Intelligent liver image sign analysis and li-rads classification system based on multi-task model

Abstract

The application discloses an intelligent liver image sign analysis and LI-RADS classification system based on a multi-task model, which inputs two-dimensional liver tumor images of multiple phases into a multi-task convolutional neural network model, automatically extracts potential features required by a classification task based on an LI-RADS standard in a main task, and extracts potential features required by main sign classification defined by the LI-RADS standard in a subtask, so that higher LI-RADS grading accuracy can be realized, and classification basis can be provided for doctors in actual applications such as clinical diagnosis. The application uses a multi-task convolutional neural network, an image-based tumor size automatic analysis method, and an end-to-end and supervised contrast learning combined mode to train a model, so that the liver tumor LI-RADS classification system which can provide judgment basis for doctors is realized.

Description

Technical Field

[0001] This invention relates to the fields of computer application technology and artificial intelligence, and in particular to an intelligent liver imaging feature analysis and LI-RADS classification system based on a multi-task model. Background Technology

[0002] The liver is a vital organ in the human body, responsible for digestion and metabolism. Epidemiological data shows that the incidence of liver diseases has been rising worldwide in recent years, with liver tumors being a major cause of death. Hepatocellular carcinoma (HCC) is the most common primary malignant liver tumor, accounting for more than 80% of clinical diagnoses.

[0003] The Liver Imaging Reporting and Data System (LI-RADS) aims to provide standardized terminology and characteristic parameters for interpreting and reporting liver CT and MRI findings in patients at risk of developing hepatocellular carcinoma. The significance and goal of this standard is to reduce the variability in physician interpretation of imaging results and improve the fluency and clarity of communication among radiologists, attending physicians, and other relevant experts.

[0004] However, in actual diagnosis, if doctors use the LI-RADS criteria to assess the risk of hepatocellular carcinoma from liver tumor images, they need to observe CT or MRI images in the arterial, portal venous, and delayed phases. They must then analyze the characteristics of each sign according to the primary and secondary signs specified in the LI-RADS criteria, and finally determine the LI-RADS grade based on the comprehensive discrimination rules established by the LI-RADS criteria. This not only requires experienced doctors to spend a significant amount of time analyzing the images, but also sometimes leads to difficulties in determining the final LI-RADS grade due to disagreements among doctors regarding certain signs.

[0005] In recent years, thanks to the development of AI technology, AI-based assisted diagnostic technologies have been advancing rapidly and have gradually appeared in actual hospital diagnoses. However, current intelligent assisted diagnostic technologies, especially the automatic LI-RADS grading method, only use black-box-like convolutional neural networks to automatically extract and evaluate medical image features, lacking analysis of individual signs. That is, the evidence supporting the evaluation results is missing, and the reliability is not high enough, making it difficult to apply to clinical assisted diagnosis. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies in practical applications by proposing an intelligent liver imaging feature analysis and LI-RADS (Liver Imaging Reporting and Data System) classification system based on a multi-task model.

[0007] The objective of this invention is achieved through the following technical solution: an intelligent liver image sign analysis and LI-RADS classification system based on a multi-task model, which includes a medical image acquisition and processing module, a region of interest extraction module, a main label assignment module, a sign sub-label setting module, a LI-RADS classification module, a non-circular arterial phase high enhancement feature analysis module, a non-marginal "clarification" feature analysis module, an enhanced "pseudocapsule" feature analysis module, and a tumor size estimation module;

[0008] The medical image acquisition and processing module is used to acquire medical images of patients to be classified, including four phases: plain scan, arterial phase, portal venous phase, and delayed phase, and to perform preprocessing.

[0009] The region of interest extraction module is used to extract the region of interest from plain scan, arterial scan, portal venous scan, and delayed scan images;

[0010] The main label assignment module is used to process the main label of the image. It merges LI-RADS level 1 and 2 into one category and assigns the label "1" to indicate benign lesions. It merges LI-RADS level 3 and 4 into one category and assigns the label "2" to indicate possible malignant lesions. It sets LI-RADS level 5 to the label "3" to indicate primary liver cancer. It sets LI-RADS level M to the label "0" to indicate possible malignancy or confirmed malignancy but not necessarily primary liver cancer.

[0011] The feature sub-label setting module is used to assign corresponding sub-labels to the main features in each image according to the main features defined by the LI-RADS standard.

[0012] The LI-RADS classification module uses the VGG16 model, merging the regions of interest (ROI) of the arterial phase, portal venous phase, and delayed phase into a single image with three channels as the model input. The main label is assigned to the module as the model output. After training the model, the last convolutional layer is discarded, and the remaining part serves as the backbone structure of the final model. A new 4-class fully connected layer is added at the end of the backbone structure for LI-RADS classification.

[0013] The tumor size estimation module uses the VGG16 model to merge the region of interest in the portal venous phase image and two identical regions of interest in the delayed phase images into a three-channel image as the model input. The tumor diameter-related labels in the sign sub-label setting module are used as the model output. After training the model, the last convolutional layer is discarded, and the remaining part is used as the first branch structure of the final model. In addition, the input feature map of the first fully connected layer of the first branch structure is also input to the first fully connected layer of the trunk structure.

[0014] The non-circular arterial phase high enhancement feature analysis module uses the VGG16 model. It merges the region of interest in the plain scan image and two identical regions of interest in the arterial image into a three-channel image as the model input. The label related to the non-circular arterial phase high enhancement in the sign sub-label setting module is used as the model output. After training the model, the last convolutional layer is discarded, and the remaining part is used as the second branch structure of the final model. A 2-class fully connected layer is added at the end for non-circular arterial phase high enhancement feature analysis. In addition, the input feature map of the first fully connected layer of the second branch structure is also input to the first fully connected layer of the main structure.

[0015] The non-marginal "clarification" feature analysis module is used with the VGG16 model. It merges the region of interest in the arterial phase, the region of interest in the portal venous phase, and the region of interest in the delayed phase into a three-channel image as the model input. The non-marginal "clarification" related labels in the sign sub-label setting module are used as the model output. After training the model, the last convolutional layer is discarded, and the remaining part is used as the third branch structure of the final model. A 2-class fully connected layer is added at the end for non-marginal "clarification" feature analysis. In addition, the input feature map of the first fully connected layer of the third branch structure is also input to the first fully connected layer of the main structure.

[0016] The enhanced "pseudocapsule" feature analysis module is used to change the output of the last fully connected layer of the VGG16 model to 2-classification. The region of interest in the portal venous phase image, the region of interest in the delayed phase image, and the tumor edge image are merged into a three-channel image as the model input. The "pseudocapsule" related labels in the sign sub-label setting module are used as the model output. After training the model, the last convolutional layer is discarded, and the remaining part is used as the fourth branch structure of the final model. A 2-classification fully connected layer is added at the end to enhance the "pseudocapsule" feature analysis. In addition, the input feature map of the first fully connected layer of the fourth branch structure is also input to the first fully connected layer of the main structure.

[0017] Furthermore, the preprocessing in the medical image acquisition and processing module includes gray value truncation and adjusting the gray value range to 0-255.

[0018] Furthermore, the specific process of feature evaluation for obtaining tumor size information from the image by the tumor size estimation module is as follows:

[0019] (1) Using reverse pixel shuffling, the portal pulsation image I PV and delayed period image I DL The image is split into four separate images, each half the size of the original image in each direction, denoted as I. pv I dl ;

[0020] (2) Put I pv I dl Duplicate the image at a 1:2 ratio and merge it into a new image by channel, denoted as I. PDD_PU ;

[0021] (3) Will I PV I DL Duplicate the image at a 1:2 ratio and merge it into a new image by channel, denoted as I. PDD ;

[0022] (4) to I PDD Perform global average pooling, then copy and concatenate three times along the channel direction. The result is denoted as I. PDD_AP ;

[0023] (5) Will I PDD_PU with I PDD_AP Subtract them, then calculate the square of each pixel;

[0024] (6) Input the feature map output in step (5) into the convolutional layer, fully connected layer, activation layer (ReLU), dropout layer and convolutional layer in sequence, and output the tumor size assessment result.

[0025] Furthermore, in the enhanced "pseudocapsule" feature analysis module, the tumor edge image acquisition process is as follows:

[0026] (1) The Fragi Filter is used to perform a convolution operation on the region of interest in the portal venous phase image to obtain the edge region of the region of interest in the portal venous phase image;

[0027] (2) The Fragi Filter is used to perform a convolution operation on the region of interest in the delayed image to obtain the edge region of the region of interest in the delayed image;

[0028] (3) Add the results of steps (1) and (2) to obtain the tumor edge image.

[0029] Furthermore, the region of interest extracted by the region of interest extraction module needs to ensure that the input image contains the complete maximum cross-section of the tumor and that the tumor edge is at least 5 pixels away from the image edge.

[0030] Furthermore, the sub-labels set by the sign sub-label setting module for each major sign are as follows: if the tumor diameter is greater than or equal to 20 mm, the label is "S1"; if the tumor diameter is less than 20 mm, the label is "S0"; if there is high enhancement in the non-circular arterial phase, the label is "N1"; if there is no high enhancement in the non-circular arterial phase, the label is "N0"; if there is an enhanced "pseudocapsule", the label is "C1"; if there is no enhanced "pseudocapsule", the label is "C0"; if there is non-marginal "clarification", the label is "W1"; if there is no non-marginal "clarification", the label is "W0".

[0031] Furthermore, all models are trained using supervised contrastive learning.

[0032] The beneficial effects of this invention are as follows: This invention proposes an intelligent liver imaging sign analysis and LI-RADS classification system that doctors can rely on. It not only ensures the accuracy of LI-RADS grading but also provides a basis for the various judgment conditions required for LI-RADS grading, namely, analyzing the status of each major sign, providing doctors with a reference for practical use. Therefore, this invention has high application value in intelligent liver imaging sign analysis and LI-RADS classification and is of great significance in promoting the development of computer-aided diagnosis. Attached Figure Description

[0033] Figure 1 A framework diagram of a multi-task model for the LI-RADS grading task of liver cancer images;

[0034] Figure 2 A schematic diagram of the reverse pixel shuffling operation;

[0035] Figure 3 This is a schematic diagram of the tumor size estimation module. Detailed Implementation

[0036] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0037] This invention proposes an intelligent liver image sign analysis and LI-RADS classification system based on a multi-task model. It constructs a multi-task convolutional neural network model to learn the potential image features required for LI-RADS grading and the potential image features required for the main sign grading defined by the LI-RADS standard.

[0038] The system of this invention includes a medical image acquisition and processing module, a region of interest extraction module, a main label assignment module, a sign sub-label setting module, a LI-RADS classification module, a non-circular arterial phase high enhancement feature analysis module, a non-marginal "clarification" feature analysis module, an enhanced "pseudocapsule" feature analysis module, and a tumor size estimation module;

[0039] The medical image acquisition and processing module is used to acquire medical images of patients to be classified, including four phases: plain scan, arterial phase, portal venous phase, and delayed phase, and to perform preprocessing, including gray value truncation and adjusting the gray value range to 0-255.

[0040] The region of interest extraction module is used to extract the region of interest from plain scan, arterial scan, portal venous scan, and delayed scan images, ensuring that the input image contains the complete maximum cross-section of the tumor and that the tumor edge is at least 5 pixels away from the image edge.

[0041] The main label assignment module is used to process the main label of the image. LI-RADS levels 1 and 2 are merged into one category and assigned label "1", LI-RADS levels 3 and 4 are merged into one category and assigned label "2", LI-RADS level 5 is assigned label "3", and LI-RADS level M is assigned label "0". The medical significance of each label is as follows:

[0042] "0" - Possibly malignant or diagnosed as malignant but not necessarily primary liver cancer.

[0043] "1" - Benign lesion,

[0044] "2" - may indicate a malignant lesion.

[0045] "3" - Primary liver cancer;

[0046] The sign sub-label setting module is used to assign corresponding sub-labels to the major signs in each image according to the major signs defined by the LI-RADS standard. The sub-labels for each major sign are as follows: if the tumor diameter is greater than or equal to 20 mm, the label is "S1"; if the tumor diameter is less than 20 mm, the label is "S0"; if there is high enhancement in the non-circular arterial phase, the label is "N1"; if there is no high enhancement in the non-circular arterial phase, the label is "N0"; if there is an enhanced "pseudocapsule", the label is "C1"; if there is no enhanced "pseudocapsule", the label is "C0"; if there is non-marginal "clarification", the label is "W1"; if there is no non-marginal "clarification", the label is "W0".

[0047] The LI-RADS classification module uses the VGG16 model, merging the regions of interest (ROI) of the arterial phase, portal venous phase, and delayed phase into a single image with three channels (similar to an RGB image) as the model input. The main label is assigned to the module as the model output. After training the model using supervised contrastive learning, the last convolutional layer is discarded, and the remaining part serves as the backbone structure of the final model. A new 4-class fully connected layer is added at the end of the backbone structure as the main classifier for LI-RADS classification.

[0048] The tumor size estimation module uses the VGG16 model to merge the region of interest (ROI) of the portal venous phase image and two identical ROIs of the delayed phase images into a single three-channel image as model input. The tumor diameter-related labels from the sign sub-labeling module are used as model output. After training the model using supervised contrastive learning, the last convolutional layer is discarded, and the remaining portion serves as the first branch structure of the final model, acting as the feature extraction module. The first fully connected layer of this first branch structure acts as a sub-classifier, and its input feature map is also fed into the first fully connected layer (main classifier) ​​of the main structure. The specific process of feature evaluation for obtaining tumor size information from the image in the tumor size estimation module is as follows: Figure 3 As shown:

[0049] (1) Using reverse pixel shuffling, the portal pulsation image I PV and delayed period image I DL The image is split into four separate images, each half the size of the original image in each direction, denoted as I. pv I dl This step reduces the feature map size to one-quarter of its original size and increases the number of channels to four times its original size.

[0050] (2) Put I pv I dl Duplicate the image at a 1:2 ratio and merge it into a new image by channel, denoted as I. PDD_PU ;

[0051] (3) Will I PV I DL Duplicate the image at a 1:2 ratio and merge it into a new image by channel, denoted as I. PDD ;

[0052] (4) to I PDD Perform global average pooling, then copy and concatenate three times along the channel direction. The result is denoted as I. PDD_AP ;

[0053] (5) Will I PDD_PU with I PDD_APSubtract, perform the Flatten operation to convert the feature map into one dimension, and calculate the square value of each pixel;

[0054] (6) Input the feature map output in step (5) into the convolutional layer, fully connected layer, activation layer (ReLU), dropout layer and convolutional layer in sequence, and output the tumor size assessment result.

[0055] The non-circular arterial phase high-intensity feature analysis module uses the VGG16 model. It merges the region of interest in the plain scan image and two identical regions of interest in the arterial image into a three-channel image as the model input. The label related to the non-circular arterial phase high enhancement in the sign sub-label setting module is used as the model output. After training the model using supervised contrastive learning, the last convolutional layer is discarded, and the remaining part is used as the second branch structure of the final model as the feature extraction module. A two-classification fully connected layer is added at the end as the sub-classifier 2 for non-circular arterial phase high enhancement feature analysis. In addition, the input feature map of the first fully connected layer of the second branch structure is also input to the first fully connected layer (main classifier) ​​of the main structure.

[0056] The non-marginal "clarification" feature analysis module uses the VGG16 model, merging the region of interest (ROI) of the arterial phase, portal venous phase, and delayed phase into a single image with three channels as the model input. The non-marginal "clarification" related labels in the sign sub-label setting module are used as the model output. After training the model using supervised contrastive learning, the last convolutional layer is discarded, and the remaining part is used as the third branch structure of the final model, serving as the feature extraction module. A two-class fully connected layer is added at the end as a sub-classifier 3 for non-marginal "clarification" feature analysis. In addition, the input feature map of the first fully connected layer of this third branch structure is also input into the first fully connected layer (main classifier) ​​of the main structure.

[0057] The enhanced "pseudocapsule" feature analysis module is used to change the output of the last fully connected layer of the VGG16 model to binary classification, and merges the region of interest in the portal venous phase image, the region of interest in the delayed phase image, and the tumor edge image into a single image with three channels as the model input. The tumor edge image acquisition process is as follows:

[0058] (1) The Fragi Filter is used to perform a convolution operation on the region of interest in the portal venous phase image to obtain the edge region of the region of interest in the portal venous phase image;

[0059] (2) The Fragi Filter is used to perform a convolution operation on the region of interest in the delayed image to obtain the edge region of the region of interest in the delayed image;

[0060] (3) Add the results of steps (1) and (2) to obtain the tumor edge image.

[0061] In the symptom sub-label setting module, the labels related to "false envelope" are enhanced as model output. After training the model using supervised contrastive learning, the last convolutional layer is discarded, and the remaining part is used as the fourth branch structure of the final model as the feature extraction module. A two-class fully connected layer is added at the end as the sub-classifier 4 to enhance the "false envelope" feature analysis. In addition, the input feature map of the first fully connected layer of the fourth branch structure is also input to the first fully connected layer (main classifier) ​​of the main structure.

[0062] Since the LI-RADS classification result is mainly determined by the performance of the main LI-RADS features, this invention adopts a multi-task processing approach, using the features of each main feature as a model branch and fusing them with the main body. The loss function is as follows:

[0063]

[0064] in For the total loss, The loss of the main trunk, The penalty terms correspond to the losses of the tumor size, non-circular and non-circular arterial phase high enhancement, non-marginal "clarification", and enhanced "pseudocapsule" feature extraction modules, respectively, with θ being the penalty term coefficient.

[0065] The final model obtained by this invention is as follows Figure 1 As shown, the topmost part is the backbone model, responsible for handling the LI-RADS classification task. The remaining parts are branch models, responsible for handling the classification tasks of each major feature, and providing auxiliary feature maps for the LI-RADS classification model in the feature fusion part to improve the classification accuracy of the backbone model.

[0066] Figure 2 This diagram illustrates the reverse pixel shuffling operation. This operation requires no parameters or calculations and can split a 2N×2N image into four N×N images, achieving both high efficiency and preservation of the original image information.

[0067] Compared to recent deep learning methods, this invention achieves higher accuracy and stronger reliability in the LI-RADS classification of liver cancer images. Table 1 shows the comparison results of classification accuracy and AUC (Area Under Curve) between this method and existing methods in the LI-RADS classification task for liver cancer images. In the experiments in Table 1, the difference between the proposed method and existing methods lies in that the proposed method extracts features entirely from medical images for classification, while existing methods not only use liver images from three phases but also require pre-measured tumor diameter data. Table 1 uses the average classification accuracy (accuracy) and AUC of the four classification methods as evaluation metrics. As can be seen from the table, the proposed method significantly improves accuracy compared to existing methods, demonstrating the effectiveness of this method.

[0068] Table 1 compares the accuracy and AUC of our method with existing methods in the LI-RADS grading task for liver cancer images.

[0069]

[0070] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. An intelligent liver imaging feature analysis and LI-RADS classification system based on a multi-task model, characterized in that, The system includes a medical image acquisition and processing module, a region of interest extraction module, a main label assignment module, a sign sub-label setting module, a LI-RADS classification module, a non-circular arterial phase high enhancement feature analysis module, a non-marginal "clarification" feature analysis module, an enhanced "pseudocapsule" feature analysis module, and a tumor size estimation module; The medical image acquisition and processing module is used to acquire medical images of patients to be classified, including four phases: plain scan, arterial phase, portal venous phase, and delayed phase, and to perform preprocessing. The region of interest extraction module is used to extract the region of interest from plain scan, arterial scan, portal venous scan, and delayed scan images; The main label assignment module is used to process the main label of the image. It merges LI-RADS level 1 and 2 into one class and assigns the label "1" to indicate benign lesions. It merges LI-RADS level 3 and 4 into one class and assigns the label "2" to indicate possible malignant lesions. It sets LI-RADS level 5 to the label "3" to indicate primary liver cancer. It sets LI-RADS level M to the label "0" to indicate possible malignancy or confirmed malignancy but not necessarily primary liver cancer. The feature sub-label setting module is used to assign corresponding sub-labels to the main features in each image according to the main features defined by the LI-RADS standard. The LI-RADS classification module uses the VGG16 model, merging the regions of interest (ROI) of the arterial phase, portal venous phase, and delayed phase into a single image with three channels as the model input. The main label is assigned to the module as the model output. After training the model, the last convolutional layer is discarded, and the remaining part serves as the backbone structure of the final model. A new 4-class fully connected layer is added at the end of the backbone structure for LI-RADS classification. The tumor size estimation module uses the VGG16 model to merge the region of interest (ROI) of the portal venous phase image and two identical ROIs of the delayed phase images into a single three-channel image as model input. The tumor diameter-related labels from the sign sub-labeling module are used as model output. After training, the last convolutional layer is discarded, and the remaining portion becomes the first branch structure of the final model. Furthermore, the input feature map of the first fully connected layer of this first branch structure is also input to the first fully connected layer of the main structure. The specific process of feature evaluation for obtaining tumor size information from the image in the tumor size estimation module is as follows: (1) Using reverse pixel shuffling, the portal venous phase image I PV and delayed period image I DL The image is split into four separate images, each half the size of the original image in each direction, denoted as I. pv I dl ; (2) Put I pv I dl Duplicate the image at a 1:2 ratio and merge it into a new image by channel, denoted as I. PDD_PU ; (3) Put I PV I DL Duplicate the image at a 1:2 ratio and merge it into a new image by channel, denoted as I. PDD ; (4) to I PDD Perform global average pooling, then copy and concatenate three times along the channel direction. The result is denoted as I. PDD_AP ; (5) Put I PDD_PU with I PDD_AP Subtract them, then calculate the square of each pixel; (6) Input the feature map output in step (5) into the convolutional layer, fully connected layer, activation layer, dropout layer and convolutional layer in sequence, and output the tumor size assessment result; The non-circular arterial phase high enhancement feature analysis module uses the VGG16 model. It merges the region of interest in the plain scan image and two identical regions of interest in the arterial image into a three-channel image as the model input. The label related to the non-circular arterial phase high enhancement in the sign sub-label setting module is used as the model output. After training the model, the last convolutional layer is discarded, and the remaining part is used as the second branch structure of the final model. A 2-class fully connected layer is added at the end for non-circular arterial phase high enhancement feature analysis. In addition, the input feature map of the first fully connected layer of the second branch structure is also input to the first fully connected layer of the main structure. The non-marginal "clarification" feature analysis module uses the VGG16 model, merging the region of interest (ROI) of the arterial phase, portal venous phase, and delayed phase into a single image with three channels as the model input. The non-marginal "clarification" related labels in the sign sub-label setting module are used as the model output. After training the model, the last convolutional layer is discarded, and the remaining part is used as the third branch structure of the final model. A 2-class fully connected layer is added at the end of each branch for non-marginal "clarification" feature analysis. In addition, the input feature map of the first fully connected layer of the third branch structure is also input to the first fully connected layer of the main structure. The enhanced "pseudocapsule" feature analysis module uses the VGG16 model, merging the region of interest (ROI) of the portal venous phase image, the ROI of the delayed phase image, and the tumor edge image into a single image with three channels as the model input. The "pseudocapsule" related labels in the sign sub-label setting module are used as the model output. After training the model, the last convolutional layer is discarded, and the remaining part is used as the fourth branch structure of the final model. A two-class fully connected layer is added at the end to enhance the "pseudocapsule" feature analysis. In addition, the input feature map of the first fully connected layer of the fourth branch structure is also input to the first fully connected layer of the main structure.

2. The intelligent liver imaging feature analysis and LI-RADS classification system based on a multi-task model according to claim 1, characterized in that, The preprocessing in the medical image acquisition and processing module includes gray value truncation and adjusting the gray value range to 0-255.

3. The intelligent liver imaging feature analysis and LI-RADS classification system based on a multi-task model according to claim 1, characterized in that, In the enhanced "pseudocapsule" feature analysis module, the tumor edge image acquisition process is as follows: (1) The edge regions of the region of interest in the portal venous phase image are obtained by performing a convolution operation on the region of interest in the portal venous phase image using the Frangi Filter; (2) The edge regions of the region of interest in the delayed image are obtained by performing a convolution operation on the region of interest in the delayed image using the Frangi Filter; (3) Add the results of step (1) and step (2) together to obtain the tumor edge image.

4. The intelligent liver imaging feature analysis and LI-RADS classification system based on a multi-task model according to claim 1, characterized in that, The region of interest extracted by the region of interest extraction module must ensure that the input image contains the complete maximum cross-section of the tumor and that the tumor edge is at least 5 pixels away from the image edge.

5. The intelligent liver imaging feature analysis and LI-RADS classification system based on a multi-task model according to claim 1, characterized in that, The sub-labels set by the sign sub-label setting module for each major sign are as follows: If the tumor diameter is greater than or equal to 20 mm, the label is "S1"; if the tumor diameter is less than 20 mm, the label is "S0"; if there is high enhancement in the non-circular arterial phase, the label is "N1"; if there is no high enhancement in the non-circular arterial phase, the label is "N0"; if there is an enhanced "pseudocapsule", the label is "C1"; if there is no enhanced "pseudocapsule", the label is "C0"; if there is non-marginal "clarification", the label is "W1"; if there is no non-marginal "clarification", the label is "W0".

6. The intelligent liver imaging feature analysis and LI-RADS classification system based on a multi-task model according to claim 1, characterized in that, All models were trained using supervised contrastive learning.

Images