Digestive tract malignant tumor metastasis risk assessment method and system based on artificial intelligence

Abstract

The invention discloses a digestive tract malignant tumor metastasis risk assessment method and system based on artificial intelligence, and relates to the field of digestive tract malignant tumor metastasis risk assessment. The method comprises the following steps: firstly, acquiring PET-CT three-dimensional image data of a colorectal cancer patient, and constructing and extracting a three-dimensional voxel matrix of a liver region according to the PET-CT three-dimensional image data; inputting the voxel matrix of the liver region into a pre-trained convolutional neural network model, and outputting a probability distribution matrix of each voxel; setting a basic confidence threshold of the malignant tumor metastasis, and extracting candidate focuses according to the probability distribution matrix and the confidence threshold of the malignant tumor metastasis; obtaining the maximum diameter of the candidate focus, dynamically and adaptively adjusting the confidence coefficient threshold value of the malignant tumor metastasis focus according to the maximum diameter, and generating a binary mask matrix; calculating a comprehensive risk score of colorectal cancer liver metastasis according to the binary mask matrix and grading; according to the invention, the false positive inhibition capability can be enhanced and the risk assessment accuracy can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of risk assessment for metastasis of digestive tract malignancies, involves artificial intelligence technology, and specifically is a method and system for risk assessment of metastasis of digestive tract malignancies based on artificial intelligence. Background Art

[0002] Colorectal cancer is a malignant tumor with high incidence and mortality rates globally. More than 50% of its fatal cases are closely related to liver metastasis. Early and accurate assessment of the risk of liver metastasis is of great significance for formulating personalized treatment plans and improving the prognosis of patients.

[0003] Currently, clinical diagnosis mainly relies on PET-CT images, but the traditional method has the following limitations: Subjectivity and complexity: It depends on doctors to manually delineate the liver region and lesions, with inter-observer differences and time-consuming. The false positive rate (such as benign hyperplasia) and the missed diagnosis rate of micro-metastasis are relatively high; Insufficient multi-modal data fusion: Fixed HU threshold segmentation is vulnerable to anatomical variations and noise interference, and single-modal analysis is difficult to capture metabolic-anatomical heterogeneity features; Single risk assessment parameter: Traditional indicators (such as the maximum SUV value, lesion volume) only reflect local features and lack quantification of global features such as metabolic heterogeneity and spatial distribution; Lack of model specificity: Existing deep learning models have a relatively high misjudgment rate for false positive regions (such as vascular shadows), and do not model the dynamic association between lesion size and diagnostic confidence; The present invention proposes a solution to the above problems. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides a method and system for risk assessment of metastasis of digestive tract malignancies based on artificial intelligence.

[0005] To achieve the above object, the technical solution of the present invention is as follows: In the first aspect, the present invention discloses a method for risk assessment of metastasis of digestive tract malignancies based on artificial intelligence, including the following steps: Obtain the PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix according to the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is a voxel, and the voxel includes position coordinates, HU value of the CT channel, and 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; Input the voxel matrix of the liver region into a pre-trained 3D convolutional neural network model to output a probability distribution matrix for each voxel, where the probability distribution matrix includes the probability Pm that the voxel belongs to a metastatic focus of malignant tumor and the probability Pf that the voxel belongs to a false positive region; Set a basic confidence threshold P0 for the metastatic focus of malignant tumor, and extract candidate lesions from the three-dimensional voxel matrix of the liver region according to the probability distribution matrix and the confidence threshold of the metastatic focus of malignant tumor; Obtain the maximum diameter of the candidate lesion, dynamically and adaptively adjust the confidence threshold P of the metastatic focus of malignant tumor 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.

[0006] In a second aspect, the present invention discloses an artificial intelligence-based risk assessment system for metastatic foci of digestive tract malignant tumors, including: A data acquisition module: used to acquire the PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix according to 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; A data extraction module: used to set an 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; A data analysis module: used to input the voxel matrix of the liver region into a pre-trained 3D convolutional neural network model to output a probability distribution matrix for each voxel, where the probability distribution matrix includes the probability Pm that the voxel belongs to a metastatic focus of malignant tumor and the probability Pf that the voxel belongs to a false positive region; A data judgment module: used to set a basic confidence threshold P0 for the metastatic focus of malignant tumor, and extract candidate lesions from the three-dimensional voxel matrix of the liver region according to the probability distribution matrix and the confidence threshold of the metastatic focus of malignant tumor; A risk assessment module: used to obtain the maximum diameter of the candidate lesion, dynamically and adaptively adjust the confidence threshold P of the metastatic focus of malignant tumor 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.

[0007] The present invention has the following beneficial effects: 1. By dynamically adjusting the HU threshold and combining the edge recognition of the U-Net network, the individual anatomical differences and image noise interference are eliminated, and the segmentation error between the liver region and the non-liver region is significantly reduced; 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 foci from interfering regions such as vascular artifacts, and the false positive suppression ability is enhanced; 3. By dynamically adjusting the confidence threshold according to the maximum diameter of the candidate lesion, the sensitivity and specificity are balanced. For example, the threshold is relaxed for large lesions to avoid false screening, and tightened for small lesions to reduce missed diagnoses, improving the detection rate of micrometastases; 4. Through a linear weighted scoring system that fuses the metabolic heterogeneity index (H), volume (V), and SUVmax, and optimizing the weights through logistic regression, the accuracy of risk grading is improved, providing a more comprehensive quantitative basis for clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0009] Figure 1 It is the overall method block diagram of Embodiment 1 of the present invention; Figure 2 It is the method flow chart of Embodiment 1 of the present invention; Figure 3 It is the flow chart for setting an adjustable HU threshold in Embodiment 1 of the present invention; Figure 4 It is the overall system block diagram of Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0010] The following will clearly and completely describe the technical solutions of the present invention in combination with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0011] Application Overview: Colorectal cancer is a malignant tumor with high incidence and mortality rates globally, and more than 50% of its fatal cases are closely related to liver metastasis. Early and accurate assessment of the risk of liver metastasis is of great significance for formulating personalized treatment plans and improving the prognosis of patients. Currently, clinical practice mainly relies on imaging examinations such as PET-CT for the detection and risk assessment of metastatic foci. PET-CT realizes lesion localization and qualitative analysis by fusing functional metabolic information and anatomical structure information, but still faces many challenges in practical applications: Subjectivity and complexity of image interpretation: Traditional methods rely on doctors to manually delineate the liver region and identify metastatic foci, which is time-consuming and has inter-observer differences. False positive problems such as benign hyperplasia and inflammatory regions have a relatively high missed diagnosis rate of small metastatic foci, affecting the diagnostic accuracy. Limitations of multi-modal data fusion: Existing algorithms mostly use fixed thresholds such as CT threshold method to segment the liver region, but anatomical variations of patients, image noise, and partial volume effects are likely to cause segmentation errors. In addition, single-modal analysis of only CT or PET is difficult to fully characterize the metabolic-anatomical heterogeneity of lesions. Simplification of risk assessment parameters: Traditional indicators such as the maximum SUV value and lesion volume only reflect local features and lack quantification of global features such as metabolic heterogeneity and spatial distribution, making it difficult to comprehensively evaluate the metastatic risk. In recent years, deep learning technology has provided new ideas for medical image analysis. 3D convolutional neural networks can extract multi-scale features from three-dimensional images through end-to-end learning and have been successfully applied to tumor detection and segmentation. However, there are still the following bottlenecks for liver metastasis risk assessment: (1) The imaging features of false positive regions and metastatic foci highly overlap, resulting in insufficient model specificity; (2) The dynamic association between lesion size and diagnostic confidence has not been effectively modeled; (3) The automated construction of multi-parameter risk scoring systems is not yet mature.

[0012] To solve the above problems, the researchers noticed that the HU value distribution of liver tissue has population differences, and fixed thresholds will lead to over-inspection or under-inspection in segmentation. Regarding false positive interference, it was found that there are subtle differences in the PET-CT dual-modal features between metastatic foci and benign regions, and multi-dimensional features need to be captured through a deep network. It was further found that there is a non-linear relationship between lesion size and diagnostic confidence, and smaller lesions require stricter thresholds to avoid misjudgment. Thus, a technical route of dynamic threshold segmentation, dual-modal feature fusion, and confidence adaptive adjustment was formed.

[0013] Therefore, the present application proposes the following technical solutions.

[0014] Example 1

[0015] As Figures 1-3 shown, an artificial intelligence-based method for assessing the metastatic risk of digestive tract malignancies includes the following steps: Obtain the PET-CT three-dimensional image data of colorectal cancer patients to construct a three-dimensional voxel matrix, and the voxels of this matrix include position coordinates, HU values, and SUV values. Set a dynamically adjustable HU threshold to extract the voxel matrix of the liver region, input it into a pre-trained 3D convolutional neural network, and output the probabilities of each voxel belonging to metastatic foci and false positive regions. After extracting candidate lesions based on the basic confidence threshold, dynamically adjust the confidence threshold according to the maximum diameter of the lesions, generate a binary mask matrix, calculate the comprehensive risk score, and grade it.

[0016] Among them, the construction of the three-dimensional voxel matrix integrates PET-CT dual-modal data through spatial alignment and interlayer interpolation. Each voxel contains anatomical and metabolic information, providing a multi-dimensional data basis for subsequent analysis. The dynamically adjusted HU threshold optimizes the segmentation range using the three-dimensional Otsu algorithm, and automatically adapts to individual differences by statistically analyzing the HU distribution of the entire image. The 3D convolutional neural network adopts an adversarial training strategy, and distinguishes true and false positive features through a generative adversarial network to improve the model's specificity. The confidence threshold dynamic adjustment mechanism automatically corrects the diagnostic criteria according to the lesion diameter. The threshold is relaxed for large lesions to avoid excessive exclusion, and tightened for small lesions to reduce misjudgment.

[0017] Specifically, the spatial alignment of the three-dimensional voxel matrix eliminates the registration error between modalities, ensuring that the corresponding HU and SUV values are included at the same voxel position. During dynamic threshold segmentation, the three-dimensional Otsu algorithm adds a spatial continuity constraint on the basis of the traditional two-dimensional algorithm to avoid generating isolated noise regions. The 3D convolutional neural network captures local details and global context simultaneously through a multi-scale feature fusion module, and the skip connection structure retains edge information to improve the detection ability of small lesions. The confidence threshold adjustment formula introduces a decay coefficient to establish an inverse relationship between the diameter and the threshold. When the diameter increases, the threshold decreases appropriately but remains higher than the baseline value, balancing sensitivity and specificity.

[0018] Compared with the prior art, the traditional HU threshold method uses a fixed range for segmentation, resulting in a large volume error. The three-dimensional Otsu algorithm in this solution can reduce the segmentation error. The traditional risk assessment only uses a single parameter of SUVmax, while this solution comprehensively considers multiple parameters such as metabolic heterogeneity and volume, improving the accuracy of risk grading.

[0019] Through the above technical solutions, this application realizes the individual adaptation of liver region segmentation, effectively eliminating the influence of image noise and partial volume effect. The dual-modal feature fusion and adversarial training strategy significantly improve the model's ability to distinguish between micro metastatic foci and false positive regions. The confidence dynamic adjustment mechanism optimizes the diagnostic criteria according to the lesion size, reducing the false positive rate while maintaining high sensitivity. The multi-parameter scoring system comprehensively evaluates the metabolic activity and spatial distribution characteristics, providing a more accurate basis for clinical risk assessment.

[0020] This application further proposes a technical solution for constructing a three-dimensional voxel matrix by spatially aligning functional metabolic imaging data and anatomical structure imaging data and generating a three-dimensional voxel matrix through interlayer interpolation.

[0021] Among them, functional metabolic imaging data refers to the standardized uptake value distribution data obtained through PET scans, which can be specifically collected using positron emission tomography equipment and is used to characterize the strength of tissue metabolic activity. Anatomical structure imaging data refers to the Hounsfield unit value distribution data obtained through CT scans, which can be specifically collected using X-ray computed tomography equipment and is used to characterize tissue density differences. Spatial alignment refers to the coordinate system normalization of PET and CT images of the same patient, which can be specifically achieved through a rigid registration algorithm and is used to eliminate spatial offsets caused by respiratory movement or equipment acquisition differences. Interpolation between slices refers to the interpolation calculation of raw data with different scan slice thicknesses to generate continuous three-dimensional voxel data, which can be specifically achieved using a cubic spline interpolation algorithm and is used to solve the stepped artifact problem caused by inconsistent slice spacing.

[0022] Specifically, PET images provide metabolic hot spot information, while CT images provide anatomical structure information. Through the rigid registration algorithm, the voxel coordinates of the PET image are mapped to the CT image coordinate system to achieve pixel-level spatial matching of the bimodal data. For slices with inconsistent slice thicknesses in the raw data, a cubic spline interpolation algorithm is used to perform interpolation calculations along the axial direction to generate a continuous three-dimensional voxel matrix with a unified spatial resolution, and each voxel contains both HU values and SUV values. The resulting three-dimensional data structure can accurately reflect the corresponding relationship between metabolic activity and anatomical location, providing spatially consistent multimodal features for subsequent dynamic threshold segmentation.

[0023] Compared with the prior art, traditional methods usually process PET and CT data separately or use fixed thresholds for unimodal segmentation, resulting in spatial misalignment between metabolic hot spots and anatomical structures and prone to segmentation errors. In this application, image offsets are eliminated through bimodal spatial alignment, and the continuity of the three-dimensional data is ensured by combining interpolation between slices, enabling subsequent analysis to accurately correlate metabolic abnormal regions with anatomical locations and avoiding false positives or missed detections caused by inconsistent data.

[0024] Through the above technical solutions, this application effectively solves the difficulty of feature fusion caused by spatial mismatch of multimodal images and improves the accuracy of liver region segmentation. The continuity and multimodal attributes of the three-dimensional voxel matrix provide a reliable data basis for subsequent dynamic threshold adjustment and lesion recognition, reducing the risk of misjudgment caused by data artifacts or spatial offsets.

[0025] This application further proposes a process for setting a dynamically adjustable HU threshold, including performing Gaussian filtering denoising on the CT channel 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, classifying voxels into foreground and background according to the comparison results of HU values and T values, and calculating: σ 2(T) = ω1(T)ω2(T)[μ1(T) - μ2(T)] 2 , where ω1 and ω2 are the voxel occupancy ratios of the foreground and background respectively, and μ1 and μ2 are the HU means of the foreground and background; Select the T value that maximizes σ 2 (T) as the initial value Tinitial; set the initial HU threshold interval [Tlow, Thigh], where Tlow is Tinitial - kσH and Thigh is Tinitial + kσH; calculate the volume Vcurrent of the liver candidate region and determine whether it exceeds the anatomical range [Vmin, Vmax], and dynamically adjust the k value to make the volume Vcurrent fall within a reasonable range.

[0026] Among them, Gaussian filtering denoising refers to convolving the CT channel data with a two-dimensional Gaussian kernel. For example, a Gaussian kernel with a standard deviation of 1.5 is used to eliminate the random noise generated during image acquisition. The foreground and background can be divided according to the comparison result of the T value and the HU value. When the T value is less than the HU value, it is divided into the foreground, and vice versa. The volume Vcurrent of the candidate region can be obtained by multiplying the number of voxels in the liver candidate region by the unit volume size corresponding to the voxels. Dynamically adjusting the k value means performing 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 the abnormal segmentation caused by the partial volume effect.

[0027] Specifically, after the CT channel data is preprocessed by Gaussian filtering, the HU value map of each voxel is statistically analyzed and the between-class variance corresponding to different HU thresholds is calculated, and the initial threshold Tinitial that maximizes the variance is selected. Centered on this initial threshold Tinitial, the dynamic threshold interval is determined in combination with the standard deviation σH of the full-image HU value and the empirical coefficient k. Calculate the volume Vcurrent of the liver candidate region at the current threshold. When it is detected that the volume Vcurrent exceeds the reasonable anatomical range of the liver tissue, the threshold interval 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 looped until the volume Vcurrent stabilizes within the preset reasonable anatomical range, such as the liver volume interval of 800 - 1200 cm³.

[0028] Compared with the prior art, the traditional fixed threshold method directly uses the fixed HU range recommended in the literature for segmentation. For example, a threshold of 40 - 90HU is uniformly adopted, which cannot adapt to the natural variation of liver density among individuals. In this solution, the threshold interval is dynamically set by combining the standard deviation. For example, for patients with liver cirrhosis, the upper threshold can be automatically extended to 110HU to cover fibrotic tissues. The prior art uses a two-dimensional Otsu algorithm that only considers the information of a single-layer slice. This solution calculates the between-class variance in three-dimensional space and can effectively distinguish adjacent organs with similar density to the liver.

[0029] Through the above technical solution, this application solves the problem of mis-segmentation of the liver region caused by individual differences and image noise in traditional segmentation methods. By analyzing in three-dimensional space, it can accurately distinguish the liver from surrounding tissues. The dynamic threshold adjustment mechanism can adapt to the anatomical structure variations of different patients and avoid abnormal candidate regions caused by partial volume effects, providing an accurate data basis for the liver region for subsequent metastasis detection.

[0030] This application further proposes a method for segmenting the liver region based on HU threshold screening and U-Net network edge recognition, which specifically includes: screening liver candidate regions according to the HU threshold, using the U-Net network to recognize the liver edge, accurately segmenting the liver region, and outputting a three-dimensional voxel matrix of the liver region.

[0031] Among them, HU threshold screening refers to preliminarily dividing liver candidate regions using the Hounsfield unit value range of the CT channel, which can be specifically implemented by using a dynamically adjusted threshold interval. For example, the initial threshold range is determined through three-dimensional histogram analysis and the between-class variance maximization algorithm, and the threshold boundary is dynamically corrected according to the volume of the liver candidate region. This step can quickly exclude regions with HU values significantly deviating from liver tissue, such as bones and fat.

[0032] Among them, U-Net network edge recognition refers to performing pixel-level segmentation using a convolutional neural network with an encoder-decoder structure. Specifically, the high-resolution features of the shallow layer and the semantic features of the deep layer can be fused through a skip connection mechanism. For example, multi-scale features are extracted through 4-level downsampling in the encoding stage, and the spatial details are restored by combining transposed convolution and skip connections in the decoding stage. When training the network, the Dice loss function can be used to optimize the continuity of the segmentation boundary, and pre-trained weights can be loaded to accelerate convergence.

[0033] Specifically, this method first generates candidate regions through HU thresholding, which reduces the scale of voxels to be processed and computational redundancy. Subsequently, the liver candidate regions are input into the 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 operations. In the decoding stage, the upsampling layer and skip connections work together to fuse low-level edge information and high-level semantic features, effectively identifying blurred boundaries caused by partial volume effects. The output layer uses the Sigmoid activation function to generate a probability map, which is thresholded into a binary mask, and finally outputs an accurate three-dimensional voxel matrix of the liver.

[0034] Compared with the prior art, traditional methods rely on fixed HU thresholds for liver segmentation and cannot adapt to density differences between individuals and image noise interference. For example, the liver density of patients with liver cirrhosis may be lower than the standard threshold range, resulting in missing segmentation regions. In contrast, this solution combines dynamic threshold screening with deep learning edge detection, which can suppress the influence of noise while maintaining a high recall rate. Experiments show that this method improves the Dice coefficient compared with the single-threshold method in the liver segmentation task, and the segmentation accuracy in low-contrast regions is significantly better than that of traditional morphological post-processing methods.

[0035] Through the above technical solutions, 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. Through the two-stage strategy of dynamic threshold initial positioning and deep learning fine correction, the influence of vascular artifacts and partial volume effects on the segmentation results is effectively suppressed, providing accurate anatomical constraint conditions for subsequent metastasis detection. While ensuring computational efficiency, this method controls the liver volume measurement error within the clinically acceptable range, providing a reliable spatial reference for the automated risk assessment system.

[0036] This application further proposes the process of outputting a two-channel probability distribution matrix through a three-dimensional convolutional neural network model, including obtaining positron emission tomography images of colorectal cancer patients and pathologically confirmed metastasis annotation data, using an adversarial training strategy to input the computed tomography images of false positive regions as negative samples into the discriminator to optimize the specificity of the feature extractor, extracting high-level semantic feature context fusion through multi-level convolution and pooling, using skip connections to combine low-level details and high-level semantic information, and outputting the probability that each voxel belongs to a malignant tumor metastasis and the probability that the voxel belongs to a false positive region through a softmax layer.

[0037] Among them, the adversarial training strategy refers to the game training between the discriminator and the feature extractor in the generative adversarial network framework. Specifically, it can be realized by alternately optimizing the parameters of the generator and the discriminator. Through the discriminator's discrimination feedback on false positive samples, the feature extractor is driven to generate more discriminative feature expressions. Multilevel convolution and pooling refer to the feature extraction architecture formed by alternately stacking multiple convolutional layers and pooling layers. Specifically, it can be realized by using a five-layer convolutional structure combined with the max pooling operation. Three-dimensional semantic features with spatial correlation are extracted from the original image layer by layer through abstraction. Skip connection refers to the operation of concatenating the feature maps of the shallow network and the deep network in channels. Specifically, it can be realized by upsampling using three-dimensional transposed convolution and then fusing with the corresponding hierarchical features. By retaining the detail information at different scales, the model's ability to recognize tiny lesions is enhanced. The sigmoid function layer refers to the non-linear function that converts the output of the neural network into a probability distribution. Specifically, it can be realized by using two-channel probability mapping and normalization calculation. The confidence of each voxel belonging to different categories is quantified through probability values.

[0038] Specifically, during the training process, the pathological annotation data serves as the supervision signal to ensure that the model learns the true metastatic focus feature distribution. The discriminator in the adversarial training framework is input with false positive region samples, forcing the feature extractor to suppress the interference patterns related to the false positive regions when generating features. Multilevel convolution extracts local texture and global morphological features layer by layer through three-dimensional convolutional kernels. The pooling operation reduces the resolution of the feature map while enhancing the robustness to deformation. Skip connection fuses the low-level high-resolution features and high-level semantic features in the decoding stage to achieve complementary optimization of multi-scale information. The sigmoid function layer normalizes the probability of the two-channel output of each voxel to obtain the probability distribution matrix of malignant tumor metastases and false positive regions, providing a reliable basis for subsequent threshold screening.

[0039] Compared with the prior art, traditional methods usually use single-modal data to train classification models, lacking targeted optimization for false positive samples, resulting in insufficient specificity. Most existing three-dimensional convolutional networks adopt a one-way encoding structure, losing low-level detail information during the feature fusion process and affecting the detection rate of tiny lesions. This solution explicitly models the feature distribution of false positive regions through the adversarial training mechanism and combines the multi-scale fusion strategy of skip connection to effectively improve the model's ability to analyze complex image patterns.

[0040] Through the above technical solutions, this application can significantly reduce the misjudgment rate of false positive regions, enhance the model's deep feature fusion ability for multi-modal image data, achieve accurate probability prediction of malignant tumor metastases, and provide a reliable quantitative basis for subsequent dynamic threshold screening.

[0041] The present application further proposes a technical solution for setting a basic confidence threshold and performing probability comparison when extracting candidate lesions. Specifically, it includes: setting a basic confidence threshold P0; comparing the probability of malignant tumor metastasis Pm of each voxel in the three-dimensional voxel matrix with P0, and when the Pm of the voxel reaches or exceeds P0, the voxel is marked as a candidate lesion.

[0042] Among them, the basic confidence threshold P0 refers to the probability critical value used to distinguish malignant tumor metastases from false positive regions, and can specifically be determined by the median of the metastasis probability distribution of true positive cases in clinically verified data, such as 0.7. This threshold is used as a preliminary screening criterion to filter out low-probability interfering voxels. The probability comparison mechanism refers to the operation of making a point-by-point judgment on the predicted probability of each voxel, and can specifically be implemented through a matrix element comparison algorithm for parallel computing. This mechanism can synchronously process millions of voxel data to ensure the operation efficiency.

[0043] Specifically, in the probability distribution matrix output by the pre-trained 3D convolutional neural network, each voxel has a quantified probability value representing its belonging to a malignant tumor metastasis. The setting of the basic confidence threshold P0 is based on the metastasis probability distribution characteristics in a large-scale clinical dataset. For example, the threshold is set to the 85th percentile of the probability distribution. When traversing all voxels, only the voxels with probability values higher than this threshold are retained to form the spatial distribution of candidate lesions. This process excludes the voxels with low probability values in the false positive region by establishing a mapping relationship between probability and anatomical location, and at the same time retains the regions with typical metastasis characteristics.

[0044] Compared with the prior art, traditional methods mostly use fixed thresholds or single-modal data for lesion extraction, such as segmentation based on a fixed interval of CT values only. Such methods cannot effectively distinguish false positive regions caused by inflammation or vascular abnormalities. This solution realizes precise filtering of false positive regions by fusing the probability prediction results of multi-modal image features and combining a dynamically optimized confidence threshold. In addition, the probability comparison mechanism is based on voxel-level analysis in the full three-dimensional space, which significantly improves the spatial continuity of detection compared with methods based on two-dimensional slices or manual delineation.

[0045] Through the above technical solution, the present application effectively reduces the false positive misjudgment rate caused by overlapping image features. The objective screening mechanism based on probability distribution avoids the subjective deviation of manual threshold setting, ensuring the repeatability and accuracy of the candidate lesion extraction process. By retaining the set of voxels with high probability characteristics, it provides a reliable data basis for subsequent lesion diameter calculation and adaptive threshold adjustment, thereby improving the overall system specificity while ensuring sensitivity.

[0046] The present application further proposes that the calculation process of the maximum diameter of the candidate lesion includes calculating the maximum diameter according to the position coordinates of the voxels corresponding to the candidate lesion through connected component analysis and convex hull algorithm.

[0047] Among them, connected component analysis refers to the technology of clustering adjacent candidate lesion voxels in three-dimensional space into independent connected regions, which can be specifically implemented by region growing algorithm or labeling algorithm based on neighborhood search, and is used to eliminate the interference of discrete noise points and distinguish different lesion individuals. The convex hull algorithm refers to the technology of calculating the minimum convex polyhedron of the three-dimensional point cloud to wrap the geometric structure of the candidate lesion, which can be specifically implemented by Graham scan method or quick convex hull algorithm, and is used to exclude the interference of the internal concave structure of the lesion on the calculation of the maximum diameter. Euclidean distance calculation refers to the geometric distance measurement method based on the spatial coordinates of voxels, which can be specifically implemented by traversing the distances between any two points in the convex hull vertex set and selecting the maximum value, and is used to accurately represent the true extension range of the lesion in three-dimensional space.

[0048] Specifically, the candidate lesion voxels may be distributed dispersedly or adhered to each other in three-dimensional space. Through connected component analysis, they can be divided into independent lesion entities to avoid miscounting the distances between different lesions into the same lesion. Further, the convex hull algorithm is applied to the voxel coordinate set of each independent lesion to construct a minimum convex polyhedron model that wraps all voxels, eliminating the measurement deviation caused by uneven internal density or morphological depression of the lesion. Finally, all possible pair-wise distances are calculated by traversing the convex hull vertex set, and the maximum Euclidean distance is selected as the true diameter of the lesion. This method overcomes the problem of dimension information loss caused by perspective limitation or slice thickness in traditional two-dimensional projection measurement.

[0049] Compared with the prior art, traditional methods usually measure the maximum diameter of two-dimensional slices or estimate the equivalent sphere diameter based on the number of voxels, and their measurement results are easily affected by the spatial orientation of the lesion and partial volume effect. This solution accurately represents the three-dimensional spatial topology of the lesion through three-dimensional connected component analysis and geometric convex hull modeling, especially suitable for metastatic lesions with lobulated or irregular shapes, and significantly improves the spatial consistency of diameter measurement.

[0050] Through the above technical solutions, the present application realizes the accurate modeling of the true three-dimensional shape of the candidate lesion, effectively solves the problem of error accumulation caused by dimension loss or morphological simplification in traditional measurement methods. By excluding the interference of internal structures through geometric convex hull, 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.

[0051] The present application further proposes a process for dynamically and adaptively adjusting the confidence threshold of metastatic foci of malignant tumors according to the maximum diameter of candidate foci, including calculating the confidence threshold P according to the maximum diameter Dmax of the candidate foci, and its expression is P = P0×(1 - Dmax / (Dmax + s)), where s is the attenuation coefficient, and the smaller s is when the maximum diameter Dmax is larger; only the voxels with the probability Pm ≥ the confidence threshold P are retained to generate a binary mask matrix for marking metastatic foci and the background.

[0052] Among them, the maximum diameter refers to the longest distance in the three-dimensional space of the candidate focus obtained by connected component analysis and convex hull algorithm. Specifically, the Euclidean distance formula can be used to calculate the voxel coordinates, and this parameter reflects the spatial distribution characteristics of the focus and is used to establish an association model between size and diagnostic reliability.

[0053] Further, the calculation formula for the maximum diameter is: , where (xi, yi, zi) and (xj, yj, zj) are the position coordinates of the voxels corresponding to the candidate focus; Among them, the attenuation coefficient s is a parameter used to adjust the dynamic change range of the confidence threshold. Specifically, the empirical value range can be determined according to clinical data statistics. For example, the value of s can be adjusted in the range of 5mm to 15mm. Its function is to balance the demand for suppressing false positives of large-sized foci and the demand for ensuring the sensitivity of small-sized foci. Among them, the dynamic confidence threshold P is a non-linear function constructed based on the basic threshold P0 and the focus size. Specifically, the focus diameter is mapped to the threshold adjustment coefficient through mathematical transformation, so that large foci need to meet higher confidence requirements, while small foci can accept relatively loose screening conditions.

[0054] Specifically, the maximum diameter of the candidate focus is input to the threshold adjustment function after three-dimensional space geometry calculation, and the attenuation coefficient s is adjusted inversely according to the focus size. For example, when it is detected that Dmax exceeds 20mm, s can be set to 5mm to enhance the threshold attenuation amplitude. At this time, the confidence threshold P rises rapidly with the increase of the diameter, prompting the system to only retain high-confidence voxels; when Dmax is less than 10mm, s can be set to 15mm to slow down the threshold rising slope and avoid the loss of tiny foci due to excessive filtering. Through this dynamic mechanism, the generation process of the binary mask matrix can adaptively adjust the screening strictness according to the focus size, while suppressing large-sized false positive regions such as vascular artifacts, and retaining tiny metastatic foci with low probability characteristics but conforming to the anatomical distribution law.

[0055] Compared with the prior art, the traditional method uses a fixed confidence threshold to screen voxels, which cannot distinguish the differences in the diagnostic reliability of lesions of different sizes, resulting in holes in large lesions due to low probability values of some voxels, or small lesions being misfiltered due to too high a threshold. This solution establishes a dynamic association model between the lesion diameter and the confidence threshold, making the voxel screening criteria have spatial self - adaptability, reducing false - positive interference while ensuring the integrity of large - lesion segmentation, and maintaining the detection sensitivity for small - size lesions at the same time.

[0056] Through the above - mentioned technical solution, this application effectively solves the problems of high false - positive rate and missed diagnosis of micro - metastatic lesions caused by the lack of dynamic association between the lesion size and the diagnostic confidence in the traditional method. Through the size - adaptive threshold adjustment mechanism, the system can differentially process the voxel screening process of lesions of different sizes. For example, for lesions with a diameter exceeding 15 mm, the screening strictness is increased to suppress the interference of vascular artifacts, while for lesions with a diameter less than 8 mm, the conditions are appropriately relaxed to avoid missing early metastasis signals, thereby achieving an optimized balance between sensitivity and specificity and improving the accuracy of liver metastasis risk assessment.

[0057] The process of calculating the comprehensive risk score and rating according to the binary mask matrix in this application includes: obtaining the maximum SUV value SUVmax of the SUVs in all voxels marked as 1 in the binary mask matrix; calculating the volume V of the metastatic lesion and the metabolic heterogeneity index H according to all voxels marked as 1 in the binary mask matrix; the volume V of the metastatic lesion is calculated according to the number of voxels marked as 1; the calculation formula of the metabolic heterogeneity index H is: H = Σ|SUV_i - SUV_avg|² / N; where SUV_i is the SUV value of the edge voxel, 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 the linear weighted sum of the volume V, the metabolic heterogeneity index H, and SUVmax, and the specific calculation formula is: R = α * SUVmax+β * V+γ * H, where α, β, and γ are weight coefficients obtained through logistic regression training; matching the comprehensive risk score with the preset grading standard and outputting low - risk, medium - risk, or high - risk signals for colorectal cancer liver metastasis.

[0058] Among them, the metabolic heterogeneity index H is a quantitative index calculated by normalizing the sum of the squared differences between the SUV values of the edge voxels and the non - edge voxels. Specifically, it can be achieved by extracting the edge voxels through an edge - detection algorithm for the lesion area in the binary mask and calculating the discrete degree of the difference in metabolic activity between the edge voxels and the internal voxels. This index is used to characterize the unevenness of the metabolic distribution inside the tumor and reflects the spatial heterogeneity characteristics of the tumor biological behavior.

[0059] The volume V refers to the three-dimensional spatial occupancy parameter of the lesion calculated by multiplying the number of voxels labeled as metastases in the binary mask by the volume of a single voxel, which can be specifically achieved by combining three-dimensional connected component analysis with voxel spatial resolution conversion. This parameter is used to evaluate the anatomical invasion range of metastases.

[0060] SUVmax refers to the maximum standardized uptake value extracted from the labeled area of the binary mask, which can be specifically achieved by traversing the PET channel data of all labeled voxels and recording the peak value. This parameter is used to characterize the local highest metabolic activity level of the lesion.

[0061] The linear weighted sum R refers to combining three independent parameters, namely volume, metabolic heterogeneity, and SUVmax, into a single scoring index according to weight coefficients, which can be specifically achieved by using a logistic regression model to train the contribution coefficients of each parameter based on clinical follow-up data. This scoring mechanism optimizes the multi-feature fusion strategy in a data-driven manner.

[0062] Specifically, first, a set of voxels labeled as metastases in the binary mask is extracted through a three-dimensional image processing algorithm, and the occupied spatial volume is calculated as a quantitative index of the anatomical invasion degree. Further, edge detection technology is used to distinguish the boundary and core regions of the lesion, and a mathematical model reflecting the internal heterogeneity of the tumor is constructed by comparing the metabolic activity differences between edge voxels and non-edge voxels. At the same time, the maximum metabolic activity value within the lesion area is recorded to capture the characteristics of local malignancy. Finally, the parameters in the three dimensions are integrated into a comprehensive score through weight coefficients optimized by machine learning, and the risk level is output according to the preset clinical grading threshold.

[0063] Compared with the prior art, traditional risk assessment methods mostly use a single parameter such as SUVmax or the diameter of the lesion for grading, and fail to integrate indicators reflecting the biological complexity of tumors such as metabolic heterogeneity. This solution designs a metabolic heterogeneity index, for the first time incorporates spatial distribution heterogeneity into the scoring system, and combines an adaptive weight allocation mechanism, enabling the evaluation model to simultaneously capture the metabolic activity, anatomical invasion range, and internal heterogeneity characteristics of the lesion. For example, when calculating metabolic heterogeneity in the prior art, the global standard deviation is usually used, while in this solution, the spatial heterogeneity pattern is more accurately characterized by calculating in the edge and non-edge partitions.

[0064] Through the above technical solutions, this application solves the problem of insufficient grading accuracy caused by parameter simplification in traditional risk assessment, and realizes the fusion analysis of three-dimensional features of metabolic activity, anatomical invasion and spatial heterogeneity. Through the calculation of edge-specific metabolic differences, the biological behavior differences between diffuse and focal metastases can be effectively distinguished; through the dynamically weighted allocation optimized by machine learning, the subjective bias of manually setting weights is avoided, making the scoring system more in line with the statistical laws of clinical prognostic data. The finally output comprehensive risk score can provide multi-dimensional quantitative basis covering morphology, functional metabolism and spatial heterogeneity for clinicians, and support personalized treatment decisions.

[0065] Example Two

[0066] As Figure 4 shown, an artificial intelligence-based risk assessment system for metastatic malignancies of the digestive tract includes: Data acquisition module: used to acquire the PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix according to the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is a voxel, and the voxel includes position coordinates, HU values of the CT channel, and SUV values of the 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, where the probability distribution matrix includes the probability Pm that the voxel belongs to a metastatic malignant tumor focus and the probability Pf that the voxel belongs to a false positive region; Data judgment module: used to set the basic confidence threshold P0 of the metastatic malignant tumor focus, and extract candidate lesions from the three-dimensional voxel matrix of the liver region according to the probability distribution matrix and the confidence threshold of the metastatic malignant tumor focus; Risk assessment module: used to obtain the maximum diameter of the candidate lesion, dynamically and adaptively adjust the confidence threshold P of the metastatic malignant tumor focus according to the maximum diameter, and generate a binary mask matrix; calculate the comprehensive risk score and grading of colorectal cancer liver metastasis according to the binary mask matrix.

[0067] The above content is only an example and illustration of the structure of the present invention. Those skilled in the art of this technology can make various modifications or supplements to the described specific embodiments 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 by this claim book, they should all belong to the protection scope of the present invention.

[0068] In the description of this specification, the descriptions referring to the terms "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0069] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the present invention to only the specific embodiments. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. An artificial intelligence-based method for assessing the risk of metastasis of malignant tumors in the digestive tract, characterized in that, The method includes the following steps: Obtain the PET-CT three-dimensional image data of a colorectal cancer patient, 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, a HU value of the CT channel, and an 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; Input the three-dimensional voxel matrix of the liver region into a pre-trained 3D convolutional neural network model, and output a probability distribution matrix for each voxel, where the probability distribution matrix includes the probability Pm that the voxel belongs to a metastatic malignant tumor focus and the probability Pf that the voxel belongs to a false positive region; Set a basic confidence threshold P0 for metastatic malignant tumor foci, and extract candidate foci from the three-dimensional voxel matrix of the liver region according to the probability distribution matrix and the confidence threshold P0 for metastatic malignant tumor foci; Obtain the maximum diameter of the candidate foci, dynamically adjust the confidence threshold P that the voxel belongs to a metastatic malignant tumor focus according to the maximum diameter, and generate a binary mask matrix; calculate a comprehensive risk score and grade for colorectal cancer liver metastasis according to the binary mask matrix.

2. The method for evaluating the metastasis risk of digestive tract malignant tumors based on artificial intelligence according to claim 1, wherein, 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: spatially aligning and performing 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 evaluating the metastasis risk of digestive tract malignant tumors based on artificial intelligence according to claim 1, wherein 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 occupancy ratios of the foreground and background respectively, and μ1 and μ2 are the mean voxel HU values of the foreground and background respectively; Select to make σ 2 (T) the maximum T value as the initial value Tinitial; Set an 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 narrow the threshold range; Repeat adjusting the value of k until the volume Vcurrent ∈ [Vmin, Vmax].

4. The method for evaluating the metastasis risk of digestive tract 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.

5. The method for evaluating the metastasis risk of digestive tract malignant tumors based on artificial intelligence according to claim 1, wherein The process of outputting a two-channel probability distribution matrix through the 3D convolutional neural network model includes: Obtain the PET-CT images of colorectal cancer patients and the labeled data of pathologically diagnosed malignant tumor metastases; Adopt an adversarial training strategy, use the CT images of false positive regions in the labeled data of malignant tumor metastases as negative samples to input into the discriminator, and optimize the specificity of the feature extractor; Extract high-level semantic feature context fusion through multi-level convolution and pooling, use skip connections to combine low-level details with high-level semantic information, and the Softmax layer outputs the probability Pm that each voxel belongs to a malignant tumor metastasis and the probability Pf that the voxel belongs to a false positive region.

6. The method for evaluating the metastasis risk of digestive tract malignant tumors based on artificial intelligence according to claim 1, wherein The process of extracting candidate lesions according to the confidence threshold of malignant tumor metastases in the probability distribution matrix includes: Set the basic confidence threshold P0; Compare the probability Pm of all voxels with the basic confidence threshold P0 respectively: if the probability Pm ≥ the basic confidence threshold P0, then mark this voxel as a candidate lesion.

7. A method for evaluating the risk of metastasis of digestive tract malignant tumors based on artificial intelligence according to claim 1, characterized in that, The calculation process of the maximum diameter of the candidate lesion includes: Calculate the maximum diameter Dmax of the candidate lesion through connected component analysis and convex hull algorithm according to the position coordinates (xi, yi, zi) and (xj, yj, zj) of the voxels corresponding to the candidate lesion; The calculation formula is as follows: .

8. The method for assessing the metastasis risk of digestive tract malignant tumors based on artificial intelligence according to claim 1, wherein The process of dynamically and adaptively adjusting the confidence threshold of malignant tumor metastases according to the maximum diameter includes: Calculate the confidence threshold P according to 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, the smaller s; Only retain the voxels with Pm ≥ the confidence threshold P, mark the voxels that meet the confidence threshold as 1, and the rest as 0, to generate a binary mask matrix.

9. The method for evaluating the risk of metastasis of malignant tumors of the digestive tract based on artificial intelligence according to claim 1 or 8, characterized in that, The process of calculating the comprehensive risk score and grading according to the binary mask matrix includes: Obtain the maximum SUV value SUVmax of the SUV in all voxels marked as 1 in the binary mask matrix; Calculate the volume V and metabolic heterogeneity index H according to all voxels marked as 1 in the binary mask matrix; The calculation formula of the metabolic heterogeneity index H is: H = Σ|SUV_i - SUV_avg|² / N; Where SUV_i is the SUV value of the edge voxel, SUV_avg is the average SUV value of non-edge voxels, and N is the number of edge voxels; The comprehensive risk score is the linear weighted sum of volume V, metabolic heterogeneity index H and SUVmax, and the specific calculation formula is: R = α*SUVmax + β*V + γ*H, where α, β, γ are weight coefficients obtained through logistic regression training; Match the comprehensive risk score with the preset grading standard, and output low-risk, medium-risk or high-risk signals of colorectal cancer liver metastases.

10. An artificial intelligence-based risk assessment system for metastasis of digestive tract malignant tumors, characterized in that: Use an artificial intelligence-based method for assessing the metastasis risk of digestive tract malignant tumors as described in any one of claims 1 to 9, including: Data acquisition module: used to obtain the PET-CT three-dimensional image data of colorectal cancer patients, and construct a three-dimensional voxel matrix according to the PET-CT three-dimensional image data; the smallest unit of the three-dimensional voxel matrix is a voxel, and the voxel includes position coordinates, HU value of the CT channel, and SUV value of the PET channel; Data extraction module: used to set a dynamically adjustable HU threshold, and extract a 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 a probability distribution matrix for each voxel, where the probability distribution matrix includes the probability Pm that the voxel belongs to a metastatic focus of malignant tumor and the probability Pf that the voxel belongs to a false positive region; Data judgment module: used to set a basic confidence threshold P0 for metastatic foci of malignant tumors, and extract candidate lesions from the three-dimensional voxel matrix of the liver region according to the probability distribution matrix and the confidence threshold of the metastatic foci of malignant tumors; Risk assessment module: used to obtain the maximum diameter of the candidate lesion, dynamically adjust the confidence threshold P that the voxel is a metastatic focus of malignant tumor according to the maximum diameter, and generate a binary mask matrix; calculate the comprehensive risk score and grade of colorectal liver metastasis according to the binary mask matrix.

Images