that this method improves the Dice coefficient compared to the single threshold method in liver segmentation tasks, and its segmentation accuracy in low-contrast regions is significantly better than traditional morphological post-processing methods.

[0062] Through the above technical solution, this application achieves accurate three-dimensional segmentation of the liver region, solving the segmentation error problem caused by anatomical variations and image noise in traditional fixed-threshold methods. By employing a two-stage strategy of initial localization using dynamic thresholding and fine-tuning using deep learning, the influence of vascular artifacts and partial volume effects on the segmentation results is effectively suppressed, providing accurate anatomical constraints for subsequent metastatic lesion detection. This method, while ensuring computational efficiency, keeps liver volume measurement errors within clinically acceptable ranges, providing a reliable spatial benchmark for automated risk assessment systems.

[0063] This application further proposes a process for outputting a dual-channel probability distribution matrix through a three-dimensional convolutional neural network model. This process includes acquiring positron emission tomography (PET) images of colorectal cancer patients and pathologically confirmed metastatic lesion annotation data. An adversarial training strategy is adopted to use PET images of false-positive regions as negative samples to input the discriminator in order to optimize the specificity of the feature extractor. High-level semantic features are extracted and contextualized through multi-level convolution and pooling. Skip connections combine low-level details with high-level semantic information. The probability of each voxel belonging to a malignant tumor metastatic lesion and the probability of each voxel belonging to a false-positive region are output through a normalized exponential function layer.

[0064] The adversarial training strategy refers to training the discriminator and feature extractor within the generative adversarial network (GAN) framework through a game-like process. This can be achieved by alternately optimizing the parameters of the generator and discriminator. Feedback from the discriminator's identification of false positive samples drives the feature extractor to generate more discriminative feature representations. Multi-level convolution and pooling refers to a feature extraction architecture formed by alternating stacks of multiple convolutional and pooling layers. This can be implemented using a five-layer convolutional structure combined with max pooling, extracting spatially correlated 3D semantic features from the original image through layer-by-layer abstraction. Skip connections involve concatenating the feature maps of shallow and deep networks. This can be achieved by upsampling 3D transposed convolutions and fusing them with features from the corresponding layer, enhancing the model's ability to identify minute lesions by preserving details at different scales. The normalized exponential function layer converts the neural network output into a non-linear function of probability distribution. This can be achieved using dual-channel probability mapping and normalization calculations, quantifying the confidence level of voxels belonging to different categories through probability values.

[0065] Specifically, during training, pathologically labeled data serves as supervisory signals, ensuring the model learns the true distribution of metastatic lesion features. In the adversarial training framework, the discriminator is input with false-positive region samples, forcing the feature extractor to suppress interference patterns associated with false-positive regions when generating features. Multi-level convolution extracts local texture and global morphological features layer by layer using 3D convolutional kernels, while pooling reduces feature map resolution while enhancing robustness to deformation. Skip connections fuse low-level high-resolution features with high-level semantic features during the decoding stage, achieving complementary optimization of multi-scale information. The normalized exponential function layer probabilistically normalizes the dual-channel output of each voxel, obtaining the probability distribution matrix of malignant tumor metastases and false-positive regions, providing a reliable basis for subsequent threshold selection.

[0066] Compared to existing technologies, traditional methods typically train classification models using single-modality data, lacking targeted optimization for false positive samples, resulting in insufficient specificity. Existing 3D convolutional networks mostly employ unidirectional encoding structures, losing low-level detail information during feature fusion, affecting the detection rate of minute lesions. This approach explicitly models the feature distribution of false positive regions through adversarial training mechanisms, combined with a multi-scale fusion strategy using skip connections, effectively improving the model's ability to resolve complex image patterns.

[0067] Through the above technical solutions, this application can significantly reduce the false positive rate, enhance the model's ability to fuse deep features of multimodal image data, achieve accurate probability prediction of malignant tumor metastases, and provide a reliable quantitative basis for subsequent dynamic threshold screening.

[0068] This application further proposes a technical solution for setting a baseline confidence threshold and performing probability comparison when extracting candidate lesions. Specifically, it includes: setting a baseline confidence threshold P0; comparing the malignant tumor metastasis probability Pm of each voxel in the three-dimensional voxel matrix with P0; and marking the voxel as a candidate lesion when Pm reaches or exceeds P0.

[0069] The baseline confidence threshold P0 is a critical probability value used to distinguish malignant tumor metastases from false-positive regions. Specifically, it can be determined using the median probability distribution of metastatic lesions in true-positive cases from clinical validation data, for example, 0.7. This threshold serves as a preliminary screening criterion to filter out low-probability interfering voxels. The probability comparison mechanism involves judging the predicted probability of each voxel point by point. This can be implemented using a parallel matrix element comparison algorithm, which can simultaneously process millions of voxel data points, ensuring computational efficiency.

[0070] Specifically, in the probability distribution matrix output by the pre-trained 3D convolutional neural network, each voxel has a quantified probability value representing that it belongs to a malignant tumor metastasis. The baseline confidence threshold P0 is set based on the probability distribution characteristics of metastatic lesions in a large-scale clinical dataset, for example, setting the threshold to the 85th percentile of the probability distribution. When traversing all voxels, only voxels with probability values ​​higher than this threshold are retained, forming the spatial distribution of candidate lesions. This process, by establishing a mapping relationship between probability and anatomical location, eliminates voxels with low probability values ​​in false positive regions while retaining regions with typical metastatic lesion characteristics.

[0071] Compared to existing technologies, traditional methods often employ fixed thresholds or single-modal data for lesion extraction, such as fixed-range segmentation based solely on CT values. These methods cannot effectively distinguish false-positive areas caused by inflammation or vascular abnormalities. This proposed solution achieves precise filtering of false-positive areas by fusing probabilistic predictions from multimodal image features and combining them with dynamically optimized confidence thresholds. Furthermore, the probabilistic comparison mechanism, based on voxel-level analysis in full three-dimensional space, significantly improves the spatial continuity of detection compared to methods based on two-dimensional slicing or manual delineation.

[0072] Through the above technical solutions, this application effectively reduces the false positive rate caused by overlapping image features. The objective screening mechanism based on probability distribution avoids the subjective bias of manual threshold setting, ensuring the repeatability and accuracy of the candidate lesion extraction process. By retaining the voxel set with high-probability features, a reliable data foundation is provided for subsequent lesion diameter calculation and adaptive threshold adjustment, thereby improving the specificity of the overall system while ensuring sensitivity.

[0073] This application further proposes a process for calculating the maximum diameter of a candidate lesion, which includes using connected component analysis and convex hull algorithm to calculate the maximum diameter based on the position coordinates of the voxel corresponding to the candidate lesion.

[0074] Connected component analysis refers to the technique of clustering adjacent candidate lesion voxels in three-dimensional space into independent connected regions. This can be implemented using region growing algorithms or neighbor-search-based labeling algorithms to eliminate discrete noise interference and distinguish between different lesions. Convex hull algorithm refers to the technique of calculating the minimum convex polyhedron that encloses the geometry of candidate lesions in the three-dimensional point cloud. This can be implemented using Graham's scan method or the fast convex hull algorithm to eliminate interference from the internal concave structure of the lesion on the calculation of the maximum diameter. Euclidean distance calculation is a geometric distance measurement method based on voxel space coordinates. It can be implemented by traversing the distance between any two points in the convex hull vertex set and selecting the maximum value, to accurately characterize the true extent of the lesion in three-dimensional space.

[0075] Specifically, candidate lesion voxels may be scattered or interconnected in three-dimensional space. Connected component analysis can classify them into independent lesion entities, preventing the distance between different lesions from being mistakenly included in the same lesion. Furthermore, a convex hull algorithm is applied to the voxel coordinate set of each independent lesion to construct a minimal convex polyhedron model that encloses all voxels, eliminating measurement biases caused by uneven density or concave shape within the lesion. Finally, the distance between all possible point pairs is calculated by traversing the vertex set of the convex hull, and the maximum Euclidean distance is selected as the true diameter of the lesion. This method overcomes the problem of dimensional information loss caused by viewing angle limitations or slice thickness in traditional two-dimensional projection measurements.

[0076] Compared to existing technologies, traditional methods typically employ maximum diameter measurement of two-dimensional slices or estimation of the equivalent sphere diameter based on the number of voxels. These measurements are easily affected by the spatial orientation of the lesion and partial volume effects. This proposed method, through three-dimensional connected domain analysis and geometric convex hull modeling, directly and accurately characterizes the three-dimensional spatial topology of the lesion. It is particularly suitable for lobed or irregularly shaped metastatic lesions, significantly improving the spatial consistency of diameter measurements.

[0077] Through the above technical solution, this application achieves accurate modeling of the true three-dimensional morphology of candidate lesions, effectively solving the problem of error accumulation caused by missing dimensions or morphological simplification in traditional measurement methods. By eliminating interference from internal structures through geometric convex hulls, it ensures that the maximum diameter calculation only reflects the external extension characteristics of the lesion, providing more reliable morphological parameters for subsequent risk scoring.

[0078] This application further proposes a process for dynamically and adaptively adjusting the confidence threshold of malignant tumor metastases based on the maximum diameter of candidate lesions. This includes calculating the confidence threshold P based on the maximum diameter Dmax of the candidate lesions, expressed as P=P0×(1-Dmax / (Dmax+s)), where s is the attenuation coefficient, and s is smaller when the maximum diameter Dmax is larger. Only voxels with a probability Pm≥confidence threshold P are retained to generate a binary mask matrix that marks metastases and the background.

[0079] The maximum diameter refers to the longest distance in the three-dimensional space of the candidate lesion obtained by connected component analysis and convex hull algorithm. Specifically, it can be achieved by calculating the voxel coordinates using the Euclidean distance formula. This parameter reflects the spatial distribution characteristics of the lesion and is used to establish a correlation model between size and diagnostic reliability.

[0080] Furthermore, the formula for calculating the maximum diameter is:

[0081] , where (xi,yi,zi) and (xj,yj,zj) are the position coordinates of the voxels corresponding to the candidate lesions;

[0082] The attenuation coefficient *s* is a parameter used to adjust the dynamic change range of the confidence threshold. Its specific range can be determined empirically based on clinical data statistics; for example, the value of *s* can be adjusted within the range of 5mm to 15mm. Its function is to balance the need for false positive suppression in large lesions with the need to ensure sensitivity in small lesions. The dynamic confidence threshold *P* is a non-linear function constructed based on the baseline threshold *P0* and the lesion size. Specifically, it uses mathematical transformations to map the lesion diameter to the threshold adjustment coefficient, ensuring that large lesions meet higher confidence requirements, while small lesions can accept relatively lenient screening conditions.

[0083] Specifically, the maximum diameter of candidate lesions is calculated using three-dimensional spatial geometry and then input into the threshold adjustment function. The attenuation coefficient s is adjusted inversely according to the lesion size. For example, when Dmax exceeds 20mm, s can be set to 5mm to enhance the threshold attenuation. In this case, the confidence threshold P rises rapidly with the diameter, prompting the system to retain only high-confidence voxels. Conversely, when Dmax is less than 10mm, s can be set to 15mm to slow down the threshold rise and prevent small lesions from being lost due to over-filtering. Through this dynamic mechanism, the generation process of the binarized mask matrix can adaptively adjust the screening strictness according to the lesion size, suppressing large false positive areas such as vascular artifacts while retaining small metastatic lesions with low-probability characteristics but conforming to anatomical distribution patterns.

[0084] Compared to existing technologies, traditional methods use fixed confidence thresholds to screen voxels, failing to differentiate the diagnostic reliability of lesions of different sizes. This leads to large lesions appearing as "holes" due to low probability values ​​for some voxels, or small lesions being falsely filtered out due to excessively high thresholds. This proposed solution establishes a dynamic correlation model between lesion diameter and confidence thresholds, enabling spatially adaptive voxel screening criteria. This reduces false positive interference while ensuring the integrity of large lesion segmentation, and maintains sensitivity for detecting small lesions.

[0085] Through the above technical solution, this application effectively solves the problems of high false positive rates and missed diagnoses of small metastases caused by the lack of dynamic correlation between lesion size and diagnostic confidence in traditional methods. Through a size-adaptive threshold adjustment mechanism, the system can differentiate the voxel screening process for lesions of different sizes. For example, it can increase the screening stringency for lesions with a diameter greater than 15 mm to suppress vascular artifact interference, while appropriately relaxing the conditions for lesions with a diameter less than 8 mm to avoid missing early metastasis signals. This achieves an optimized balance between sensitivity and specificity, improving the accuracy of liver metastasis risk assessment.

[0086] This application further proposes a process for calculating and rating a comprehensive risk score based on a binary mask matrix, including: obtaining the maximum value of SUV (SUVmax) among all voxels marked as 1 in the binary mask matrix; calculating the volume V of the metastatic lesion and the metabolic heterogeneity index H based on all voxels marked as 1 in the binary mask matrix; the volume V of the metastatic lesion is calculated based on the number of voxels marked as 1; the formula for calculating the metabolic heterogeneity index H is: H = Σ|SUV_i -SUV_avg|² / N; where SUV_i is the SUV value of the marginal voxels, SUV_avg is the average SUV value of the non-marginal voxels, and N is the number of marginal voxels; the comprehensive risk score is a linear weighted sum of volume V, metabolic heterogeneity index H, and SUVmax, specifically calculated as: R = α*SUVmax + β*V + γ*H, where α, β, and γ are weighting coefficients obtained through logistic regression training; matching the comprehensive risk score with a preset grading standard, and outputting low-risk, intermediate-risk, or high-risk signals for colorectal cancer liver metastases.

[0087] The metabolic heterogeneity index H is a quantitative indicator calculated by normalizing the sum of squared differences in the SUV values ​​of edge voxels and non-edge voxels. Specifically, it can be achieved by extracting edge voxels from the lesion region in a binary mask using an edge detection algorithm and calculating the degree of dispersion of their metabolic activity differences from those of internal voxels. This index is used to characterize the non-uniformity of metabolic distribution within the tumor and reflects the spatial heterogeneity of tumor biological behavior.

[0088] Volume V refers to the three-dimensional spatial occupancy parameter of the lesion, calculated by multiplying the number of voxels marked as metastases in a statistically binary mask by the volume of a single voxel. Specifically, this can be achieved using three-dimensional connected component analysis combined with voxel spatial resolution conversion. This parameter is used to assess the anatomical extent of the metastatic lesion.

[0089] SUVmax refers to the maximum normalized uptake value extracted from the binary masked labeled region, which can be achieved by traversing the PET channel data of all labeled voxels and recording the peak values. This parameter is used to characterize the highest level of metabolic activity in the lesion area.

[0090] Linear weighted sum R refers to combining three independent parameters—volume, metabolic heterogeneity, and SUVmax—into a single scoring index by weighting coefficients. Specifically, this can be achieved by training the contribution coefficients of each parameter using a logistic regression model based on clinical follow-up data. This scoring mechanism optimizes the multi-feature fusion strategy through a data-driven approach.

[0091] Specifically, firstly, a set of voxels labeled as metastatic lesions is extracted from a binary mask using a 3D image processing algorithm, and the spatial volume occupied by these voxels is calculated as a quantitative indicator of the degree of anatomical invasion. Next, edge detection technology is employed to distinguish the boundaries and core regions of the lesions. By comparing the metabolic activity differences between edge and non-edge voxels, a mathematical model reflecting the heterogeneity within the tumor is constructed. Simultaneously, the maximum metabolic activity value within the lesion region is recorded to capture local malignancy characteristics. Finally, the parameters from the three dimensions are integrated into a comprehensive score using machine learning-optimized weighting coefficients, and a risk level is output based on a preset clinical grading threshold.

[0092] Compared to existing technologies, traditional risk assessment methods often use single parameters such as SUVmax or lesion diameter for grading, failing to integrate indicators reflecting the biological complexity of tumors, such as metabolic heterogeneity. This proposed solution, by designing a metabolic heterogeneity index, incorporates spatial distribution heterogeneity into the scoring system for the first time. Combined with an adaptive weighting mechanism, this allows the assessment model to simultaneously capture the metabolic activity, anatomical invasion extent, and internal heterogeneity characteristics of lesions. For example, existing technologies typically use global standard deviation to calculate metabolic heterogeneity, while this solution calculates it through edge and non-edge partitions, more accurately characterizing spatial heterogeneity patterns.

[0093] Through the above technical solution, this application solves the problem of insufficient accuracy in grading caused by the simplification of parameters in traditional risk assessment, and achieves the integrated analysis of three-dimensional characteristics: metabolic activity, anatomical invasion, and spatial heterogeneity. By calculating edge-specific metabolic differences, it effectively distinguishes the biological behavioral differences between diffuse and focal metastases; through dynamic weight allocation optimized by machine learning, it avoids the subjective bias of manually setting weights, making the scoring system more consistent with the statistical regularities of clinical prognostic data. The final comprehensive risk score output can provide clinicians with multidimensional quantitative evidence covering morphology, functional metabolism, and spatial heterogeneity, supporting personalized treatment decisions.

[0094] Example 2

[0095] like Figure 4 As shown, an artificial intelligence-based system for assessing the risk of metastasis of gastrointestinal malignant tumors includes:

[0096] Data acquisition module: used to acquire PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix based on the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is ​​a voxel, and the voxel contains position coordinates, HU value of CT channel and SUV value of PET channel;

[0097] Data extraction module: used to set a dynamically adjustable HU threshold, and extract the three-dimensional voxel matrix of the liver region from the three-dimensional voxel matrix based on the HU threshold;

[0098] Data analysis module: used to input the voxel matrix of the liver region into a pre-trained 3D convolutional neural network model and output the probability distribution matrix of each voxel, the probability distribution matrix including the probability Pm that the voxel belongs to the malignant tumor metastasis and the probability Pf that the voxel belongs to the false positive region;

[0099] Data judgment module: used to set the basic confidence threshold P0 for malignant tumor metastases, and extract candidate lesions from the three-dimensional voxel matrix of the liver region based on the probability distribution matrix and the confidence threshold for malignant tumor metastases;

[0100] Risk assessment module: used to obtain the maximum diameter of the candidate lesions, dynamically and adaptively adjust the confidence threshold P of the malignant tumor metastasis based on the maximum diameter, and generate a binary mask matrix; calculate and classify the comprehensive risk score of colorectal cancer liver metastasis based on the binary mask matrix.

[0101] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0102] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0103] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence, characterized in that, The steps include: Obtain the PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix based on the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is a voxel, and the voxel includes a position coordinate, the HU value of the CT channel, and the SUV value of the PET channel; Set a dynamically adjustable HU threshold, and extract the three-dimensional voxel matrix of the liver region from the three-dimensional voxel matrix based on the HU threshold; The process of setting the dynamically adjustable HU threshold includes: Perform Gaussian filtering denoising on the HU values of the voxels in the three-dimensional voxel matrix, extract the HU values of all voxels, and calculate the standard deviation σH of all HU values; Set different T values, divide the voxels into foreground and background according to the comparison result of the HU value and the T value, and calculate: σ 2 (T)=ω1(T)ω2(T)[μ1(T)-μ2(T)] 2 , where ω1 and ω2 are the voxel percentages of the foreground and background, respectively, and μ1 and μ2 are the mean HU values ​​of the foreground and background voxels, respectively; Choose to make σ 2 (T) The maximum value of T is used as the initial value Tinitial; Set the initial HU threshold [Tlow, Thigh]; where Tlow is Tinitial - k * σH, Thigh is Tinitial + k * σH, σH is the standard deviation of all HU values, and k is an empirical coefficient; Obtain all voxels with HU values within the HU threshold [Tlow, Thigh] and label them as the liver candidate region, calculate the volume Vcurrent of the liver candidate region, and determine whether it exceeds the preset anatomical range [Vmin, Vmax]: If the volume Vcurrent > Vmax: increase the value of k and expand the threshold range; If the volume Vcurrent < Vmin: decrease the value of k and shrink the threshold range; Repeat adjusting the value of k until the volume Vcurrent ∈ [Vmin, Vmax]; Input the three-dimensional voxel matrix of the liver region into a pre-trained 3D convolutional neural network model, and output the probability distribution matrix of each voxel. The probability distribution matrix includes the probability Pm that the voxel belongs to a malignant tumor metastasis focus and the probability Pf that the voxel belongs to a false positive region; Set the basic confidence threshold P0 for malignant tumor metastasis foci, and extract candidate lesions from the three-dimensional voxel matrix of the liver region according to the probability distribution matrix and the confidence threshold P0 for malignant tumor metastasis foci; Obtain the maximum diameter of the candidate lesion, dynamically adjust the confidence threshold P that the voxel belongs to a malignant tumor metastasis focus according to the maximum diameter, and generate a binary mask matrix; calculate the comprehensive risk score and grade of colorectal cancer liver metastasis according to the binary mask matrix.

2. The method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence according to claim 1, characterized in that, The PET-CT three-dimensional image data includes: functional metabolic imaging data and anatomical structure imaging data; The process of constructing the three-dimensional voxel matrix includes: performing spatial alignment and interlayer interpolation on the functional metabolic imaging data and the anatomical structure imaging data to obtain a three-dimensional voxel matrix.

3. The method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence according to claim 1, characterized in that, The process of extracting the three-dimensional voxel matrix of the liver region from the three-dimensional voxel matrix includes: Screen out the liver candidate region according to the HU threshold, use the U-Net network to identify the edge of the liver candidate region, accurately segment the liver region, and output the three-dimensional voxel matrix of the liver region.

4. The method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence according to claim 1, characterized in that, The process of outputting the probability distribution matrix of each voxel through the 3D convolutional neural network model includes: Obtain PET-CT images of colorectal cancer patients and labeled data of pathologically confirmed malignant tumor metastases; An adversarial training strategy was adopted, using CT images of false positive regions in the annotated data of malignant tumor metastases as negative samples to input the discriminator and optimize the specificity of the feature extractor. High-level semantic features are extracted through multi-level convolution and pooling, and context fusion is used. Skip connections combine low-level details with high-level semantic information. The Softmax layer outputs the probability Pm that each voxel belongs to a malignant tumor metastasis and the probability Pf that each voxel belongs to a false positive region.

5. The method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence according to claim 1, characterized in that, The process of extracting candidate lesions from the three-dimensional voxel matrix of the liver region based on the probability distribution matrix and the confidence threshold P0 of the malignant tumor metastases includes: Set the baseline confidence threshold P0; The probability Pm of each voxel is compared with the baseline confidence threshold P0: if the probability Pm ≥ the baseline confidence threshold P0, then the voxel is marked as a candidate lesion.

6. The method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence according to claim 1, characterized in that, The calculation process for the maximum diameter of the candidate lesion includes: The maximum diameter Dmax of the candidate lesion is calculated by using connected component analysis and convex hull algorithm, based on the position coordinates (xi,yi,zi) and (xj,yj,zj) of the voxels corresponding to the candidate lesion. The calculation formula is: .

7. The method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence according to claim 1, characterized in that, The process of dynamically adjusting the confidence threshold P for malignant tumor metastases based on the maximum diameter includes: Calculate the confidence threshold P based on the maximum diameter Dmax of the candidate lesion: P = P0 × (1 - Dmax / (Dmax + s)), where s is the attenuation coefficient, and the larger the maximum diameter Dmax is, the smaller s is; Only voxels with Pm ≥ confidence threshold P are retained. Voxels that meet the confidence threshold are marked as 1, and the rest are marked as 0, thus generating a binary mask matrix.

8. The method for assessing the risk of metastasis of gastrointestinal malignant tumors based on artificial intelligence according to claim 7, characterized in that, The process of calculating and classifying the comprehensive risk score for colorectal cancer liver metastasis based on the aforementioned binarized mask matrix includes: Obtain the maximum value of SUV among all voxels marked as 1 in the binarized mask matrix: SUVmax; Calculate the volume V and the metabolic heterogeneity index H based on all voxels marked as 1 in the binarized mask matrix; The formula for calculating the metabolic heterogeneity index H is: H = Σ|SUV_i - SUV_avg|² / N; Where SUV_i is the SUV value of the edge voxels, SUV_avg is the average SUV value of the non-edge voxels, and N is the number of edge voxels; The comprehensive risk score is a linear weighted sum of volume V, metabolic heterogeneity index H, and SUVmax, calculated using the following formula: R = α*SUVmax + β*V + γ*H, where α, β, and γ are weight coefficients obtained through logistic regression training; The comprehensive risk score is matched with the preset grading criteria to output low-risk, medium-risk, or high-risk signals for colorectal cancer liver metastasis.

9. An artificial intelligence-based system for assessing the risk of metastasis of gastrointestinal malignant tumors, characterized in that, The method for assessing the metastatic risk of gastrointestinal malignancies based on artificial intelligence, as described in any one of claims 1 to 8, includes: Data acquisition module: used to acquire PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix based on the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is ​​a voxel, and the voxel contains position coordinates, HU value of CT channel and SUV value of PET channel; Data extraction module: used to set a dynamically adjustable HU threshold, and extract the three-dimensional voxel matrix of the liver region from the three-dimensional voxel matrix based on the HU threshold; Data analysis module: used to input the voxel matrix of the liver region into a pre-trained 3D convolutional neural network model and output the probability distribution matrix of each voxel, the probability distribution matrix including the probability Pm that the voxel belongs to the malignant tumor metastasis and the probability Pf that the voxel belongs to the false positive region; Data judgment module: used to set the basic confidence threshold P0 for malignant tumor metastases, and extract candidate lesions from the three-dimensional voxel matrix of the liver region based on the probability distribution matrix and the confidence threshold for malignant tumor metastases; Risk assessment module: used to obtain the maximum diameter of the candidate lesions, dynamically adjust the confidence threshold P of the voxel as a malignant tumor metastasis based on the maximum diameter, and generate a binary mask matrix; calculate and classify the comprehensive risk score of colorectal cancer liver metastasis based on the binary mask matrix.

Images