Early liver cancer postoperative recurrence prediction model and method based on radiomics phenotype

Abstract

The invention relates to the technical field of medical imaging omics analysis, and discloses an early-stage liver cancer postoperative recurrence prediction model and method based on imaging omics phenotypes, and the method comprises the following steps: data collection and preprocessing, feature screening and standardization, model construction, model verification and risk layering. The data acquisition and preprocessing is to obtain preoperative enhanced CT image data and clinical information of an early-stage hepatocellular carcinoma (HCC) patient meeting the agralan standard, and the clinical information comprises serum alpha fetoprotein (AFP) level and tumor number. According to the image omics phenotype-based early liver cancer postoperative recurrence prediction model and method, pre-operation and post-operation dual models are constructed by integrating pre-operation enhanced CT image omics characteristics (including first-order statistics, texture and wavelet characteristics), clinical parameters (such as serum alpha fetoprotein and tumor number) and post-operation pathological variables (such as microvascular invasion and satellite nodules).

Description

Technical Field

[0001] This invention relates to the field of medical radiomics analysis technology, specifically to a model and method for predicting early postoperative recurrence of liver cancer based on radiomics phenotypes. Background Technology

[0002] Hepatocellular carcinoma (HCC) is the sixth most common cancer and the second leading cause of cancer death worldwide. For early-stage HCC patients meeting the Milan criteria (single nodule ≤5 cm or up to three nodules with a maximum diameter ≤3 cm, without major vascular invasion or extrahepatic metastasis), hepatectomy or liver transplantation is the preferred treatment. Although liver transplantation can eradicate both the tumor and the diseased liver, donor shortages limit its application. Therefore, hepatectomy is the preferred treatment for early-stage HCC patients with good liver function, while liver transplantation is recommended for patients with decompensated cirrhosis. However, the 5-year recurrence rate in these patients remains as high as 50%-60%.

[0003] Existing HCC staging systems (such as the Barcelona Clinical Hepatocellular Carcinoma Staging System [BCLC], the Italian Clinical Hepatocellular Carcinoma Program [CLIP], and the TNM system) have limitations in prognostic assessment and cannot provide quantitative information on recurrence risk. While the Korean model and the Early Recurrence After Surgery (ERASL) model proposed in recent years are designed for predicting HCC recurrence after surgery, they have not been validated in the optimal indication population for early-stage HCC.

[0004] Radiomics, by transforming medical images into quantitative features, can reveal tumor heterogeneity. Previous studies have reported the predictive value of radiomics for HCC recurrence, but external validation is lacking. This study aims to establish and externally validate a radiomics-based recurrence risk model, comparing its performance with existing staging systems and other predictive models. Therefore, a radiomics-based model and method for predicting early-stage liver cancer recurrence after surgery are proposed to address the aforementioned issues. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a predictive model and method for early-stage liver cancer recurrence based on radiomics phenotypes. It has advantages such as high prediction accuracy, low error, effective risk stratification, and clinical applicability, and solves the problems of insufficient accuracy, lack of external validation, and lack of individualized decision support in the risk assessment of early-stage liver cancer recurrence by existing staging systems and prediction models.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A model and method for predicting postoperative recurrence of early-stage liver cancer based on radiomics phenotypes, comprising the following steps: data acquisition and preprocessing, feature screening and standardization, model construction, model validation, and risk stratification. As a preferred embodiment of the present invention, the data acquisition and preprocessing involves obtaining preoperative enhanced CT image data and clinical information of patients with early-stage hepatocellular carcinoma (HCC) meeting the Milan criteria. The clinical information includes serum alpha-fetoprotein (AFP) levels and the number of tumors. The tumor and the 4 mm peritumoral region are manually delineated using 3D Slicer software. The radiomics features of each lesion in the arterial and portal venous phases are extracted using the Pyradiomics software package, including first-order statistical features, texture features, and wavelet features. The radiomics features include a total of 3384 features of the tumor and the peritumoral region in the arterial and portal venous phases. The wavelet features are extracted through discretization (fixed bin width 25 HU).

[0007] As a preferred technical solution of the present invention, the feature screening and standardization involves a three-step screening of the extracted radiomics features, including: retaining high-stability features with intragroup correlation coefficients > 0.80; screening features related to tumor recurrence (P < 0.05) through univariate Cox regression; selecting the final features using the LASSO Cox regression algorithm with 10-fold cross-validation; and standardizing the feature values ​​of the development cohort using Z-score, and standardizing the features of the validation cohort based on the mean and standard deviation of the development cohort.

[0008] As a preferred technical solution of the present invention, the model is constructed by multivariate Cox regression analysis to build a preoperative model and a postoperative model. The preoperative model integrates preoperative radiomics labels, the natural logarithm of serum AFP level and the number of tumors. The postoperative model further integrates postoperative pathological variables, including microvascular invasion and satellite nodules.

[0009] As a preferred technical solution of the present invention The formula for calculating the risk score of the preoperative model is as follows: Risk score = 0.13 × ln (serum AFP level) + 0.91 × number of tumors (0: solitary; 1: multiple) + 1.56 × radiomics label; The postoperative risk score calculation formula is: Risk score = 0.12 × ln (serum AFP level) + 0.96 × number of tumors (0: single; 1: multiple) + 1.42 × radiomics label + 0.68 × microvascular invasion (0: absent; 1: present) + 1.13 × satellite nodules (0: absent; 1: present).

[0010] As a preferred embodiment of the present invention, the model validation is performed by using an independent external validation cohort to evaluate the model’s discriminability, calibration and prediction error, wherein the discriminability is evaluated by the consistency index (C-index) and the area under the time dependence curve (AUC), and the prediction error is measured by the comprehensive Brier score.

[0011] As a preferred technical solution of the present invention, the risk stratification is based on X-tile analysis to determine the risk score cutoff value, and patients are divided into low-risk, intermediate-risk and high-risk recurrence risk levels, and the differences in recurrence-free survival among the layers are verified by the Kaplan-Meier method.

[0012] As a preferred embodiment of the present invention, the cutoff values ​​for risk stratification are: preoperative model total scores of 37 and 57, corresponding to low risk, intermediate risk, and high risk; and postoperative model total scores of 47 and 62, corresponding to low risk, intermediate risk, and high risk.

[0013] This invention also provides an application of a postoperative recurrence prediction model in clinical decision-making. It is applied to the above-mentioned early liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes, including: recommending patients to receive liver transplantation as a priority treatment based on the high-risk stratification results of the preoperative model; formulating an enhanced follow-up monitoring and adjuvant treatment plan based on the intermediate-high-risk stratification results of the postoperative model; and simplifying the postoperative follow-up frequency based on the low-risk stratification results.

[0014] (III) Beneficial Effects Compared with existing technologies, this invention provides a model and method for predicting early-stage liver cancer recurrence after surgery based on radiomics phenotypes, which has the following beneficial effects: This radiomics-based model and method for predicting postoperative recurrence in early-stage liver cancer integrates preoperative enhanced CT radiomics features (including first-order statistics, texture, and wavelet features) with clinical parameters (such as serum alpha-fetoprotein and tumor number) and postoperative pathological variables (such as microvascular invasion and satellite nodules). It constructs a dual-model approach for both preoperative and postoperative recurrence, demonstrating significantly superior predictive performance compared to existing staging systems and competing models (C-index of 0.77-0.82 and 0.78-0.88 in the development and validation cohorts, respectively, with a comprehensive Brier score ≤0.14). Accurate risk stratification was achieved through X-tile analysis (5-year recurrence rates of 17.9%, 61.2%, and 100.0% for low-, intermediate-, and high-risk patients, respectively). Calibration curves and decision analysis validated the model's reliability and clinical net benefit. This model can guide high-risk patients to prioritize liver transplantation, optimize postoperative follow-up intensity, and optimize adjuvant therapy strategies, providing a personalized recurrence prediction tool for early-stage liver cancer patients and effectively improving clinical decision-making efficiency and treatment outcomes. Attached Figure Description

[0015] Figure 1 The diagram shows the discrimination performance and prediction error of all prediction models and staging systems in this invention. Figure 2 The above figure shows the decision curves for 2-year (top) and 5-year (bottom) relapse-free survival obtained by applying existing models and staging systems in this invention. Figure 3This is a graph showing the cumulative tumor recurrence rate corresponding to the risk stratification (preoperative model and postoperative model) defined by the radiomics model in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0018] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0019] Please see Figure 1-3 , Example 1: A model and method for predicting postoperative recurrence of early-stage liver cancer based on radiomics phenotypes, including the following steps: data acquisition and preprocessing, feature screening and standardization, model construction, model validation, and risk stratification. In this embodiment, data acquisition and preprocessing involved obtaining preoperative enhanced CT imaging data and clinical information of patients with early-stage hepatocellular carcinoma (HCC) meeting the Milan criteria. Clinical information included serum alpha-fetoprotein (AFP) levels and the number of tumors. The tumor and the 4 mm peritumoral region were manually delineated using 3D Slicer software. The arterial and portal venous phase radiomics features of each lesion were extracted using the Pyradiomics software package, including first-order statistical features, texture features, and wavelet features. The radiomics features included a total of 3384 features of the tumor and the peritumoral region in the arterial and portal venous phases. The wavelet features were extracted through discretization (fixed bin width 25 HU).

[0020] It should be noted that data acquisition involved obtaining preoperative enhanced CT images and clinical data (including serum AFP levels and tumor number) from patients with early-stage hepatocellular carcinoma who met the Milan criteria. This ensured that the study subjects met the strict definition of early-stage hepatocellular carcinoma and enhanced the clinical applicability of the results. In the preprocessing stage, the tumor and the 4 mm peritumoral region were manually delineated using a 3D Slicer to accurately locate the lesion. Radiomic features (including first-order statistical features, texture features, and wavelet features) in the arterial and portal venous phases were extracted using the Pyradiomics software package. Among them, the wavelet features were discretized with a fixed bin width of 25 HU to reduce noise interference. Finally, 3384 features were extracted to comprehensively quantify tumor heterogeneity and imaging differences in surrounding tissues.

[0021] In this embodiment, feature screening and standardization involves a three-step screening process for the extracted radiomics features, including: retaining highly stable features with intragroup correlation coefficients > 0.80; screening features associated with tumor recurrence using univariate Cox regression (P < 0.05); selecting final features using the LASSO Cox regression algorithm with 10-fold cross-validation; standardizing the feature values ​​of the development cohort using Z-scores; and standardizing the features of the validation cohort based on the mean and standard deviation of the development cohort.

[0022] It should be noted that the feature selection adopts a three-step method to optimize feature quality. First, highly stable features with intragroup correlation coefficients > 0.80 are retained, while unreliable features caused by differences in image acquisition or delineation are excluded. Second, features significantly associated with recurrence are screened through univariate Cox regression (P < 0.05) to initially identify potential predictors. Finally, LASSO Cox regression combined with 10-fold cross-validation is used to further compress feature dimensions and avoid overfitting, ultimately generating radiomics labels. In the standardization stage, Z-score is used to normalize the features of the development cohort, and the cohort data is uniformly validated based on the mean and standard deviation of the development cohort to ensure the model's cross-institutional generalization ability.

[0023] In this embodiment, the model was constructed by multivariate Cox regression analysis to build a preoperative model and a postoperative model. The preoperative model integrated preoperative radiomics labels, the natural logarithm of serum AFP level and the number of tumors. The postoperative model further integrated postoperative pathological variables, including microvascular invasion and satellite nodules.

[0024] In this embodiment, The formula for calculating the risk score of the preoperative model is: Risk score = 0.13 × ln (serum AFP level) + 0.91 × number of tumors (0: solitary; 1: multiple) + 1.56 × radiomics label; The postoperative risk score calculation formula is: Risk score = 0.12 × ln (serum AFP level) + 0.96 × number of tumors (0: single; 1: multiple) + 1.42 × radiomics label + 0.68 × microvascular invasion (0: absent; 1: present) + 1.13 × satellite nodules (0: absent; 1: present).

[0025] It should be noted that the model was constructed based on multivariate Cox regression analysis, with preoperative and postoperative models established separately: the preoperative model integrated radiomics labels, the natural logarithmic value of serum AFP (reflecting tumor activity), and the number of tumors (single or multiple), and quantified the contribution of each factor to recurrence through weighted coefficients (e.g., AFP weight 0.13, tumor number weight 0.91); the postoperative model further incorporated postoperative pathological variables (microvascular invasion and satellite nodules), and strengthened the model's ability to capture pathological risks through weighted coefficients (e.g., microvascular invasion 0.68, satellite nodules 1.13), ultimately forming a comprehensive scoring formula to achieve multidimensional risk assessment.

[0026] In this embodiment, model validation is performed by using an independent external validation cohort to evaluate the model’s discrimination, calibration and prediction error. Discrimination is evaluated by the consistency index (C-index) and the area under the time dependence curve (AUC), and prediction error is measured by the comprehensive Brier score.

[0027] It should be noted that model validation used independent external cohorts (from different institutions) to evaluate performance: the model's ability to distinguish between relapsed and non-relapsed patients was measured by the consistency index (C-index ≥ 0.77) and time-dependent AUC (median ≥ 0.82); the Brier score (≤ 0.14) was used to quantify the prediction error, reflecting how close the model is to the actual relapse probability; and the calibration curve verified the consistency between the predicted risk and the actual Kaplan-Meier survival curve, ensuring the reliability of the model.

[0028] In this embodiment, risk stratification is based on X-tile analysis to determine the risk score cutoff value, and patients are divided into low-risk, intermediate-risk and high-risk recurrence risk levels. The differences in recurrence-free survival among the different levels are verified by the Kaplan-Meier method.

[0029] In this embodiment, the cutoff values ​​for risk stratification are: preoperative model total scores of 37 and 57, corresponding to low, intermediate, and high risk; and postoperative model total scores of 47 and 62, corresponding to low, intermediate, and high risk.

[0030] It should be noted that risk stratification is based on X-tile analysis to determine the cutoff values ​​of the preoperative model (37 and 57 points) and the postoperative model (47 and 62 points), classifying patients into three levels: low-risk, intermediate-risk, and high-risk. The Kaplan-Meier method is used to verify the differences in relapse-free survival among the levels (e.g., the 5-year relapse rate of the high-risk group is 100.0%), clarifying the clinical significance of the stratification. This stratification directly guides treatment decisions (e.g., prioritizing liver transplantation for high-risk patients), achieving closed-loop management from prediction to intervention.

[0031] Example 2: Application of a postoperative recurrence prediction model in clinical decision-making, which is applied to a model and method for predicting postoperative recurrence of early-stage liver cancer based on radiomics phenotypes, including: recommending patients to receive liver transplantation as a priority treatment based on the high-risk stratification results of the preoperative model; formulating enhanced follow-up monitoring and adjuvant treatment plans based on the intermediate-to-high-risk stratification results of the postoperative model; and simplifying the postoperative follow-up frequency based on the low-risk stratification results.

[0032] Beneficial effects: This radiomics-based model and method for predicting postoperative recurrence in early-stage liver cancer integrates preoperative enhanced CT radiomics features (including first-order statistics, texture, and wavelet features) with clinical parameters (such as serum alpha-fetoprotein and tumor number) and postoperative pathological variables (such as microvascular invasion and satellite nodules). It constructs a dual-model approach for both preoperative and postoperative recurrence, demonstrating significantly superior predictive performance compared to existing staging systems and competing models (C-index of 0.77-0.82 and 0.78-0.88 in the development and validation cohorts, respectively, with a comprehensive Brier score ≤0.14). Accurate risk stratification was achieved through X-tile analysis (5-year recurrence rates of 17.9%, 61.2%, and 100.0% for low-, intermediate-, and high-risk patients, respectively). Calibration curves and decision analysis validated the model's reliability and clinical net benefit. This model can guide high-risk patients to prioritize liver transplantation, optimize postoperative follow-up intensity, and optimize adjuvant therapy strategies, providing a personalized recurrence prediction tool for early-stage liver cancer patients and effectively improving clinical decision-making efficiency and treatment outcomes.

[0033] Working principle: By acquiring preoperative enhanced CT imaging data and clinical information (including serum AFP levels and tumor number) of patients with early-stage liver cancer, the tumor and peritumoral region were delineated using 3D Slicer. First-order statistical, textural, and wavelet features (a total of 3384 features) were extracted using Pyradiomics for the arterial and portal venous phases. A three-step screening process was then performed (retaining highly stable features, initial screening using univariate Cox regression, and LASSO...). Radiomics labels were constructed using Cox regression optimization. Combining clinical parameters and postoperative pathological variables (such as microvascular invasion and satellite nodules), multivariate Cox regression was used to establish preoperative models (integrating radiomics labels, natural logarithm of AFP, and tumor number) and postoperative models (further incorporating pathological variables). Performance was evaluated through independent external validation (C-index ≥ 0.77, overall Brier score ≤ 0.14). Risk score cutoff values ​​were determined based on X-tile analysis (e.g., 37 and 57 points in the preoperative model, and 47 and 62 points in the postoperative model), classifying patients into low, intermediate, and high recurrence risk levels (5-year recurrence rates of 17.9%, 61.2%, and 100.0%, respectively). The stratification effect was validated using the Kaplan-Meier method, guiding clinical decisions (e.g., prioritizing liver transplantation for high-risk patients and intensive follow-up for intermediate- and high-risk patients), thereby achieving individualized recurrence risk prediction and treatment optimization.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A model and method for predicting postoperative recurrence of early-stage liver cancer based on radiomics phenotypes, characterized in that, Includes the following steps: Data acquisition and preprocessing, feature selection and standardization, model building, model validation, and risk stratification.

2. The early-stage liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes according to claim 1, characterized in that: The data acquisition and preprocessing involved obtaining preoperative enhanced CT imaging data and clinical information of patients with early-stage hepatocellular carcinoma (HCC) meeting the Milan criteria. The clinical information included serum alpha-fetoprotein (AFP) levels and the number of tumors. The tumor and the 4mm peritumoral region were manually delineated using 3D Slicer software. The arterial and portal venous phase radiomics features of each lesion were extracted using the Pyradiomics software package, including first-order statistical features, texture features, and wavelet features. The radiomics features included a total of 3384 features of the tumor and the peritumoral region in the arterial and portal venous phases. The wavelet features were extracted through discretization (fixed bin width 25 HU).

3. The early-stage liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes according to claim 1, characterized in that: The feature selection and standardization process involves a three-step screening of the extracted radiomics features, including: retaining highly stable features with intragroup correlation coefficients > 0.80; screening features associated with tumor recurrence using univariate Cox regression (P < 0.05); selecting the final features using the LASSOCox regression algorithm with 10-fold cross-validation; standardizing the feature values ​​of the development cohort using Z-scores; and standardizing the features of the validation cohort based on the mean and standard deviation of the development cohort.

4. The early-stage liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes according to claim 1, characterized in that: The model was constructed using multivariate Cox regression analysis to create a preoperative and a postoperative model. The preoperative model integrated preoperative radiomics labels, the natural logarithm of serum AFP levels, and the number of tumors. The postoperative model further integrated postoperative pathological variables, including microvascular invasion and satellite nodules.

5. The early-stage liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes according to claim 4, characterized in that: The formula for calculating the risk score of the preoperative model is as follows: Risk score = 0.13 × ln(serum AFP level) + 0.91 × number of tumors (0: solitary; 1: multiple) + 1.56 × radiomics label; The formula for calculating the risk score of the postoperative model is as follows: Risk score = 0.12 × ln(serum AFP level) + 0.96 × number of tumors (0: single; 1: multiple) + 1.42 × radiomics label + 0.68 × microvascular invasion (0: absent; 1: present) + 1.13 × satellite nodules (0: absent; 1: present).

6. The early-stage liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes according to claim 1, characterized in that: The model validation assesses the model’s discrimination, calibration and prediction error using an independent external validation cohort. Discrimination is evaluated using the consistency index (C-index) and the area under the time dependence curve (AUC), and prediction error is measured using the combined Brier score.

7. The early-stage liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes according to claim 1, characterized in that: The risk stratification is based on X-tile analysis to determine the risk score cutoff value, classifying patients into low-risk, intermediate-risk, and high-risk recurrence risk levels, and verifying the differences in recurrence-free survival among the different strata using the Kaplan-Meier method.

8. The early-stage liver cancer postoperative recurrence prediction model and method based on radiomics phenotypes according to claim 7, characterized in that: The cutoff values ​​for the risk stratification are: preoperative model total scores of 37 and 57, corresponding to low, intermediate, and high risk; and postoperative model total scores of 47 and 62, corresponding to low, intermediate, and high risk.

9. The application of a postoperative recurrence prediction model in clinical decision-making, wherein it is applied to the postoperative recurrence prediction model and method for early-stage liver cancer based on radiomics phenotypes as described in any one of claims 1-8, characterized in that, include: Based on the high-risk stratification results of the preoperative model, liver transplantation is recommended as the first-line treatment for patients. Based on the results of the intermediate-to-high risk stratification of the postoperative model, an enhanced follow-up monitoring and adjuvant treatment plan was developed. Based on the low-risk stratification results, the postoperative follow-up frequency was simplified.

Images