Liver cancer ablation postoperative early recurrence identification method and storage medium

Abstract

The liver cancer ablation postoperative early recurrence identification method comprises the following steps: acquiring a computed tomography image of a target object, and inputting the computed tomography image into a liver cancer ablation postoperative early recurrence prediction model to obtain an early recurrence prediction result corresponding to the computed tomography image, the generation method of the early recurrence prediction model after the liver cancer ablation operation comprises the steps that a computed tomography image used for model training is obtained, features are extracted based on traditional radiomics and deep learning, and the features comprise multi-region and multi-phase features; the features are screened, a first prediction score and a second prediction score are calculated according to the screened features, the first prediction score is a prediction score of traditional radiomics, and the second prediction score is a prediction score of deep learning; and according to the first prediction score and the second prediction score, adding and fusing two types of features of traditional radiomics and deep learning, and constructing a liver cancer ablation postoperative early recurrence prediction model based on the fused features.

Description

Technical Field

[0001] The present application relates to the field of medical technology, and more specifically to a method and storage medium for identifying early recurrence of liver cancer after ablation surgery. Background Art

[0002] Primary liver cancer is the second most lethal malignant tumor in the world, and its incidence is on the rise. By 2025, the number of new cases worldwide will exceed 1 million. Hepatocellular carcinoma (HCC) accounts for about 90% of primary liver cancer, and its treatment effect is poor, becoming a major global health challenge. Ablation therapy has been proven to be a minimally invasive and effective first-line treatment for early HCC, achieving treatment effects consistent with surgical resection, but the 5-year recurrence rate is around 70%, which is higher than surgical resection, limiting its development. Compared with late recurrence, early recurrence (ER) leads to higher mortality and worse survival prognosis. Since HCC patients are mostly diagnosed by imaging without histopathological examination and lack risk factors related to recurrence such as MVI, predicting early recurrence after ablation has always been a difficult point in clinical research. Summary of the invention

[0003] The present application is proposed to solve the above problems. According to one aspect of the present application, a method for identifying early recurrence of liver cancer after ablation is provided, characterized in that the method comprises: A computer tomography image of a target object is obtained, and the computer tomography image is input into an early recurrence prediction model after liver cancer ablation to obtain an early recurrence prediction result corresponding to the computer tomography image, wherein the method for generating the early recurrence prediction model after liver cancer ablation includes: Obtaining computerized tomography images for model training, respectively extracting features based on traditional radiomics and extracting features based on deep learning, wherein the features include multi-region and multi-phase features; Screening the features, and calculating a first prediction score and a second prediction score respectively according to the screened features, wherein the first prediction score is a prediction score based on traditional imaging omics, and the second prediction score is a prediction score based on deep learning; According to the first prediction score and the second prediction score, the two types of features, traditional imaging genomics and deep learning, are added and fused, and a prediction model for early recurrence after liver cancer ablation is constructed based on the fusion features.

[0004] In one embodiment of the present application, obtaining an early recurrence prediction result corresponding to the computed tomography image includes: The computed tomography images are predicted, and the maximum Youden index in the validation set is used as the cutoff value of the prediction results. The patients' recurrence-free survival rate and overall survival rate are risk stratified, and the patients are divided into a low-risk group and a high-risk group.

[0005] In one embodiment of the present application, acquiring a computed tomography image for model training includes: The computer tomography image is in DICOM format; Delineating the computed tomography image, and delineating the three-dimensional region of interest of the tumor and the remaining liver in the arterial phase, portal phase and delayed phase of the computed tomography image in the software; The computed tomography images after delineation were subjected to peritumoral processing, and the peritumoral region of interest was processed using the Python morphological erosion and expansion algorithm, which automatically shrunk the boundary of each lesion inward by 5 mm and expanded outward by 3, 5, and 10 mm, respectively.

[0006] In one embodiment of the present application, the features are extracted based on traditional imaging omics and based on deep learning, and the features include multi-region and multi-phase features, including: Based on the traditional imaging omics feature extraction, the first feature was extracted, and the features of six regions and three phases were extracted. The three phases included arterial phase, venous phase and delayed phase. The six regions included: tumor, remaining liver, 5 mm within the tumor, 3 mm around the tumor, 5 mm around the tumor and 10 mm around the tumor. Six groups of features were extracted for each patient, including stage III tumor, stage III remaining liver, stage III intratumor 5 mm, stage III around the tumor 3 mm, stage III around the tumor 5 mm and stage III around the tumor 10 mm. Based on deep learning feature extraction, the pre-trained model was used to extract the second feature. The same six regions and three phases were extracted. Six groups of features were also extracted for each patient, including stage III tumor, stage III remaining liver, stage III intratumor 5mm, stage III peritumor 3mm, stage III peritumor 5mm and stage III peritumor 10mm. The multi-region, multi-phase characteristics include a first characteristic and a second characteristic.

[0007] In one embodiment of the present application, extracting features based on traditional imaging omics to extract the first feature includes: The characteristic differences between the arterial phase and the venous phase and the characteristic differences between the delayed phase and the venous phase of the six regions are calculated respectively, and the characteristic differences are screened. The first prediction score is calculated based on the screened characteristic differences and the screened features to improve the prediction ability of the prediction model.

[0008] In one embodiment of the present application, extracting features based on traditional imaging omics to extract the first feature includes: resampling the coronal and sagittal planes of the computed tomography image; extracting a first feature from the resampled computed tomography image; The first feature type extracted includes original image features and features extracted by applying multiple filters to the original image, and the filters used include square, square root, logarithmic, exponential, gradient operation, Gaussian-Laplacian filtering and wavelet filtering.

[0009] In one embodiment of the present application, the extracting features based on deep learning and using a pre-trained model to extract the second feature includes: Adapting features of an input layer of the pre-trained model to receive the computed tomography image as input; The network structure of the pre-trained model is optimized, and the last fully connected layer of the network structure is removed, so that the pre-trained model can output a second feature containing 512 feature vectors.

[0010] In one embodiment of the present application, the features are screened using a three-step screening method, comprising: Screen the six groups of features with intra-group and inter-group correlation coefficients greater than 0.8; The features are standardized and features with non-zero coefficients are selected by regularization; If the number of features screened by regularization exceeds one, recursive feature elimination is used with the decision tree as the kernel to determine the number of features at each stage.

[0011] In one embodiment of the present application, respectively calculating the first prediction score and the second prediction score according to the screened features includes: Calculating the mean of the screening feature, and calculating the standard deviation of the screening feature based on the mean; Using regularization to calculate the feature weight of the screening feature, and standardizing the feature weight by the standard deviation to obtain a standardized and adjusted feature weight; Using regularization to calculate the intercept of the screening feature, adjusting the intercept of the original model by the mean and standard deviation to obtain a standardized and adjusted intercept; A first prediction score and a second prediction score are calculated according to the standardized adjusted feature weights and the standardized adjusted intercept.

[0012] According to another aspect of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the processor executes the above-mentioned method for identifying early recurrence after liver cancer ablation.

[0013] The method for identifying early recurrence after liver cancer ablation of the present invention combines deep learning and traditional imaging genomics to construct a prediction model for early recurrence after liver cancer ablation. It adopts a multi-region and multi-phase feature extraction method to extract more comprehensive features and better prediction efficiency, so as to select more accurate treatment plans for patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0015] Figure 1 A schematic flow chart showing a method for identifying early recurrence of liver cancer after ablation according to an embodiment of the present application; Figure 2 A schematic flow chart showing a method for generating a prediction model for early recurrence after liver cancer ablation according to an embodiment of the present application; Figure 3 A multi-period and multi-regional image mapping diagram of radiomics according to an embodiment of the present application is shown; Figure 4 A comparison diagram of model performance according to an embodiment of the present application is shown; Figure 5 A generalization study diagram of a model for predicting early recurrence of liver cancer after ablation according to an embodiment of the present application is shown; Figure 6 The calibration curve and DCA curve diagram of the early recurrence prediction model after liver cancer ablation according to the embodiment of the present application are shown; Figure 7 A survival curve diagram of the early recurrence prediction model after liver cancer ablation according to the model of an embodiment of the present application is shown; Figure 8 A schematic diagram showing the prediction results of the early recurrence prediction model after liver cancer ablation according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present application more obvious, the example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application described in the present application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present application.

[0017] First, refer to Figure 1 The following describes a method for identifying early recurrence of liver cancer after ablation for implementing an embodiment of the present invention. Figure 1 FIG. 2 shows a schematic flow chart of a method for identifying early recurrence of liver cancer after ablation according to an embodiment of the present application. Figure 1 As shown, the method for identifying early recurrence of liver cancer after ablation according to an embodiment of the present application may include the following steps: In step S10, a computerized tomography image of the target object is acquired; In step S20, the computed tomography image is input into a model for predicting early recurrence after liver cancer ablation to obtain an early recurrence prediction result corresponding to the computed tomography image.

[0018] Figure 2 FIG. 1 is a schematic flow chart showing a method for generating a prediction model for early recurrence of liver cancer after ablation according to an embodiment of the present application. Figure 2 As shown, the method for generating a prediction model for early recurrence of liver cancer after ablation according to an embodiment of the present application may include the following steps: In step S100, a computerized tomography image for model training is obtained, and features are extracted based on traditional radiomics and based on deep learning, respectively, wherein the features include multi-region and multi-phase features; In step S200, the features are screened, and a first prediction score and a second prediction score are calculated according to the screened features, wherein the first prediction score is a prediction score based on traditional imaging omics, and the second prediction score is a prediction score based on deep learning; In step S300, according to the first prediction score and the second prediction score, the two types of features, traditional radiomics and deep learning, are added and fused, and a deep learning prediction model is constructed based on the fused features.

[0019] The method for identifying early recurrence after liver cancer ablation of the present invention combines deep learning and traditional imaging genomics to construct a prediction model for early recurrence after liver cancer ablation. It adopts a multi-region and multi-phase feature extraction method to extract more comprehensive features and better prediction efficiency, so as to select more accurate treatment plans for patients.

[0020] In an embodiment of the present application, a computer tomography image of the target object is acquired in step S10.

[0021] In one embodiment, before obtaining the computed tomography image of the target object, the target object is determined and the patient is selected. 222 early HCC patients who received radical ablation in the hospital were systematically included and randomly divided into a training group (155 cases) and a validation group (67 cases) in a ratio of 7:3. Ablation was performed under ultrasound guidance by a surgeon or interventional physician with 10 years of experience. All patients underwent contrast-enhanced computed tomography (CECT) within 1 month before ablation, and the clinical information of the patients was collected through the hospital case system, and the imaging data of the patients was collected through the PACS system.

[0022] Inclusion criteria were: (1) clinically diagnosed HCC, following the non-invasive criteria defined by the American Association for the Study of Liver Diseases based on specific imaging features; (2) single tumor diameter ≤5 cm, multiple tumors ≤3, each ≤3 cm in diameter; (3) patients were treated with ablation therapy (radiofrequency or microwave); (4) had good liver function, and Child-Pugh score was A. Exclusion criteria were: (1) previous HCC treatment history (such as liver resection, transarterial chemoembolization (TACE), targeted therapy, radiotherapy, etc.); (2) lack of CECT imaging data or no CECT examination 1 month before ablation; (3) postoperative ablation treatment was evaluated as failure or incomplete ablation; (4) follow-up was less than 2 years.

[0023] In one embodiment, before data collection began, all participating staff were trained in data extraction, the medical records of eligible patients were carefully reviewed, clinical and imaging data were extracted in a standardized form, and the most recent examination results within one week before the ablation treatment were provided to ensure the relevance and consistency of the data. Another reviewer randomly evaluated and verified 30% of the extracted data. Considering the different measurement units of laboratories in different participating centers, the values ​​were converted to be consistent, and extreme outliers were marked for review. These outliers were re-evaluated by the hospital's principal investigator or designated supervising physician to determine the validity of the data and exclude data entry errors.

[0024] Two physicians (Reader 1 and Reader 2) with 10 years of experience in abdominal CT reading independently evaluated the CECT images of each patient. During the evaluation process, the following six semantic features were mainly focused on: (1) tumor margin; (2) tumor capsule; (3) intratumoral blood vessels; (4) tumor growth pattern; (5) intratumoral necrosis; (6) peritumoral enhancement. When the patient had multiple lesions in the liver, the largest tumor was used as the evaluation object.

[0025] In an embodiment of the present application, a computed tomography image for model training is acquired in step S100, and features are extracted based on traditional imaging omics and based on deep learning, respectively, and the features include multi-region and multi-phase features.

[0026] like Figure 3 As shown, in one embodiment, a computed tomography image for model training is obtained, including: the computed tomography image is in DICOM format. The computed tomography image is delineated, and the three-dimensional regions of interest of the tumor and the remaining liver in the arterial phase, portal phase and delayed phase are delineated in the software. The delineated computed tomography image is processed around the tumor, and the Python morphological erosion and expansion algorithm is used to process the region of interest around the tumor, and the boundary of each lesion is automatically contracted inward by 5 mm and expanded outward by 3, 5, and 10 mm respectively. Figure 3 The VOI in the middle is the region of interest, and A, P, and D are the arterial phase, venous phase, and delayed phase, respectively.

[0027] The patient's CECT images were downloaded from the hospital's PACS system in DICOM format, and a physician A with 5 years of experience in reading abdominal CT films outlined the 3D volume of interest (VOI) of the tumor and the remaining liver in the arterial phase, portal phase, and delayed phase layer by layer in ITK-SNAP software. If the patient had multiple lesions, the largest lesion was selected as the representative lesion for segmentation. VOI was defined as the entire tumor or liver without peritumoral blood vessels or bile ducts. Physician A was unaware of the clinical data of the outlined cases during the entire process of outlining.

[0028] To extract peritumoral radiomic features, the Python morphological erosion and expansion algorithm was used to process the peritumoral VOI, automatically shrinking the boundary of each lesion by 5 mm inward and expanding it by 3, 5, and 10 mm outward, respectively. After automatic expansion, the part outside the liver parenchyma was manually removed. To ensure the repeatability of radiomic features, 50 patients were randomly selected after an interval of two weeks to be delineated again by physician A and another physician B with 10 years of experience in reading abdominal CT scans, and the Interclass Correlation Coefficient (ICC) was calculated.

[0029] In one embodiment, features are extracted based on traditional radiomics and features are extracted based on deep learning, and the features include features of multiple regions and multiple phases, including: based on traditional radiomics, features are extracted by extracting the first feature, and features of six regions (tumor, remaining liver, 5 mm within the tumor, 3, 5, and 10 mm around the tumor) and three phases (arterial phase A, venous phase P, and delayed phase D) are extracted. The three phases include arterial phase, venous phase, and delayed phase. The six regions include: tumor, remaining liver, 5 mm within the tumor, 3, 5, and 10 mm around the tumor mm, tumor periphery 5mm and tumor periphery 10mm. Six groups of features were extracted for each patient, including stage III tumor (A1, P1, D1), stage III remaining liver (A2, P2, D2), stage III intratumor 5mm (A-005mm, P-005mm, D-005mm), stage III tumor periphery 3mm (A-3mm, P-3mm, D-3mm), stage III tumor periphery 5mm (A-5mm, P-5mm, D-5mm) and stage III tumor periphery 10mm (A-10mm, P-10mm, D-10mm).

[0030] Based on deep learning feature extraction, the pre-trained model was used to extract the second feature. A total of the same six regions (tumor, remaining liver, 5mm within the tumor, 3, 5, and 10mm around the tumor) and three phases (arterial phase A, venous phase P, and delayed phase D) were extracted. The three phases included the arterial phase, venous phase, and delayed phase. The six regions were tumor, remaining liver, 5mm within the tumor, 3mm around the tumor, 5mm around the tumor, and 10mm around the tumor. Six groups of features were also extracted for each patient, including stage III tumor (A1, P1, D1), stage III remaining liver (A2, P2, D2), stage III intratumor 5mm (A-005mm, P-005mm, D-005mm), stage III tumor 3mm (A-3mm, P-3mm, D-3mm), stage III tumor 5mm stage III tumor 5mm (A-5mm, P-5mm, D-5mm), and stage III tumor 10mm (A-10mm, P-10mm, D-10mm). The multi-region and multi-phase characteristics include a first characteristic and a second characteristic.

[0031] In one embodiment, based on traditional radiomics feature extraction, a first feature is extracted, including: respectively calculating the feature difference between the arterial phase and the venous phase, and the feature difference between the delayed phase and the venous phase of the six regions, performing feature screening on the feature differences, and calculating a first prediction score based on the screened feature differences and the screened features to improve the predictive ability of the prediction model.

[0032] In one embodiment, extracting features based on traditional radiomics, extracting a first feature includes: resampling the coronal plane and the sagittal plane of the computed tomography image. Extracting the first feature from the resampled computed tomography image. The first feature type extracted includes original image features and features extracted by applying a plurality of filters to the original image, and the filters used include square, square root, logarithm, exponential, gradient operation, Gaussian-Laplace filtering, and wavelet filtering.

[0033] Radiomic features were extracted using the Pyradiomics (v3.1.0) package in Python (version 3.8.4 https: / / www.python.org / ). Before feature extraction, the following settings were made: First, the coronal and sagittal planes were resampled to 1 mm × 1 mm × 5 mm based on the original image layer thickness of the patient's original image of 5 mm. The image feature types included original image features and radiomic features extracted by applying multiple filters to the original image. The filters used included: Square, SquareRoot, Logarithm, Exponential, Gradient, Gaussian-Laplacian (LoG), and Wavelet. These settings allow further processing and analysis of the original image to extract more relevant feature information. The extracted feature categories include 19 first-order features, 17 3D shape features, 10 2D shape features, 24 gray-level co-occurrence matrix (GLCM) features, 16 gray-level local region matrix (GLSZM) features, 16 gray-level run-length matrix (GLRLM) features, 5 neighboring gray-level difference matrix (NGTDM) features, and 14 gray-level dependence matrix (GLDM) features.

[0034] Number of features extracted from each region: 1539 features were extracted from each region. After removing features with variables close to 0, 1515 features were retained from the tumor and remaining liver regions, and 1507 features were retained from other regions. A total of 45290 features were obtained for each patient.

[0035] In one embodiment, extracting features based on deep learning, using a pre-trained model to extract a second feature includes: performing adaptive features on an input layer of the pre-trained model to receive a computed tomography image as an input. Optimizing the network structure of the pre-trained model, removing the last fully connected layer of the network structure, so that the pre-trained model can output a second feature containing 512 feature vectors.

[0036] The settings before 3Dresnet-18 feature extraction are as follows: Pre-trained 3D ResNet-18 (Pytorch2.1.2) was used for deep feature extraction. Since the model is designed to process three-channel 3D input images, and CT images are single-channel 3D grayscale images, the input layer of 3D ResNet-18 was adaptively adjusted to receive single-channel images as input. In addition, the network structure was optimized and the last fully connected layer was removed, so that the model can output deep learning features containing 512 feature vectors.

[0037] In the deep learning feature extraction stage, the number of features extracted from each region was 512, and 9216 deep learning features were extracted from each patient, ensuring the consistency and comparability of the data.

[0038] In an embodiment of the present application, features are screened in step S200, and a first prediction score and a second prediction score are calculated respectively according to the screened features, wherein the first prediction score is a prediction score based on traditional imaging genomics, and the second prediction score is a prediction score based on deep learning.

[0039] In one embodiment, a three-step screening method is used to screen features, including: screening six groups of features with intra-group and inter-group correlation coefficients greater than 0.8. The features are Z-score standardized, and regularized LASSO is used to screen out features with non-zero coefficients. If the number of features screened by regularized LASSO exceeds 1, a decision tree (DT) is used as the kernel, and recursive feature elimination (RFE) is used to determine the number of features in each stage, and the number of features is the optimal number of features. In order to obtain more representative features, the features extracted from each region are subjected to the above feature screening process.

[0040] In one embodiment, the first prediction score and the second prediction score are calculated respectively according to the screened features, including: calculating the prediction score based on traditional radiomics by extracting the screened features based on traditional radiomics, and calculating the prediction score based on deep learning by extracting the screened features based on deep learning. The calculation process is as follows: calculating the mean of the screened features, and calculating the standard deviation of the screened features according to the mean: ; This formula represents the characteristic The average value of all data points, where Features The mean of . For the Of the data points The value of the feature, For all samples The mean of the features, is the total number of samples; ; This formula is used to measure the dispersion of eigenvalues ​​around the mean, where Features The standard deviation of .

[0041] Use regularization to calculate the feature weights of the screening features, standardize the feature weights by standard deviation, and get the standardized adjusted feature weights: ; This formula normalizes the weight of a feature by its standard deviation. is the feature weight calculated by LASSO, is the feature weight after standardization adjustment.

[0042] Use regularization to calculate the intercept of the screening feature, adjust the intercept of the original model by the mean and standard deviation, and get the standardized adjusted intercept: ; The intercept of the original model is adjusted by the mean and standard deviation, where is the number of selected features, is the intercept calculated by LASSO, is the re-standardized intercept.

[0043] Calculate the first prediction score based on the standardized adjusted feature weights and the standardized adjusted intercept and the second prediction score : ; The calculation method and steps are the same as above.

[0044] In an embodiment of the present application, in step S300, according to the first prediction score and the second prediction score, the two types of features of traditional imaging omics and deep learning are added and fused, and a prediction model for early recurrence after liver cancer ablation is constructed based on the fused features. The features of the traditional imaging omics class are the first features, and the features of the deep learning class are the second features. The two types of features of traditional imaging omics and deep learning are directly spliced ​​together to form a large feature vector, and a prediction model for early recurrence after liver cancer ablation is constructed based on the fused feature vector.

[0045] In the embodiment of the present application, in step S20, the computed tomography image is input into the early recurrence prediction model after liver cancer ablation to obtain an early recurrence prediction result corresponding to the computed tomography image. Figure 8 As shown, the data displayed are age, sex, and early recurrence.

[0046] In the recognition process, given new imaging data, the model first extracts traditional imaging omics and deep learning features and inputs these features into the early recurrence prediction model for liver cancer after ablation. The model makes predictions based on these input features and obtains classification results or risk assessments.

[0047] like Figure 7 As shown, in one embodiment, an early recurrence prediction result corresponding to a computed tomography image is obtained, including: predicting the computed tomography image, using the maximum Youden index in the validation set as the cutoff value of the prediction result, and using the optimal cutoff value to risk stratify the patient's recurrence-free survival rate and overall survival rate, and divide the patients into a low-risk group and a high-risk group.

[0048] In one embodiment, the features used for model construction include clinical-imaging features, traditional radiomics features and deep learning features. A clinical-imaging model (Cli) is constructed by clinical-imaging features, the first prediction score Rad-Score calculated by traditional radiomics features is used to construct a traditional radiomics model (Rad), and the second prediction score DL-Score calculated by deep learning features is used to construct a deep learning model (DLR).

[0049] The model building algorithms include six machine learning algorithms: support vector machine (SVM), logistic regression (LR), random forest (RF), K-nearest neighbor (KNN), light gradient boosting machine (LightGBM), and extreme gradient boosting (XGBoost). These six algorithms are applied to each set of data, and the algorithm with the highest AUC value in the validation set is selected as the best model. In order to improve the generalization ability of the model, 5-fold cross validation is used for hyperparameter tuning and model training. This process is only performed on the training set to avoid data leakage. In addition, models with a difference of more than 0.1 in the AUC of the training set and the validation set are excluded to prevent overfitting or underfitting.

[0050] First, we used the three features to construct three single models, namely the clinical-imaging model (Cli), the traditional radiomics model (Rad), and the deep learning model (DLR). Then, in order to identify potential better prediction models, we constructed four combined models based on the clinical-imaging features, traditional radiomics features and deep learning features using the feature fusion method. These four combined models include a combined model of clinical-imaging features and traditional radiomics features (CR), a combined model of clinical-imaging features and deep learning features (CDL), a combined model of traditional radiomics features and deep learning features (DLRR), and a comprehensive model combining all features (Cli + DLR + Rad, Combined).

[0051] To demonstrate the calibration ability and clinical applicability of the model, calibration curve and DCA analysis were performed on the best model in the training set and validation set. To evaluate whether the best model can effectively predict progression-free survival (PFS) and overall survival (OS) and perform risk stratification, we used the maximum Youden index in the validation set as the optimal cutoff value for prediction results and performed predictions in the training and validation cohorts, respectively. Patients were divided into low-risk and high-risk groups, and Kaplan-Meier survival curves were used to analyze 2-year PFS and 5-year OS.

[0052] Patients were followed up regularly after discharge, and local efficacy was evaluated in the first month after ablation. Serum AFP level, liver function and abdominal ultrasound were examined every 3 or 6 months thereafter. CECT, MRI or ultrasound angiography were performed to monitor recurrence if necessary. The starting point of this study was set as the time of ablation surgery; the termination event was early recurrence; complete ablation was defined as no enhancement of the tumor ablation lesion in the arterial phase after dynamic enhanced CT, MRI or ultrasound angiography, indicating complete tumor necrosis; early recurrence was defined as the appearance of new intrahepatic lesions or metastases within 2 years after surgery (lesions had typical imaging features of HCC or were confirmed by histopathology).

[0053] This study used R software (version 4.3.0) for statistical analysis. In the training cohort, missing parameters with more than 20% were excluded from the analysis, and missing data were imputed using a multiple imputation algorithm. To increase the interpretability of the model, continuous variables were converted into binary variables using the ROC curve threshold as the cutoff value. Categorical variables were expressed as frequencies and percentages, and chi-square tests or Fisher's exact tests were used. Variables with a P value less than 0.1 in the univariate analysis were included in the multivariate analysis for variable screening, such as Figure 4 As shown in the figure, the predictive performance of these models was evaluated by the area under the curve (AUC), and the predictive accuracy (ACC), positive predictive value (PPV) and negative predictive value (NPV) were also presented. Figure 4A is the training set and B is the validation set. The true positive rate is the number of true positive samples detected divided by the number of all true positive samples. The false positive rate is the number of false positive samples detected divided by the number of all true negative samples. Cli is the clinical-imaging model, Rad is the radiomics model, DLR is the deep learning model, CR is the combination model of clinical-imaging features and traditional radiomics features, CDL is the combination model of clinical-imaging features and deep learning features, DLRR is the combination model of traditional radiomics features and deep learning features, and Combined is the comprehensive model combining all features. The best performance is the model combining all features, with an AUC of 0.996 on the training set and 0.914 on the validation set. As shown in Table 1, the comparison between models was performed using the NRI and IDI tests. NRI (Net Reclassification Improvement) is the net reclassification index, which is used to evaluate the classification improvement of the new and old models at a specific threshold. IDI (Integrated Discrimination Improvement) is the integrated discrimination improvement index, which is used to evaluate the difference in overall predictive ability between the new and old models. P .value is the recurrence rate. The comprehensive model with added clinical imaging features is not better than the DLRR model. DLRR (Rad and DL combined model) is selected as the optimal model for the prediction of early recurrence after liver cancer ablation. Figure 5 As shown, the DLRR model is effective in the training set ( Figure 5 The AUC on the validation set ( Figure 5 The value of B in the figure is 0.910. The DLRR model has good generalization performance.

[0054] The comparison results of NRI and IDI between different models are shown in Table 1: .

[0055] like Figure 6 As shown in Figure 2, the calibration curve and DCA curve were used to evaluate the calibration and clinical applicability of the model. Figure 6 A and B in the DCA curve correspond to Figure 6 C and D in the figure. The calibration curve shows that the model has good fitting ability and stability. Figure 6 A is the training set, and B is the validation set. The DCA curve is the decision curve, which shows that the DLRR model has a good overall net benefit (clinical application value). ALL means all, and None means none. Figure 6 C is the training set and D is the validation set.

[0056] like Figure 7As shown, the survival curves were calculated using the Kaplan-Meier method and compared using the Log-rank test. Strata is the stratification of high risk and low risk. The Log-rank test is a nonparametric statistical test method used in survival analysis, and p is the recurrence rate value.

[0057] The KM curve shows that the DLRR model prediction value can accurately stratify the risk of recurrence-free survival and overall survival of HCC patients undergoing ablation therapy (dividing patients into low-risk and high-risk groups). The statistical significance level reported is bilateral and set at 0.05. The training set and validation set KM curves of the patient's 2-year PFS ( Figure 7 A and B in the figure), KM curves of the training set and validation set of patients' 5-year OS ( Figure 7 C, D in it).

[0058] The method for identifying early recurrence after liver cancer ablation of the present invention combines deep learning and traditional imaging genomics to construct a prediction model for early recurrence after liver cancer ablation. It adopts a multi-region and multi-phase feature extraction method to extract more comprehensive features and better prediction efficiency, so as to select more accurate treatment plans for patients.

[0059] In addition, the present application also provides a storage medium on which a computer program is stored, and when the computer program is run by a processor, the processor executes the method for identifying early recurrence after liver cancer ablation according to the embodiment of the present application described above. The storage medium may include, for example, a memory card of a smart phone, a storage component of a tablet computer, a hard disk of a personal computer, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disk read-only memory (CD-ROM), a USB memory, or any combination of the above storage media. The computer-readable storage medium may be any combination of one or more computer-readable storage media.

[0060] Although example embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above example embodiments are merely exemplary and are not intended to limit the scope of the present application to this. Those of ordinary skill in the art may make various changes and modifications therein without departing from the scope and spirit of the present application. All these changes and modifications are intended to be included within the scope of the present application as required by the appended claims.

[0061] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0062] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed.

[0063] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0064] Similarly, it should be understood that in order to streamline the present application and help understand one or more of the various inventive aspects, in the description of the exemplary embodiments of the present application, the various features of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, the method of the present application should not be interpreted as reflecting the following intention: the claimed application requires more features than the features clearly stated in each claim. More specifically, as reflected in the corresponding claims, the inventive point is that the corresponding technical problem can be solved with features less than all the features of a single disclosed embodiment. Therefore, the claims following the specific embodiment are hereby explicitly incorporated into the specific embodiment, wherein each claim itself serves as a separate embodiment of the present application.

[0065] Those skilled in the art will understand that, except for mutually exclusive features, all features disclosed in this specification (including the accompanying claims, abstract and drawings) and all processes or units of any method or device disclosed in this specification may be combined in any combination. Unless otherwise explicitly stated, each feature disclosed in this specification (including the accompanying claims, abstract and drawings) may be replaced by an alternative feature that provides the same, equivalent or similar purpose.

[0066] In addition, those skilled in the art will appreciate that, although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present application and form different embodiments. For example, in the claims, any one of the claimed embodiments may be used in any combination.

[0067] The various component embodiments of the present application can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or a digital signal processor (DSP) can be used in practice to implement some or all functions of some modules according to the embodiments of the present application. The present application can also be implemented as a program (e.g., a computer program and a computer program product) for executing part or all of the methods described herein. Such a program implementing the present application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0068] It should be noted that the above embodiments illustrate the present application rather than limit the present application, and those skilled in the art may design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets should not be constructed as a limitation to the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of multiple such elements. The present application can be implemented by means of hardware including several different elements and by means of appropriately programmed computers. In the enumerated several unit claims, these several can be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words can be interpreted as names.

[0069] The above is only a specific implementation or description of a specific implementation of the present application, and the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. The protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A method for identifying early recurrence of liver cancer after ablation, characterized in that: The method comprises: A computer tomography image of a target object is obtained, and the computer tomography image is input into an early recurrence prediction model after liver cancer ablation to obtain an early recurrence prediction result corresponding to the computer tomography image, wherein the method for generating the early recurrence prediction model after liver cancer ablation includes: Obtaining computerized tomography images for model training, respectively extracting features based on traditional radiomics and extracting features based on deep learning, wherein the features include multi-region and multi-phase features; Screening the features, and calculating a first prediction score and a second prediction score respectively according to the screened features, wherein the first prediction score is a prediction score based on traditional imaging omics, and the second prediction score is a prediction score based on deep learning; According to the first prediction score and the second prediction score, the two types of features, traditional imaging genomics and deep learning, are added and fused, and a prediction model for early recurrence after liver cancer ablation is constructed based on the fusion features.

2. The method according to claim 1, characterized in that The obtaining of an early recurrence prediction result corresponding to the computer tomography image comprises: The computed tomography images are predicted, and the maximum Youden index in the validation set is used as the cutoff value of the prediction results. The patients' recurrence-free survival rate and overall survival rate are risk stratified, and the patients are divided into a low-risk group and a high-risk group.

3. The method according to claim 1, characterized in that The step of acquiring a computer tomography image for model training comprises: The computer tomography image is in DICOM format; Delineating the computed tomography image, and delineating the three-dimensional region of interest of the tumor and the remaining liver in the arterial phase, portal phase and delayed phase of the computed tomography image in the software; The computed tomography images after delineation were subjected to peritumoral processing, and the peritumoral region of interest was processed using the Python morphological erosion and expansion algorithm, which automatically shrunk the boundary of each lesion inward by 5 mm and expanded outward by 3, 5, and 10 mm, respectively.

4. The method according to claim 1, characterized in that The features are extracted based on traditional imaging omics and deep learning, respectively, and the features include multi-region and multi-phase features, including: Based on the traditional imaging omics feature extraction, the first feature was extracted, and the features of six regions and three phases were extracted. The three phases included arterial phase, venous phase and delayed phase. The six regions included: tumor, remaining liver, 5 mm within the tumor, 3 mm around the tumor, 5 mm around the tumor and 10 mm around the tumor. Six groups of features were extracted for each patient, including stage III tumor, stage III remaining liver, stage III intratumor 5 mm, stage III around the tumor 3 mm, stage III around the tumor 5 mm and stage III around the tumor 10 mm. Based on deep learning feature extraction, the pre-trained model was used to extract the second feature, and the same six regions and three phases were extracted. Six sets of features were also extracted for each patient, including stage III tumor, stage III remaining liver, stage III intratumor 5mm, stage III peritumoral 3mm, stage III peritumoral 5mm, and stage III peritumoral 10mm. The multi-region, multi-phase characteristics include a first characteristic and a second characteristic.

5. The method according to claim 4, characterized in that The extracting of features based on traditional radiomics to extract the first feature includes: The characteristic differences between the arterial phase and the venous phase and the characteristic differences between the delayed phase and the venous phase of the six regions are calculated respectively, and the characteristic differences are screened. The first prediction score is calculated based on the screened characteristic differences and the screened features to improve the prediction ability of the prediction model.

6. The method according to claim 4, characterized in that The extracting of features based on traditional radiomics to extract the first feature includes: resampling the coronal and sagittal planes of the computed tomography image; extracting a first feature from the resampled computed tomography image; The first feature type extracted includes original image features and features extracted by applying multiple filters to the original image, and the filters used include square, square root, logarithmic, exponential, gradient operation, Gaussian-Laplacian filtering and wavelet filtering.

7. The method according to claim 4, characterized in that The method of extracting features based on deep learning and using a pre-trained model to extract a second feature includes: Adapting features of an input layer of the pre-trained model to receive the computed tomography image as input; The network structure of the pre-trained model is optimized, and the last fully connected layer of the network structure is removed, so that the pre-trained model can output a second feature containing 512 feature vectors.

8. The method according to claim 4, characterized in that The features were screened using a three-step screening process, including: Screen the six groups of features with intra-group and inter-group correlation coefficients greater than 0.8; The features are standardized and features with non-zero coefficients are selected by regularization; If the number of features screened by regularization exceeds one, recursive feature elimination is used with the decision tree as the kernel to determine the number of features at each stage.

9. The method according to claim 1, characterized in that The calculating the first prediction score and the second prediction score respectively according to the screened features comprises: Calculating the mean of the screening feature, and calculating the standard deviation of the screening feature based on the mean; Using regularization to calculate the feature weight of the screening feature, and standardizing the feature weight by the standard deviation to obtain a standardized and adjusted feature weight; Using regularization to calculate the intercept of the screening feature, adjusting the intercept of the original model by the mean and standard deviation to obtain a standardized and adjusted intercept; A first prediction score and a second prediction score are calculated according to the standardized adjusted feature weights and the standardized adjusted intercept.

10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is run, the method for identifying early recurrence of liver cancer after ablation according to any one of claims 1 to 9 is executed.

Images