System and method for predicting risk of recurrence of hepatocellular carcinoma in patient after surgery

Abstract

The present invention relates to a method for predicting the risk of recurrence of hepatocellular carcinoma (HCC). Specifically, the present invention proposes a deep learning model capable of integrating information of a multi-phase CT image and clinical data to predict an HCC recurrence risk within 1-5 years after treatment. Experimental results show that the models are superior to a traditional histological microvascular invasion (MVI)-based prediction method in the aspect of predicting the HCC recurrence risk.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 588,757 filed on October 9, 2023, and the disclosure of which is incorporated herein by reference in its entirety. Technical Field

[0002] The present invention relates to the field of medical technology, in particular to the treatment and prognosis of liver cancer, with a focus on hepatocellular carcinoma, which is a major type of primary liver cancer. Background Art

[0003] Primary liver cancer is the seventh most common cancer and the third leading cause of cancer death worldwide. Hepatocellular carcinoma (HCC) is the main type of primary liver cancer. Treatment strategies including liver resection, liver transplantation and local ablation are feasible first-line treatment options for early-stage HCC. However, in the Asia-Pacific region, liver resection is still considered the main curative treatment, mainly due to the scarcity of organ donations. Despite the experience of curative surgical treatment, recurrence of early-stage HCC within two years after surgery remains common, and the overall risk of recurrence within five years exceeds 50%. Therefore, the ability to predict postoperative recurrence is crucial for risk stratification and may affect subsequent surveillance strategies. High-risk patients can also be candidates for neoadjuvant or adjuvant therapy, although most such strategies are currently still in the clinical trial stage.

[0004] Medical data including clinical records, computed tomography (CT) images, magnetic resonance imaging (MRI), and multispectral imaging (MSI) have been used to model the risk of HCC recurrence. For example, Shiyin Hong et al. reported an artificial intelligence model for predicting the risk of HCC recurrence based on six tumor types and related regions. 1 Ailian Liu et al. 2 and Sijie Niu et al. 3 reported machine learning and deep learning models for MRI images, respectively. Huang, Y. et al. 4 and Zeng, J. et al. 5 The aim is to identify patients at risk after surgical resection using machine learning and Cox models for clinical variables, respectively.

[0005] To improve the prediction accuracy of the model, features are usually extracted from the region of interest by combining CT images and clinical variables, and a machine learning model or Cox regression model is constructed for survival risk prediction. 6,7,8 .

[0006] Histological features are often used in training datasets to ensure that the model is predictive after surgery. 9The limitation of extracting features from the region of interest is that the information that can be extracted is limited, and potential tumors or high-risk areas may be missed due to human errors when marking the regions.

[0007] Therefore, there is a need in the art for more advanced and accurate prediction models that can minimize human errors by analyzing the entire imaging dataset rather than just a specific region of interest. Summary of the invention

[0008] The present invention aims to calculate the risk of HCC recurrence after curative surgery. Given that deep learning has demonstrated its capabilities in HCC research, the present invention developed a novel deep learning model for predicting the risk of HCC recurrence after curative surgery based on preoperative CT images.

[0009] The present invention helps in prognostic assessment in the preoperative stage and can predict early and late recurrence. To improve the effectiveness of feature extraction from the entire CT image, two strategies are adopted: applying liver segmentation technology to identify slices with intrahepatic lesions; and applying a two-branch neural network architecture with hepatic artery phase and portal vein phase as input respectively to improve contrast for better identification of tumor lesions.

[0010] In addition, the present invention also presents several adapted versions of Recurr-NET to facilitate prediction in different scenarios: using only CT data, combining CT with primary clinical data, and integrating CT with comprehensive clinical data.

[0011] In a first aspect, the present invention provides a method for predicting the risk of HCC recurrence in a patient after surgery. The method comprises obtaining one or more preoperative CT images from the patient's liver, wherein the one or more preoperative CT images at least comprise a hepatic artery phase image and a portal vein phase image, and the images comprise features associated with tumor lesions; inputting the one or more preoperative CT images into a multiphase deep learning model, and training the multiphase deep learning model with the one or more preoperative CT images; correspondingly extracting image features to generate a comprehensive image-based representation of the one or more preoperative CT images; and calculating a predicted risk score for the HCC based on the extracted image features to plan postoperative treatment and monitoring within a predetermined time period after the surgery.

[0012] In another embodiment, the method further comprises the step of segmenting the liver region in the one or more preoperative CT images using a liver segmentation model before inputting the one or more preoperative CT images into the multiphase deep learning model.

[0013] In one embodiment, the multi-phase deep learning model is configured to extract image features from the one or more pre-operative CT images.

[0014] In one embodiment, the multiphase deep learning model comprises a first residual network architecture, the first residual network architecture applies the hepatic artery phase image; a second residual network architecture, the second residual network architecture applies the portal vein phase image; and a multiphase residual network random survival forest (RSF) model. The RSF model combines the extracted image features to generate the predicted risk score. The RSF model uses log-rank splitting and bootstrapping to analyze right-censored survival data.

[0015] In one embodiment, each of the two residual network architectures includes at least one backbone block, at least one convolutional block, and at least one identity residual module.

[0016] In another embodiment, the method further comprises the step of inputting preoperative clinical parameters of the patient into the multi-phase deep learning model.

[0017] In one embodiment, the RSF model combines the extracted image features with the preoperative clinical parameters to generate the predicted risk score.

[0018] In one embodiment, the preoperative clinical parameter is selected from age, gender, hepatitis B surface antigen status, hepatitis C virus antibody status, history of fatty liver disease, alpha-fetoprotein level and Model for End-Stage Liver Disease (MELD) score.

[0019] In another embodiment, the preoperative clinical parameters further comprise smoking status, comorbidities, use of antiviral therapy for hepatitis B, baseline blood test results comprising platelet count, prothrombin time, albumin, alpha-fetoprotein level, or a combination thereof.

[0020] In one embodiment, the multi-phase deep learning model is optimized using an Adam optimizer in combination with batch normalization techniques.

[0021] In one embodiment, the predetermined time period ranges from 1 year to 5 years.

[0022] In a second aspect, the present invention provides a system for predicting the risk of recurrence of hepatocellular carcinoma. The system includes a processor, a memory storing instructions, and a user interface for displaying the output of the predicted risk score. The memory storing instructions causes the system to perform the following steps when executed by the processor: obtaining one or more preoperative CT images from the patient's liver, wherein the one or more preoperative CT images at least include a hepatic artery phase image and a portal vein phase image, and the image is characterized by characteristics associated with tumor lesions; inputting the one or more preoperative CT images into a multiphase deep learning model, and training the multiphase deep learning model with the one or more preoperative CT images; generating a comprehensive image-based representation of the preoperative CT image by cascading the features of the extracted images; and calculating the predicted risk score of the HCC based on the visual information of the comprehensive image-based representation to plan postoperative treatment and monitoring within a predetermined time period after the surgical operation.

[0023] In one embodiment, the memory storing instructions further causes the system to use a liver segmentation model to segment liver regions in the one or more preoperative CT images before inputting the one or more preoperative CT images into the multiphase deep learning model.

[0024] In one embodiment, the multiphase deep learning model comprises an image model based on two residual network architectures and a multiphase residual network RSF model, wherein the RSF model combines the extracted image features to generate the predicted risk score. The two residual network architectures each apply the hepatic artery phase image and the portal vein phase image, respectively.

[0025] In one embodiment, each of the two residual network architectures includes at least one backbone block, at least one convolutional block, and at least one identity residual module (ID).

[0026] In one embodiment, the RSF model uses the log-rank split method and the bootstrap method to analyze right-censored survival data.

[0027] In one embodiment, the memory storing instructions further causes the system to input preoperative clinical parameters of the patient into the multi-phase deep learning model.

[0028] In one embodiment, the preoperative clinical parameters include age, gender, hepatitis B surface antigen status, hepatitis C virus antibody status, history of fatty liver disease, alpha-fetoprotein level, MELD score, or a combination thereof.

[0029] In another embodiment, the preoperative clinical parameters further comprise smoking status, comorbidities, use of antiviral therapy for hepatitis B, baseline blood test results comprising platelet count, prothrombin time, albumin, alpha-fetoprotein level, or a combination thereof.

[0030] The present invention has the following advantages: (1) All clinical parameters used in CT scans and model training were preoperative, thus excluding any postoperative information and thus enabling predictions in the preoperative stage; (2) The model uses the entire CT image as input instead of extracting features from labeled tumor regions, thereby eliminating the possibility of missing valuable information and reducing the risk of human error; (3) The robustness of the model in predicting HCC recurrence was validated on an internal validation dataset consisting of 135 samples from Hong Kong and an external testing dataset consisting of 560 samples from Taiwan. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Embodiments of the present invention are described in more detail below with reference to the accompanying drawings, in which:

[0032] Figure 1 A schematic diagram of the system of the present invention is shown;

[0033] Figure 2 Recurr-NET CT The image embedding block represents the overall structure of the network, and the image embedding block consists of a backbone block, a convolutional block, and an identity residual module;

[0034] Figure 3 shows the architecture of the RSF model;

[0035] Figures 4A-4D The cumulative probability of recurrence for both the internal validation cohort and the external test cohort at year 2 and year 5 is shown; and

[0036] Figures 5A-5D The ability of the model to predict all-cause mortality in patients with HCC at 2 and 5 years in both the internal validation cohort and the external testing cohort is shown. DETAILED DESCRIPTION

[0037] In the following description, a system and method for predicting the risk of HCC recurrence in a patient after surgery is proposed as a preferred example. It will be apparent to those skilled in the art that modifications including additions and / or substitutions may be made without departing from the scope and spirit of the invention. Specific details may be omitted so as not to obscure the invention; however, this disclosure is written to enable those skilled in the art to practice the teachings herein without excessive experimentation.

[0038] like Figure 1 As shown, system 100 includes a processor 101 and a memory 102 storing instructions. When executed by processor 101, instructions 102 cause system 100 to perform a series of steps aimed at predicting HCC recurrence. The core functionality focuses on processing preoperative CT images and combining extracted image features with relevant clinical data to generate an accurate prediction of recurrence risk. Finally, system 100 includes a user interface 104 for displaying the output of the predicted risk score, thereby providing clinicians with easy access to key information that affects treatment decisions.

[0039] The system 100 begins by obtaining a preoperative CT image of the patient's liver. These CT images contain at least the hepatic artery phase and the portal vein phase, both of which are essential for identifying features associated with tumor lesions. The CT scan typically covers the area from the lung base to the iliac crest, ensuring that all relevant liver tissue is included.

[0040] The memory 102 storing instructions further causes the system 100 to perform liver segmentation on the preoperative CT image before the image is processed by the multiphase deep learning model 103. The segmentation utilizes a segmentation model trained on a public dataset such as the Liver Tumor Segmentation Challenge (LiTS) to separate liver-specific slices from the full CT scan. By focusing on the liver region, the system 100 improves the accuracy of the deep learning model 103, thereby ensuring that only relevant areas are analyzed for predicting recurrence.

[0041] The system 100 includes a multi-phase deep learning model 103 for processing segmented CT images. This model includes an image model based on two residual network architectures, each of which applies a hepatic artery phase image and a portal vein phase image, respectively. The two network branches extract corresponding image features from different phases, and then cascade the image features to generate a representation based on a comprehensive image.

[0042] The RSF model combines the extracted image features to generate a predicted risk score using the log-rank splitting method and the bootstrap method to analyze right-censored survival data, thereby ensuring that the predicted risk score accurately reflects the patient's risk of recurrence within a specified time period. The model is trained on a data set comprising both an internal validation cohort and an external test cohort. The loss function used is based on the Cox proportional hazard function with regularization to consider recurrence data.

[0043] The memory 102 storing instructions also causes the system 100 to input the patient's preoperative clinical parameters into the multi-phase deep learning model 103. These clinical parameters are combined with the image-based features to improve the overall predictive accuracy of the system 100.

[0044] Clinical data can include primary factors such as the patient's age, sex, hepatitis B surface antigen status, hepatitis C virus antibody status, history of fatty liver disease, alpha-fetoprotein level, and MELD score. In some cases, additional clinical data such as smoking status, comorbidities, and baseline blood test results (including platelet count, prothrombin time, albumin level, and alpha-fetoprotein level) are also incorporated.

[0045] After processing the CT images and clinical data, the system 100 calculates a predicted risk score for HCC recurrence. This risk score is used to guide postoperative treatment planning and ongoing monitoring of the patient. Finally, the system 100 outputs the predicted risk score to a user interface 104 for display, providing clinicians with easy access to key information that influences treatment decisions. This system 100 provides a powerful tool for preoperative prediction of HCC outcomes, enabling more informed clinical decisions and potentially improving patient outcomes through targeted postoperative interventions.

[0046] In addition, the method of the present invention describes how to use a deep learning model to integrate information from multi-phase CT images, extract tumor-related features, and ultimately predict the risk of recurrence of HCC.

[0047] The method of the present invention involves training three different versions of deep learning models to predict the risk of HCC recurrence 1 to 5 years after treatment. These models include: Recurr-NET CT 、Recurr-NET LITE and the complete Recurr-NET model, which will be described in detail below.

[0048] I.Recurr-NET CT

[0049] Recurr-NET CT is a deep learning model that predicts HCC recurrence using only preoperative CT images.

[0050] CT images are stored in the Digital Imaging and Communications in Medicine (DICOM) format, which is a standard for processing, storing, and transmitting medical imaging information. CT images are segmented using a liver segmentation model trained on a public dataset such as LiTS (Liver Tumor Segmentation Challenge) to focus on liver-specific slices. These CT images are then normalized and resized to a standard size of 512×512×32. Data augmentation techniques with a 15-degree rotation were applied to enhance the robustness of the model. The deep learning model is built on a residual network (ResNet) architecture, utilizing the arterial and portal venous phase information in CT images. The model generates a comprehensive image-based representation by extracting features from these two phases and concatenating them. The training process is optimized using batch normalization and the Adam optimizer to achieve faster convergence.

[0051] AUROC (Area Under the Receiver Operating Characteristic Curve) is a metric used to evaluate the performance of a binary classification model. The model is evaluated for its ability to distinguish between classes at different threshold settings. AUROC values ​​range from 0 to 1, where a value of 0.5 indicates that the model has no discriminative ability, meaning that its performance is no better than random guessing, while a value of 1.0 indicates perfect classification, meaning that the model is able to perfectly distinguish between classes. The AUROC value reflects the performance of the model in predicting HCC recurrence, where higher AUROC values ​​indicate better performance in identifying recurrence.

[0052] Recurr-NET CT The Recurr-NET CT AUROC values ​​between 0.756 and 0.796 were achieved for years 1-5, while in the external cohort, the AUROC values ​​were between 0.693 and 0.755. The high and stable AUROC values ​​obtained from the relatively large external test group demonstrate the robustness of the model of the present invention.

[0053] II.Recurr-NET LITE

[0054] By incorporating additional baseline clinical parameters and preoperative CT images, the Recurr-NET CT Built Recurr-NET LITE .

[0055] Basic clinical parameters can include age, gender, hepatitis B surface antigen, hepatitis C virus antibody, imaging history of fatty liver, alpha-fetoprotein and MELD score model. All clinical parameters are preoperative data.

[0056] With Recurr-NET CT Similarly, CT images are segmented, normalized, and resized, and then input into a ResNet-based architecture. The extracted image features are combined with clinical parameters to enhance the predictive power of the model.

[0057] With Recurr-NET CT In comparison, Recurr-NET LITE A slightly improved predictive performance was achieved. In the internal validation cohort, the model achieved diagnostic accuracy in predicting HCC recurrence from 1 to 5 years with AUROC values ​​ranging from 0.742 to 0.811, while in the external cohort, AUROC values ​​ranged from 0.711 to 0.760. The integration of clinical data with image features enhanced the accuracy of the model in predicting recurrence, thereby providing patients with more detailed risk stratification.

[0058] III. Complete Recurr-NET

[0059] The full Recurr-NET model is an advanced version that incorporates both preoperative CT images and comprehensive clinical parameters.

[0060] The complete Recurr-NET is a multi-phase residual network RSF deep learning model. LITE In addition to the baseline clinical parameters used in the , the full Recurr-NET also includes factors such as smoking status, comorbidities, use of antiviral therapy for hepatitis B, and additional blood test results (eg, platelet count, prothrombin time, albumin level).

[0061] The full Recurr-NET model was trained on multiphase contrast CT liver images of patients with histologically confirmed HCC from four centers in Hong Kong (internal cohort). The internal cohort was randomly divided into 80% for training and 20% for internal validation. In addition, an independent cohort from Taiwan was also used for external validation of Recurr-NET. The area under the curve (AUC), positive and negative predictive values ​​(PPV / NPV) of Recurr-NET were compared with microvascular invasion (MVI) by diagnostic accuracy and survival analysis, respectively.

[0062] For image and data processing, CT images were obtained from major global CT manufacturers (Siemen, General Electric (GE), Philips, and Toshiba) and stored in the Digital Imaging and Communications in Medicine (DICOM) format. The Hounsfield unit (HU) of the image was windowed to the range of [-160, 240], and the image was normalized to the range of [-0.5, 0.5] to remove irrelevant features. The original CT image size is 512 (height) × 512 (width) × N, where N represents the number of slices. The CT scan spans from the base of the lung to the iliac crest and contains three phases (non-contrast phase, hepatic artery phase, and portal vein phase). In order to enhance the performance of the deep learning model, a liver segmentation model trained on the public dataset LiTS (Liver Tumor Segmentation Challenge) was used to extract liver-related CT slices.

[0063] To facilitate computation, the CT scans were resized to 512 × 512 × 32. Two data augmentation methods were used: 15-degree clockwise rotation and 15-degree counterclockwise rotation. Multiple imputation was used to handle missing clinical data. Continuous variables were normalized to the range of [0, 1], and discrete variables were encoded as one-hot vectors.

[0064] Deep learning was performed on PyTorch 1.12.1, and the computing platform was powered by NVIDIA Tesla V100 graphics processing units and Intel(R) Xeon(R) Gold 6146CPU@3.20GHz with 12,256GB DDR4 memory. Using processed three-phase CT scans and preoperative clinical data, a multimodal multiphase residual network RSF deep learning model Recurr-NET was developed to predict the risk of HCC recurrence within 5 years for each patient.

[0065] For the model architecture, Recurr-NET consists of an image model based on the residual network (ResNet) structure and an RSF model. Figure 2 Recurr-NET CT First, the arterial and portal venous phase images of the patient are input into the two branches of the ResNet architecture to obtain the corresponding image features. The separate encoding path of each branch consists of a backbone block and three residual blocks, which contain convolution blocks and identity residual modules.

[0066] Each branch starts with a backbone block, which usually contains an initial convolutional layer and batch normalization. This block processes raw image data and prepares it for deeper layers.

[0067] After the backbone block, the image passes through a series of convolutional blocks, which are responsible for extracting increasingly abstract features from the image. These blocks consist of convolutional layers, activation functions (e.g., ReLU), and batch normalization. Convolutional blocks play a crucial role in learning the spatial hierarchy in image data.

[0068] Each convolutional block is followed by an identity residual module that contains skip connections (residual connections) that help alleviate the vanishing gradient problem during training. These skip connections allow the network to learn more efficiently by preserving gradients across layers.

[0069] Once the images from both phases pass through their respective residual networks, the extracted features are concatenated. This combined image-based representation integrates information from both the hepatic artery phase and the portal vein phase, thereby enhancing the model’s ability to detect subtle patterns associated with tumor lesions and other risk factors for recurrence. Risk scores are generated based on visual information transmitted through fully connected blocks. In the initial stage, Adam with an initial learning rate (lr) of 0.001 is used to optimize the training process. Batch normalization is used after each convolutional layer. This technique ensures that the activation output of the convolutional layer is correct, thereby achieving faster convergence.

[0070] At the same time, batch normalization makes it possible for each layer to learn independently of other layers and reduce data loss between processing layers. The average negative log partial likelihood with regularized Cox proportional hazard function is used as the loss function: Where N E＝1 is the number of patients with a history of recurrence. i 、E i and x i Respectively represent i th Observed event times, event indicators, and baseline data. Risk set represents the set of patients who are still at risk of recurrence at time t. is the log-hazard parameterized by β.

[0071] According to the principle of transfer learning, the 64-dimensional vector ( Figure 2 ) is used to represent image information.

[0072] This representation is combined with the patient's clinical data as needed before input into the RSF model ( Figure 3). RSF is a tree model for analyzing right-censored survival data. It uses the log-rank splitting method, which splits nodes by maximizing the difference between nodes. The model uses a bootstrap method to randomly select samples from the original data with replacement, thereby establishing a sample subset and an out-of-bag (OOB) data set. Secondly, it randomly selects the features of each sample to construct its corresponding survival tree. The Nelson-Aalen method is used to estimate the total cumulative risk of the RSF model. OOB data is used to estimate the prediction accuracy. The RSF model ultimately predicts the risk score for each patient The risk scores were then fitted into the Kaplan-Meier estimator to obtain the corresponding survival curves.

[0073] The Kaplan-Meier estimator based on empirical and censored data is used to calculate the survival function for each patient. The probability that a patient will not relapse at time t is defined as: where d i is the time t i The number of relapse cases at the time of i is at time t i The number of patients previously at risk of recurrence.

[0074] At time t j The survival probability S(t j ) can be iteratively calculated as follows: The Youden index was used to select the cutoff point and determine whether there was recurrence.

[0075] In the internal validation cohort, the complete Recurr-NET achieved excellent diagnostic accuracy in predicting 1-5 year HCC recurrence with AUROC values ​​ranging from 0.770 to 0.857, whereas in the external cohort, AUROC values ​​ranged from 0.758 to 0.798.

[0076] The combination of comprehensive clinical data and CT image features makes the full Recurr-NET model outperform Recurr-NET CT and Recurr-NET LITE , thus providing a highly accurate tool for preoperative prediction.

[0077] Examples

[0078] Internal validation and external testing of the full Recurr-NET were performed to compare recurrence outcomes predicted by MVI Comparison

[0079] More than 70% of HCC patients may relapse within 5 years after curative resection. Although MVI in histology can predict recurrence, it is based on the analysis of resection specimens and cannot provide preoperative prediction.

[0080] The present invention provides a deep learning model for predicting HCC recurrence using preoperative CT and clinical parameters. Chinese patients with resected histologically confirmed HCC recruited from four centers in Hong Kong and Taiwan were used for training, internal validation, and external testing, respectively.

[0081] This analysis included 1,231 patients (83.1% male, age 62.4+ / -10.7 years, median follow-up 65.1 months). 536 (43.9%), 135 (11.2%), and 560 (44.9%) were assigned to the training group, internal validation group, and external test group, respectively. The cumulative HCC recurrence rates at 2 and 5 years were 41.8% and 56.4%, respectively.

[0082] Figures 4A-4D Cumulative probabilities of HCC recurrence are shown for the internal validation cohort and the external testing cohort at years 2 and 5. These figures compare the performance of the full Recurr-NET model with that of the traditional histological MVI in predicting HCC recurrence. Figure 4A The cumulative probability of HCC recurrence at year 2 for the internal validation cohort is shown. This figure highlights the ability of Recurr-NET to predict early recurrence, demonstrating superior performance compared to MVI in identifying high-risk patients within two years after surgery. Figure 4B The cumulative power of HCC recurrence at 2 years in the external test cohort is presented. Figure 4A Similarly, it demonstrated that Recurr-NET had higher predictive accuracy than MVI, and the results were obtained in an independent patient population, thus validating the robustness of the model. Figure 4C The cumulative probability of HCC recurrence at year 5 is presented for the internal validation cohort. This long-term prediction shows that the full Recurr-NET consistently outperforms the MVI in recurrence risk stratification, indicating the validity of this model in predicting late recurrence. Figure 4D The cumulative probability of HCC recurrence at 5 years in the external test cohort is shown. Figure 4C Likewise, the full Recurr-NET maintained superior predictive performance over the MVI across different cohorts over a five-year period.

[0083] Figures 5A-5D A comparison between predictions made by the full Recurr-NET model and those based on histological MVI, a traditional method for assessing risk, is provided. Figure 5AThe cumulative probability of all-cause mortality at year 2 for the internal validation cohort is shown. The performance of Recurr-NET is compared with that of MVI, highlighting the superior ability of Recurr-NET to stratify patients according to their mortality risk within two years after surgery. Figure 5B The cumulative probability of all-cause mortality at 2 years in the external testing cohort is presented. Figure 5A Similarly, Recurr-NET outperformed MVI in predicting mortality, but the data were from a different, independent cohort of patients. Figure 5C Cumulative probabilities of all-cause mortality at year 5 are presented for the internal validation cohort. This long-term projection shows that the full Recurr-NET continues to provide more accurate results for mortality risk stratification compared with the MVI at the five-year point. Figure 5D The cumulative probability of all-cause mortality at 5 years for the external test cohort is shown, reinforcing the Figure 5C , showing that even over a five-year period, the full Recurr-NET maintained superior predictive performance over the MVI in an external patient population.

[0084] In the internal validation cohort, the complete Recurr-NET achieved an AUC of 0.857 (95% CI 0.783-0.921, PPV 0.744, NPV 0.807) and 0.770 (95% CI 0.681-0.849, PPV 0.884, NPV 0.543) in predicting recurrence at 2 and 5 years, which was significantly better than the predictive value of MVI (AUC 0.570 [95% CI 0.491-0.654], PPV 0.500, NPV 0.670 at 2 years; AUC 0.518 [95% CI 0.44-0.594], PPV 0.636, NPV 0.407 at 5 years) (p < 0.01 for both).

[0085] In the external testing cohort, the complete Recurr-NET achieved an AUC of 0.781 (95% CI 0.735-0.819, PPV 0.668, NPV 0.789) and 0.760 (95% CI 0.718-0.801, PPV 0.863, NPV 0.513) in predicting recurrence at 2 and 5 years, which was significantly superior to MVI (AUC 0.605, 95% CI 0.568-0.643, PPV 0.476, NPV 0.769 at 2 years; AUC 0.557, 95% CI 0.515-0.597, PPV 0.678, NPV 0.459 at 5 years) (p<0.05 for both). In both internal validation and external testing, the complete Recurr-NET was significantly more effective at year 2 (internal: 72.5% in the complete Recurr-NET vs. 50.0% in the MVI; external: 65.3% in the complete Recurr-NET vs. 46.6% in the MVI; p < 0.001 for both, Figures 4A-4B ) and year 5 (internal: 86.4% in full Recurr-NET vs. 62.5% in MVI; external: 81.4% in full Recurr-NET vs. 63.8% in MVI; p<0.001 for both, Figure 4C-4D ) was superior to MVI in terms of stratified recurrence risk.

[0086] Comparable patterns of all-cause mortality were observed in the internal validation cohort and the external testing cohort (data were available for both), with complete Recurr-NET at year 2 (internal: 31.9% vs. 14.3%; external: 32.7% vs. 18.9%, p<0.001 for both, Figure 5A-5B ) and year 5 (internal: 72.9% vs. 34.3%; external: 66.8% vs. 37.9%; p < 0.001 for both; Figure 5C-5D ) was superior to MVI in predicting all-cause mortality.

[0087] In conclusion, the complete Recurr-NET outperforms MVI in predicting early and late HCC recurrence and has the potential to become a novel tool for preoperative prediction of HCC outcome.

[0088] definition

[0089] Throughout this specification, unless the context requires otherwise, the word "comprise" or variations such as "comprises" or "comprising" will be understood to imply the inclusion of the stated integer or groups of integers, but not the exclusion of any other integer or groups of integers. It should also be noted that in this disclosure and particularly in the claims and / or paragraphs, terms such as "comprises", "comprised", "comprising", etc. may have the meaning ascribed to them in U.S. patent law; for example, such that it allows elements not expressly recited, but excludes elements found in the prior art or elements that affect the basic or novel characteristics of the invention.

[0090] Furthermore, throughout the specification and claims, unless the context requires otherwise, the word "include" or variations such as "includes" or "including", will be understood to imply the inclusion of a stated integer or group of integers but not the exclusion of any other integer or group of integers.

[0091] References in this specification to "one embodiment", "an embodiment", "example embodiment", etc. indicate that the described embodiment may include certain features, structures, or characteristics, but not every embodiment necessarily includes the certain features, structures, or characteristics. In addition, such phrases do not necessarily refer to the same embodiment. Further, when a certain feature, structure, or characteristic is described in conjunction with an embodiment, it is considered that it is within the knowledge of those skilled in the art to affect such feature, structure, or characteristic in conjunction with other embodiments, whether or not explicitly described.

[0092] As used herein, the terms "approximately", "basically", "substantially" and "about" are used to describe and explain small changes. When used in combination with an event or situation, the term can refer to the situation where the event or situation occurs exactly, as well as the situation where the event or situation occurs closely. As used herein with respect to a given value or a given range, the term "about" generally means within the range of ±10%, ±5%, ±1% or ±0.5% of the given value or given range. Ranges may be indicated herein as one end value to another end value or between two end values. Unless otherwise specified, all ranges within the ranges disclosed in the present disclosure include the end values. When referring to "substantially" the same numerical value or characteristic, the term may refer to a value within ±10%, ±5%, ±1% or ±0.5% of the average value of the value.

[0093] In the preparation method described herein, except for the case of explicitly describing the time or operation sequence, the steps can be performed in any order without departing from the principles of the present invention. The description of a step being performed first in a certain claim, and then subsequently performing several other steps should be considered to mean that the first step is performed before any other step, but the other steps can be performed in any suitable order unless the order is further described in other steps. For example, the claim element describing "step A, step B, step C, step D and step E" should be interpreted as meaning that step A is performed first and step E is performed last, and step B, step C and step D can be performed in any order between step A and step E, and the order still falls within the literal scope of the process claimed for protection. A given step or a subset of steps can also be repeated. In addition, unless the clear claim language describes the separate execution of the specified steps, the specified steps can be performed simultaneously.

[0094] Hepatocellular carcinoma (HCC) is a type of liver cancer that originates from hepatocytes, the main cells of the liver. HCC is the most common primary liver cancer and is often associated with chronic liver disease, including cirrhosis and hepatitis B or C infection.

[0095] The term "preoperative" refers to the period of time before a surgical procedure is performed. In terms of time, it generally means any time from the initial diagnosis of the condition to the day of surgery. For imaging like CT scans, "preoperative" generally means that the images were taken a few weeks to months before surgery, depending on the nature of the condition and the urgency of the procedure.

[0096] Other definitions of selected terms used herein can be found in the detailed description of the present invention and are applicable throughout the text. Unless otherwise defined, all other technical terms used herein have the same meaning as those generally understood by those of ordinary skill in the art to which the present invention belongs.

[0097] References: The disclosures of the following references are incorporated by reference 1.Shiyin Hong, WD, Mengxin Tian, ​​Weiren Liu, Zheng Tang, Kun Qian, Zhixun Wang, Hao Zhou, Yucheng Guo, Jintao Qiu, AI-based analysis method and system for prognosis of postoperative recurrence of early hepatocellular carcinoma. 2021-06-29: CN113591919A. 2. Ailian Liu, YG, Yin Zhao, Nan Wang, Xin Li, Jinjun Wu, Tingfan Wu, Xinhe Zhang, Imaging omics feature processing method for predicting recurrence of hepatocellular carcinoma after surgical resection based on multi-modal MRI (magnetic resonance imaging) images. 2019-12-11: CN110916666A. 3.Sijie Niu,FL,Xizhan Gao,Yiqi Wang,Qiang Liu,Jiwen Dong,MR image liver cancer postoperative recurrence risk prediction method and system based on deep learning.2021-04-28:CN113180633A. 4. Huang, Y. et al., Development and Validation of a Machine Learning Prognostic Model for Hepatocellular Carcinoma Recurrence After Surgical Resection. Front Oncol, 2020.10: 593741. 5. Zeng, J. et al., Development of a machine learning model to predict early recurrence for hepatocellular carcinoma after curative resection. Hepatobiliary Surg Nutr, 2022.11(2): 176-187. 6. Zhou, Y. et al., CT-based radiomics signature: a potential biomarker for preoperative prediction of early recurrence in hepatocellular carcinoma. Abdominal Radiol (New York (NY)), 2017. 42(6): 1695-1704. 7. Ji, GW, et al., Machine-learning analysis of contrast-enhanced CT radiomics predicts recurrence of hepatocellular carcinoma after resection: A multi-institutional study. EBioMedicine, 2019. 50: 156–165. 8. Liu, Y. et al., CT radiomics combined with clinical variables for predicting the overall survival of hepatocellular carcinoma patients after hepatectomy. Translational Oncology, 2022.26. 9. Lee, IC, et al., Evolutionary Learning-Derived Clinical-Radiomic Models for Predicting Early Recurrence of Hepatocellular Carcinoma after Resection. Liver Cancer, 2021.10(6): 572-582.

Claims

1. A method for predicting the risk of recurrence of hepatocellular carcinoma (HCC) in a patient after surgery, characterized in that: The method comprises: obtaining one or more preoperative computed tomography (CT) images from the patient's liver, wherein the one or more preoperative CT images include at least a hepatic artery phase image and a portal vein phase image, wherein the hepatic artery phase image and the portal vein phase image include features associated with a tumor lesion; Inputting the one or more preoperative CT images into a multiphase deep learning model, and training the multiphase deep learning model using the one or more preoperative CT images; generating a representation of a composite image based on the one or more preoperative CT images by concatenating the extracted image features; and A predictive risk score for the HCC is calculated based on the integrated image-based characterization to plan postoperative treatment and monitoring within a predetermined time period after the surgical procedure.

2. The method according to claim 1, wherein the method further comprises a step of using a liver segmentation model to segment the liver region in the one or more preoperative CT images before inputting the one or more preoperative CT images into the multiphase deep learning model.

3. The method of claim 1, wherein the multiphase deep learning model is configured to extract image features from the one or more preoperative CT images.

4. The method according to claim 1, wherein the multi-phase deep learning model comprises: a first residual network architecture, wherein the first residual network architecture applies the hepatic artery phase image; a second residual network architecture, wherein the second residual network architecture applies the portal vein phase image; as well as A multi-phase residual network random survival forest (RSF) model is provided to combine the extracted image features to generate the predicted risk score.

5. The method according to claim 4, wherein the first residual network architecture and the second residual network architecture both include at least one backbone block, at least one convolutional block and at least one identical residual module.

6. The method of claim 4, wherein the RSF model uses a log-rank split method and a bootstrap method to analyze right-censored survival data.

7. The method of claim 1, wherein the method further comprises inputting preoperative clinical parameters of the patient into the multiphase deep learning model.

8. The method of claim 7, wherein the RSF model combines the extracted image features with the preoperative clinical parameters to generate the predicted risk score.

9. The method of claim 8, wherein the preoperative clinical parameters are selected from age, gender, hepatitis B surface antigen status, hepatitis C virus antibody status, history of fatty liver disease, alpha-fetoprotein level, and Model for End-Stage Liver Disease (MELD) score.

10. The method of claim 9, wherein the preoperative clinical parameters further include smoking status, comorbidities, use of antiviral therapy for hepatitis B, baseline blood test results including platelet count, prothrombin time, albumin, alpha-fetoprotein level, or a combination thereof.

11. The method according to claim 1, wherein the multi-phase deep learning model is optimized using an Adam optimizer in combination with a batch normalization technique.

12. The method according to claim 1, wherein the predetermined period of time ranges from 1 year to 5 years.

13. A system for predicting the risk of recurrence of hepatocellular carcinoma (HCC), characterized in that: The system comprises: processor; A memory storing instructions, wherein the memory storing instructions, when executed by the processor, causes the system to perform the following steps: obtaining one or more preoperative computed tomography (CT) images from the patient's liver, wherein the one or more preoperative CT images include at least a hepatic artery phase image and a portal vein phase image, the hepatic artery phase image and the portal vein phase image being characterized by characteristics associated with a tumor lesion; Inputting the one or more preoperative CT images into a multiphase deep learning model, and training the multiphase deep learning model using the one or more preoperative CT images; generating a composite image-based representation of the one or more preoperative CT images by concatenating the extracted image features; and calculating a predictive risk score for the HCC based on the integrated image-based characterization to plan postoperative treatment and monitoring within a predetermined time period after the surgical procedure; and A user interface is provided for displaying the output of the predicted risk score.

14. The system of claim 13, wherein the memory storing instructions further causes the system to use a liver segmentation model to segment liver regions in the one or more preoperative CT images before inputting the one or more preoperative CT images into the multiphase deep learning model.

15. The system of claim 13, wherein the multi-phase deep learning model comprises: An image model based on two residual network architectures, wherein the two residual network architectures are respectively applied to the hepatic artery phase image and the portal vein phase image; A multi-phase residual network random survival forest (RSF) model is provided to combine the extracted image features to generate the predicted risk score.

16. The system of claim 15, wherein each of the two residual network architectures comprises at least one backbone block, at least one convolutional block, and at least one identity residual module.

17. The system of claim 15, wherein the RSF model uses a log-rank split method and a bootstrap method to analyze right-censored survival data.

18. The system of claim 13, wherein the memory storing instructions further causes the system to input preoperative clinical parameters of the patient into the multi-phase deep learning model.

19. The system of claim 18, wherein the preoperative clinical parameters include age, gender, hepatitis B surface antigen status, hepatitis C virus antibody status, history of fatty liver disease, alpha-fetoprotein level, MELD score, or a combination thereof.

20. The system of claim 19, wherein the preoperative clinical parameters further include smoking status, comorbidities, use of antiviral therapy for hepatitis B, baseline blood test results including platelet count, prothrombin time, albumin, alpha-fetoprotein levels, or a combination thereof.

Images