Radiomics feature acquisition method based on gadoxetate disodium enhanced MRI for predicting histological grade of hepatocellular carcinoma

Abstract

A method for collecting imaging omics features based on gadoxetate disodium enhanced MRI for predicting the histological grade of hepatocellular carcinoma. There is a lack of an imaging omics feature extraction method based on gadoxetate disodium enhanced MRI. A method for collecting imaging omics features based on gadoxetate disodium enhanced MRI for predicting the histological grade of hepatocellular carcinoma, which sets inclusion and exclusion criteria for patients with HCC pathological information on MRI, and counts the total number of patients who can be included in the imaging omics feature collection, and the patients who meet the inclusion criteria undergo conventional Gd‑EOB‑DTPA enhanced MRI examinations; the imaging omics features of the obtained gadoxetate disodium enhanced MRI are extracted. As a data basis for predicting the histological grade of hepatocellular carcinoma, the present invention has the advantage of accurate feature extraction.

Description

Technical Field

[0001] This invention relates to a feature extraction method for predicting the histological grade of liver cancer, specifically a radiomics feature acquisition method based on gadoxetate disodium-enhanced MRI for predicting the histological grade of hepatocellular carcinoma. Background Technology

[0002] HCC is one of the most frequently diagnosed cancers worldwide and a leading cause of cancer-related death [1]. Surgical resection and liver transplantation are potential curative treatments for HCC patients [2], but postoperative recurrence remains common. Histological grade is one of the most important predictors of postoperative recurrence and prognosis in HCC patients [3-6]. Poorly differentiated HCC generally indicates poorer survival compared to moderately and well-differentiated HCC [7]. Oishi et al. do not recommend liver transplantation for patients with poorly differentiated HCC and a diameter >3 cm [8]. Okusaka et al. pointed out that patients with poorly differentiated HCC require a wider resection margin to obtain a larger safe margin, and they require more frequent postoperative follow-up [9]. Biopsy is the only preoperative method to obtain histopathological information of tumor tissue, but it is invasive and subject to sampling errors. Therefore, accurate assessment of the histological grade of HCC based on preoperative imaging is crucial for treatment planning, prognosis, and postoperative management.

[0003] Radiomics is an emerging research field dedicated to high-throughput extraction of large amounts of complex, statistically-based information from medical images

[10] . Radiomics has been used to assess a variety of biological behaviors in various types of tumors, including tumor recurrence, prognosis, and histological grade[11-13]. Texture features of CT, radiomics labels of noncontrast-enhanced MRI, and radiomics based on gadopentetate dimeglumine-enhanced MRI have been explored for preoperative prediction of HCC histological grade[14-16].

[0004] Gd-EOB-DTPA is a hepatocellular-specific contrast agent. Some retrospective studies have confirmed that Gd-EOB-DTPA-enhanced MRI plays an important role in predicting the histological grade of HCC [17-19]. Recently, Kim et al. reported that preoperative use of Gd-EOB-DTPA-enhanced MRI can effectively distinguish between early recurrence (≤2 years) and late recurrence (>2 years) after radical resection in patients with solitary HCC

[20] . However, to date, no one has attempted to use Gd-EOB-DTPA-enhanced MRI radiomics features to assess the histological grade of HCC.

[0005] Artificial neural networks (ANNs) are a commonly used machine learning method consisting of a set of highly interconnected processing units (neurons) connected together by weights

[21] . The network consists of an input layer, an output layer, and one or more hidden layers. Hidden units enable ANNs to encode nonlinear relationships between the input and output layers

[22] . The interactions between many factors related to clinical decision-making are complex and nonlinear [23, 24]. Traditional linear discriminant analysis struggles to distinguish between categories. For multidimensional radiomic features, ANNs can fit nonlinear mappings of arbitrary relationships with arbitrary precision, making them a potential tool for predicting clinical endpoint events [25, 26].

[0006] Therefore, the purpose of this invention is to design a feature extraction method for establishing a model for preoperative prediction of HCC histological grade based on Gd-EOB-DTPA enhanced MRI radiomics features, in order to make up for the problem that the feature extraction method is not perfect.

[0007] References:

[0008] [1]TorreLA, BrayF, SiegelRL, et al.Globalcancerstatistics, 2012[J].CACancerJClin, 2015, 65(2):87-108.

[0009] [2] KanwalF, BefelerA, ChariRS, etal. Potentially curative treatment in patients with hepatocellular cancer--results from the liver cancer research network [J]. Aliment Pharmacol Ther, 2012, 36 (3): 257-265.

[0010] [3]LauwersGY,TerrisB,BalisUJ,etal.Prognostichistologicindicatorsofcurativelyresectedhepatocellularcarcinomas:amulti-institutionalanalysisof425patientswithdefinitionofahistologicprognosticindex[J].AmJSurgPathol,2002,26(1):25-34.

[0011] [4]JonasS,BechsteinWO,SteinmullerT,etal.Vascularinvasionandhistopathologicgradingdetermineoutcomeafterlivertransplantationforhepatocellularcarcinomaincirrhosis[J].Hepatology,2001,33(5):1080-1086.

[0012] [5]ZhouL,RuiJA,WangSB,etal.Factorspredictiveforlong-termsurvivalofmalepatientswithhepatocellularcarcinomaaftercurativeresection[J].JSurgOncol,2007,95(4):298-303.

[0013] [6]ZhouL,RuiJA,YeDX,etal.Edmondson-SteinergradingincreasesthepredictiveefficiencyofTNMstagingforlong-termsurvivalofpatientswithhepatocellularcarcinomaaftercurativeresection[J].WorldJSurg,2008,32(8):1748-1756.

[0014] [7]CourtCM,Harlander-LockeMP,MarkovicD,etal.Determinationofhepatocellularcarcinomagradebyneedlebiopsyisunreliableforlivertransplantcandidateselection[J].LiverTranspl,2017,23(9):1123-1132.

[0015] [8]OishiK,ItamotoT,AmanoH,etal.Clinicopathologicfeaturesofpoorlydifferentiatedhepatocellularcarcinoma[J].JSurgOncol,2007,95(4):311-316.

[0016] [9]OkusakaT,OkadaS,UenoH,etal.Satellitelesionsinpatientswithsmallhepatocellularcarcinomawithreferencetoclinicopathologicfeatures[J].Cancer,2002,95(9):1931-1937.

[0017]

[10] LambinP,Rios-VelazquezE,LeijenaarR,etal.Radiomics:extractingmoreinformationfrommedicalimagesusingadvancedfeatureanalysis[J].EurJCancer,2012,48(4):441-446.

[0018]

[11] JiangY,ChenC,XieJ,etal.Radiomicssignatureofcomputedtomographyimagingforpredictionofsurvivalandchemotherapeuticbenefitsingastriccancer[J].EBioMedicine,2018,36:171-182.

[0019]

[12] ThawaniR,McLaneM,BeigN,etal.Radiomicsandradiogenomicsinlungcancer:Areviewfortheclinician[J].LungCancer,2018,115:34-41.

[0020]

[13] HuangX,ChengZ,HuangY,etal.CT-basedRadiomicsSignaturetoDiscriminateHigh-gradeFromLow-gradeColorectalAdenocarcinoma[J].AcadRadiol,2018,25(10):1285-1297.

[0021]

[14] OhJ,LeeJM,ParkJ,etal.HepatocellularCarcinoma:TextureAnalysisofPreoperativeComputedTomographyImagesCanProvideMarkersofTumorGradeandDisease-FreeSurvival[J].KoreanJRadiol,2019,20(4):569-579.

[0022]

[15] WuM,TanH,GaoF,etal.Predictingthegradeofhepatocellularcarcinomabasedonnon-contrast-enhancedMRIradiomicssignature[J].EurRadiol,2019,29(6):2802-2811.

[0023]

[16] ZhouW,ZhangL,WangK,etal.Malignancycharacterizationofhepatocellularcarcinomasbasedontextureanalysisofcontrast-enhancedMRimages[J].JMagnResonImaging,2017,45(5):1476-1484.

[0024]

[17] KimHY,ChoiJY,KimCW,etal.Gadoliniumethoxybenzyldiethylenetriaminepentaaceticacid-enhancedmagneticresonanceimagingpredictsthehistologicalgradeofhepatocellularcarcinomaonlyinpatientswithChild-PughclassAcirrhosis[J].LiverTranspl,2012,18(7):850-857.

[0025]

[18] AnC,ParkMS,JeonHM,etal.Predictionofthehistopathologicalgradeofhepatocellularcarcinomausingqualitativediffusion-weighted,dynamic,andhepatobiliaryphaseMRI[J].EurRadiol,2012,22(8):1701-1708.

[0026]

[19] KogitaS,ImaiY,OkadaM,etal.Gd-EOB-DTPA-enhancedmagneticresonanceimagesofhepatocellularcarcinoma:correlationwithhistologicalgradingandportalbloodflow[J].EurRadiol,2010,20(10):2405-2413.

[0027]

[20] KimS,ShinJ,KimDY,etal.RadiomicsonGadoxeticAcid-EnhancedMagneticResonanceImagingforPredictionofPostoperativeEarlyandLateRecurrenceofSingleHepatocellularCarcinoma[J].ClinCancerRes,2019,25(13):3847-3855.

[0028]

[21] CROSSSS,HARRISONRF,KENNEDYRL.INTRODUCTIONTONEURALNETWORKS[J].LANCET,1995,346(8982):1075-1079.

[0029]

[22] KempRA,MacAulayC,PalcicB.Openingtheblackbox:therelationshipbetweenneuralnetworksandlineardiscriminantfunctions[J].AnalCellPathol,1997,14(1):19-30.

[0030]

[23] PawlikTM,DelmanKA,VautheyJN,etal.Tumorsizepredictsvascularinvasionandhistologicgrade:Implicationsforselectionofsurgicaltreatmentforhepatocellularcarcinoma[J].LiverTranspl,2005,11(9):1086-1092.

[0031]

[25] PennyW,FrostD.Neuralnetworksinclinicalmedicine[J].MedDecisMaking,1996,16(4):386-398.

[0032]

[26] CucchettiA,VivarelliM,HeatonND,etal.ArtificialneuralnetworkissuperiortoMELDinpredictingmortalityofpatientswithend-stageliverdisease[J].Gut,2007,56(2):253-258.

[0033]

[29] GuanY,LiW,JiangZ,etal.Whole-LesionApparentDiffusionCoefficient-BasedEntropy-RelatedParametersforCharacterizingCervicalCancers:InitialFindings[J].AcadRadiol,2016,23(12):1559-1567.

[0034]

[30] AminMB.AJCCcancerstagingmanual[M].Eightedition / editor-in-chief,MahulB.Amin,MD,FCAP;editors,StephenB.Edge,MD,FACS[and16others];DonnaM.Gress,RHIT,CTR-Technicaleditor;LauraR.Meyer,CAPM-Managingeditor.2017.

[0035]

[32] TorlayL,Perrone-BertolottiM,ThomasE,etal.Machinelearning-XGBoostanalysisoflanguagenetworkstoclassifypatientswithepilepsy[J].BrainInform,2017,4(3):159-169.

[0036]

[33] DeLongER,DeLongDM,Clarke-PearsonDL.Comparingtheareasundertwoormorecorrelatedreceiveroperatingcharacteristiccurves:anonparametricapproach[J].Biometrics,1988,44(3):837-845. Summary of the Invention

[0037] The purpose of this invention is to address the lack of a feature extraction method for predicting the histological grade of hepatocellular carcinoma using gadoxetate disodium-enhanced MRI radiomics, and to propose a radiomics feature acquisition method for predicting the histological grade of hepatocellular carcinoma using gadoxetate disodium-enhanced MRI.

[0038] A method for acquiring radiomics features based on gadoxetine disodium-enhanced MRI for predicting the histological grade of hepatocellular carcinoma, the method being implemented through the following steps:

[0039] Step 1: Establish inclusion and exclusion criteria for patients with HCC pathological information from MRI, and count the total number of patients eligible for radiomics feature acquisition.

[0040] The inclusion criteria are as follows:

[0041] (1) There is an Edmondson-Steiner grade confirmed by postoperative pathology; the Edmondson-Steiner grade is the internationally used pathological grading of liver cancer.

[0042] (2) Complete gadoxetate disodium enhanced MRI images within 3 weeks prior to surgery;

[0043] The exclusion criteria are as follows:

[0044] (1) Receive any systemic or local anti-tumor treatment before the procedure, such as radiofrequency or microwave ablation, TACE or molecular targeted therapy;

[0045] (2) There is no complete preoperative clinical data;

[0046] (3) Poor MRI image quality due to artifacts; where poor image quality refers to images that score low by PSNR, SSIM or multi-scale structural similarity methods.

[0047] Step 2: Patients who met the inclusion criteria in Step 1 underwent routine Gd-EOB-DTPA enhanced MRI examination; specifically:

[0048] Each patient was required to fast for 6 hours and undergo breathing training before the examination;

[0049] MRI was performed using a whole-body 3.0T scanner with a 32-channel dStream body coil;

[0050] The scan range extends from the right diaphragm to the level of the renal hilum;

[0051] Using T1-weighted 3D turbine field echo (TFE) sequences and multi-echo Dixon fat-water separation technology;

[0052] Images of the arterial phase, portal venous phase, equilibrium phase, and hepatobiliary phase were obtained by injecting d-EOB-DTPA injection solution and normal saline.

[0053] MRI parameters included repetition time: 3.7ms, echo time: 1.32ms, 2.4ms, matrix size: 268×236, slice thickness: 5mm, reconstruction slice thickness: 2.5mm, field of view: 400mm×352mm, flip angle: 10 degrees, and number of excitations: 1.

[0054] Step 3: The obtained gadoxetine disodium enhanced MRI images are divided into two parts according to the total number of statistics. One part is used as the training group to construct a histological grade model for predicting hepatocellular carcinoma, and the other part is used as the validation group to verify the predictive effect of the histological grade model for predicting hepatocellular carcinoma.

[0055] Step 4: Extract radiomics features.

[0056] Preferably, the pathological information described in step one is obtained from an electronic pathology system, and the pathological information includes: gender, age, serum AFP level, serum carcinoembryonic antigen level, serum carbohydrate antigen 19-9 level, ALT, AST, TNM stage, and ES grade.

[0057] Preferably, the process of extracting radiomics features in step four is as follows:

[0058] Step 41: Obtain the MRI radiomics features enhanced by disodium gadoxetate;

[0059] Gadoxetine disodium enhanced MRI images were obtained, and the Region of Interest (ROI) was delineated layer by layer along the tumor margin on the arterial and hepatobiliary phase MRI images, respectively. The ROI included necrotic or cystic areas within the tumor, while avoiding the surrounding liver parenchyma, blood vessels, and adjacent intra-abdominal adipose tissue.

[0060] Multiple ROIs in each phase were reconstructed to obtain the VOI of the tumor; 123 radiomics features could be extracted from each VOI, including 19 grayscale histogram features: mean, standard deviation, skewness, kurtosis, entropy, max frequency, minimum, maximum, histogram width, percentile 5%, percentile 10%, percentile 25%, percentile 50%, percentile 75%, percentile 90%, mode, AUClow, s-sDlowest, and s-sDavdistribution. Witdh; 8 shape features: volume, surface area, sphericity, spherical disproportion, surface volume ratio, compactness1, compactness2, volume mediameter; 24×4 texture features, each texture feature originating from four directions: 0°, 45°, 90°, and 135°, of which 16×4 texture features come from the gray-level co-occurrence matrix; 8×4 texture features come from the gray-level run length matrix; the average value of each texture feature in the four directions is selected as the final value of the texture feature, thus 24 texture features are obtained in each phase, and a total of 2×(19+8+24)=102 imagemic features are obtained in two phases;

[0061] Step 4.2: Use ICC to assess the inter-observer consistency of the extracted MRI radiomics features for each group; radiomics features with an ICC value below 0.75 are considered to have poor stability and are removed; then, the average of the obtained consistency values ​​is used as the final result; that is:

[0062] Features with poor stability were removed from the arterial phase MRI radiomics dataset by inter-observer repeatability calculations, leaving 47 arterial phase MRI radiomics features with ICC values ​​ranging from 0.840 to 1.000. Among the removed features were: AUClow (ICC value 0.638), sphericity (ICC value 0.290), compactness2 (ICC value 0.513), and clustershade (ICC value 0.655).

[0063] Meanwhile, features with poor stability were removed from the hepatobiliary MRI radiomics dataset through inter-observer repeatability calculations, leaving 50 hepatobiliary MRI radiomics features with ICC values ​​ranging from 0.757 to 1.000. Among the features removed, sphericity with an ICC value of 0.428 was included.

[0064] The beneficial effects of this invention are as follows:

[0065] This invention is used to construct a model that distinguishes between high-grade and low-grade HCC. Specifically, it provides a data foundation for distinguishing between high-grade and low-grade HCC, and makes up for the technical deficiencies in feature selection used in the model construction. Specifically, it is a comprehensive model composed of clinical parameters and Gd-EOB-DTPA-enhanced MRI radiomics features.

[0066] This invention establishes a model for preoperative prediction of HCC histology based on MRI radiomics features (based on arterial phase, hepatobiliary phase, arterial phase, and hepatobiliary phase) and clinical parameters.

[0067] The method of this invention selects radiomics features based on the repeatability between and among observers. It establishes a repeatability test, including intra-observer and inter-observer ICC calculations, which can significantly improve the functional clustering reproducibility of radiomic features. It can screen out important features that are related to distinguishing between high-level and low-level HCC and has the advantage of accurate feature selection, thus providing a foundation for building predictive models. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation Specific implementation method one:

[0070] This embodiment describes a method for acquiring radiomics features based on gadoxetate disodium-enhanced MRI for predicting the histological grade of hepatocellular carcinoma, such as... Figure 1 As shown, the method is implemented through the following steps:

[0071] Step 1: Establish inclusion and exclusion criteria for patients with HCC pathological information from MRI, and count the total number of patients eligible for radiomics feature acquisition.

[0072] The inclusion criteria are as follows:

[0073] (1) There is an Edmondson-Steiner (ES) grade confirmed by postoperative pathology; the Edmondson-Steiner grade is the internationally used pathological grading of liver cancer.

[0074] (2) Complete gadoxetine disodium-enhanced Gd-EOB-DTPA-enhanced MRI images within 3 weeks prior to surgery;

[0075] The exclusion criteria are as follows:

[0076] (1) Any systemic or local anti-tumor treatment received before the procedure, such as radiofrequency or microwave ablation, TACE or molecular targeted therapy (n=63);

[0077] (2) No complete preoperative clinical data (n=21);

[0078] (3) Poor MRI image quality due to artifacts (n=9); where poor image quality refers to images that score low by PSNR, SSIM or multi-scale structural similarity methods.

[0079] Ultimately, a total of 122 HCC patients were included, including 105 males and 17 females, with a mean age of 58 years (range: 35–84 years). Patients were randomly assigned to the training group (n = 85, 70%) and the test group (n = 37, 30%) in a 7:3 ratio.

[0080] Step 2: Patients who met the inclusion criteria in Step 1 underwent routine Gd-EOB-DTPA enhanced MRI examination; specifically:

[0081] Each patient was required to fast for 6 hours and undergo breathing training before the examination;

[0082] MRI was performed using a whole-body 3.0T scanner with a 32-channel dStream body coil (Ingenia 3.0T, Philips Healthcare, Best, The Netherlands);

[0083] The scan range extends from the right diaphragm to the level of the renal hilum;

[0084] Using T1-weighted 3D turbine field echo (TFE) sequences and multi-echo Dixon fat-water separation technology;

[0085] Gd-EOB-DTPA injection solution and normal saline were injected. 0.025 mmol / kg body weight (0.1 mL / kg) of Gd-EOB-DTPA injection solution (Primovist, Bayer Schering AG, Berlin, Germany) was rapidly injected via a peripheral vein at a rate of 1 mL / s. Then, 10 mL of normal saline was injected at a rate of 1 mL / s using an automated power injector (Medrad Spectris Solaris EPMRI injector System; One Medrad Drive Indianola, PA, US). Plain scan images were obtained before contrast agent injection, and images of the arterial phase (3 phases in total: with a fixed 18-s delay scan, each phase lasting 8 seconds), portal venous phase (60 seconds), equilibrium phase (180 seconds), and hepatobiliary phase (15 min) were obtained after contrast agent injection.

[0086] MRI parameters included repetition time (TR): 3.7ms, echo time (TE): 1.32ms, 2.4ms, matrix size: 268×236, slice thickness: 5mm, reconstruction slice thickness: 2.5mm, field of view (FOV): 400mm×352mm, flip angle: 10 degrees, and number of excitations: 1.

[0087] Step 3: The obtained gadoxetine disodium enhanced MRI images are divided into two parts according to the total number of statistics. One part is used as the training group to construct a histological grade model for predicting hepatocellular carcinoma, and the other part is used as the validation group to verify the predictive effect of the histological grade model for predicting hepatocellular carcinoma.

[0088] Step 4: Extract radiomics features. Specific implementation method two:

[0090] Unlike Specific Implementation Method 1, this implementation method uses gadoxetate disodium-enhanced MRI to acquire radiomics features for predicting the histological grade of hepatocellular carcinoma.

[0091] The pathological information described in step one is obtained from the electronic pathology system. The pathological information includes: gender, age, serum AFP level (>10ng / mL or ≤10ng / mL), serum carcinoembryonic antigen (CEA) level (>5ng / mL or ≤5ng / mL), serum carbohydrate antigen 19-9 (CA19-9) level (>39U / mL or ≤39U / mL), ALT (U / L), AST (U / L), TNM stage, and ES grade. Specific implementation method three:

[0093] Unlike Specific Implementation Method 2, this implementation method uses gadoxetate disodium-enhanced MRI to acquire radiomics features for predicting the histological grade of hepatocellular carcinoma.

[0094] The process of extracting radiomics features as described in step four is as follows:

[0095] Step 41: Obtain the MRI radiomics features enhanced by disodium gadoxetate;

[0096] Gadoxetine disodium-enhanced Gd-EOB-DTPA-enhanced MRI images were acquired. Arterial (phase 2) and hepatobiliary phase images with a slice thickness of 2.5 mm were obtained from the Picture Archiving and Communication System (PACS) for all HCC patients, all conforming to the Digital Imaging and Imaging and Communications in Medicine (DICOM) protocol.

[0097] Regions of interest (ROIs) were manually delineated layer by layer along the tumor margin on arterial and hepatobiliary MRI images, including necrotic or cystic areas within the tumor, avoiding surrounding liver parenchyma, blood vessels, and adjacent intra-abdominal adipose tissue. Multiple ROIs in each phase were reconstructed to obtain the voicing of tumor (VOI). From each VOI, 123 radiomics features were extracted, including 19 grayscale histogram features: mean, standard deviation, skewness, kurtosis, entropy, max frequency, minimum, maximum, histogramwidth, percentile 5%, percentile 10%, percentile 25%, percentile 50%, percentile 75%, percentile 90%, mode, AUClow, s-sDlowest, s-sDavdistributionwithdh; and 8 shape features: volume, surface area, sphericity, spherical disproportion, surface volume ratio, compactness1, compactness2. ,volumediameter; 24×4 texture features (each texture feature originates from four directions: 0°, 45°, 90° and 135°), of which 16×4 texture features come from the gray-level co-occurrence matrix (correlation, entropy(H), IDMN, IDN, dissimilarity, energy, homogeneity1, homogeneity2, contrast, variation, inverse variation, maxprobability, autocorrelation, clusterprominence, clustershade, clustertendency); 8×4 texture features come from the gray-level run length matrix (GLN, HGLRE, LRHGLE, LRLGLE, LGLRE, RLN, SRHGLE, SRLGLE). The average value of each texture feature in the four directions is selected as the final value of the texture feature

[29] . Therefore, 24 texture features can be obtained in each period. A total of 2×(19+8+24)=102 image omics features were obtained in the two periods.

[0098] The Shapiro-Wilk test and Levene test were used for normality and homogeneity of variance tests, respectively. The permutation test was used to compare the Gd-EOBDTPA-enhanced MRI radiomics features and preoperative clinical parameters of low-grade and high-grade tumors, and the decision tree was used for further parameter selection

[32] . The selected parameters were used to construct the ANN and logistic regression models for predicting the histological grade of HCC. The probability of the event of interest was obtained by the formula of the logistic regression model: p = 1 / (1 + e-logit(p)), logit(p) = b0 + b1x1 + b2x2 + b3x3 + b4x4 + ... bkxk. The ANN is a three-layer feedforward neural network with input nodes (clinical parameters and radiomics features), a hidden layer with 23 neurons and an output neuron (high-grade HCC or low-grade HCC). The predictive efficacy of different models was evaluated by ROC curves, and the AUC, sensitivity and specificity of the models were recorded. The Delong test

[33] was used to compare the differences in AUC between the predictive models. Decision curve analysis was used to evaluate the net benefit of different predictive models. Python 3.7 (Amsterdam, Netherlands) is used to build ANN and logistic regression models. R software (R Foundation for Statistical Computing, version 3.4.1; https: / / www.r-project.org / ) is used for other statistical analyses. A two-tailed p-value < 0.05 is considered statistically significant.

[0099] Table 1 presents the clinicopathological information of 122 HCC patients. The 122 HCCs were divided into 79 low-grade tumors (14 E-SI tumors and 65 E-SII tumors) and 43 high-grade tumors (41 E-SIII tumors and 2 E-SIV tumors);

[0100] Table 1. Clinicopathological information of HCC patients

[0101]

[0102]

[0103] Step 4.2: Use ICC to assess the inter-observer consistency of the extracted MRI radiomics features for each group; radiomics features with an ICC value below 0.75 are considered to have poor stability and are removed; then, the average of the obtained consistency values ​​is used as the final result; that is:

[0104] Features with poor stability were removed from the arterial phase MRI radiomics dataset by inter-observer repeatability calculations, leaving 47 arterial phase MRI radiomics features with ICC values ​​ranging from 0.840 to 1.000. Among the removed features were: AUClow (ICC value 0.638), sphericity (ICC value 0.290), compactness2 (ICC value 0.513), and clustershade (ICC value 0.655).

[0105] Meanwhile, features with poor stability were removed from the hepatobiliary MRI radiomics dataset through inter-observer repeatability calculations, leaving 50 hepatobiliary MRI radiomics features with ICC values ​​ranging from 0.757 to 1.000. Among the features removed, sphericity with an ICC value of 0.428 was removed. The ICC values ​​of all radiomics features are shown in Table 2.

[0106] Table 2. ICC values ​​of Gd-EOB-DTPA enhanced MRI radiomics features in the arterial and hepatobiliary phases.

[0107]

[0108]

[0109]

[0110]

[0111] Note: Gd-EOB-DTPA, disodium gadoxetate; ICC, intragroup correlation coefficient; *, ICC < 0.750.

Claims

1. A method for acquiring imaging features based on gadoxetate disodium enhanced MRI for predicting the histological grade of hepatocellular carcinoma, characterized by: The method is implemented by the following steps: Step 1: Set inclusion and exclusion criteria for patients with HCC pathological information on MRI, and count the total number of patients who can be included in the imaging feature collection. The inclusion criteria are: (1) The Edmondson-Steiner grade is confirmed by postoperative pathology; the Edmondson-Steiner grade is the international pathological grade for liver cancer; (2) Complete gadoxetate disodium-enhanced MRI images within 3 weeks before surgery; The exclusion criteria are: (1) Any systemic or local anti-tumor treatment before surgery, including radiofrequency or microwave ablation, TACE or molecular targeted therapy; (2) Lack of complete preoperative clinical data; (3) Poor MRI image quality due to artifacts; poor image quality refers to images with low scores assessed by PSNR, structural similarity SSIM or multi-scale structural similarity methods; Step 2: Patients who meet the criteria in step 1 will receive routine gadoxetate disodium enhanced MRI examination; specifically: Each patient fasted for 6 hours before the examination and received breathing training; MRI examinations were performed using a whole-body 3.0T scanner with a 32-channel dStream body coil; The scan range was from the right diaphragm to the level of the renal hilum; A T1-weighted 3D turbo-field echo sequence and multi-echo Dixon fat-water separation technique were used; After the injection of d-EOB-DTPA injection and saline, images were obtained during the arterial phase, portal phase, equilibrium phase, and hepatobiliary phase; MRI parameters included repetition time: 3.7 ms, echo time: 1.32 ms, 2.4 ms, matrix size: 268 × 236, slice thickness: 5 mm, reconstruction slice thickness: 2.5 mm, field of view: 400 mm × 352 mm, flip angle: 10 degrees, number of excitations: 1; Step 3, the obtained gadoxetate disodium enhanced MRI is divided into two parts according to the total number of statistics, one part is used as a training group to build a model for predicting the histological grade of hepatocellular carcinoma, and the other part is used as a validation group to verify the prediction effect of predicting the histological grade of hepatocellular carcinoma; Step 4: Extract radiomics features; The process of extracting radiomics features described in step 4 is as follows: Step 41: Obtaining the radiomic features of gadoxetate disodium-enhanced MRI; Obtain gadoxetate disodium-enhanced MRI images, and delineate ROI layer by layer along the edge of the tumor on the arterial phase and hepatobiliary phase MRI images. The ROI includes the necrotic or cystic areas within the tumor and avoids the liver parenchyma tissue, blood vessels, and adjacent abdominal fat tissue surrounding the tumor. Multiple ROIs of each phase were reconstructed to obtain the VOI of the tumor; 123 radiomics features were extracted from each VOI, including 19 grayscale histogram features: mean, standard deviation, skewness, kurtosis, entropy, max frequency, minimum, maximum, histogram width, percentile 5%, percentile 10%, percentile 25%, percentile 50%, percentile 75%, percentile 90%, mode, AUC low, s-sD lowest, s-sD av distribution witdh; 8 shape features: volume, surface area, sphericity, spherical disproportion, surface volume, compactness 1, compactness 2, volume mediameter; 24×4 texture features, each texture feature comes from the four directions of 0°, 45°, 90° and 135°, of which 16×4 texture features come from the gray level co-occurrence matrix; 8×4 texture features come from the gray level run length matrix; the average value of each texture feature in the four directions is selected as the final value of the texture feature, so 24 texture features are obtained in each period, and a total of 2×(19+8+24)=102 imaging features are obtained in two periods; Step 42: Use ICC to evaluate the inter-observer consistency of the extracted MRI imaging features of each group; imaging features with ICC values ​​lower than 0.75 are considered to have poor stability and are removed; then the average value of each consistency obtained is used as the final result; that is: The features with poor stability in the arterial phase MRI radiomics dataset were removed by reproducibility calculation between observers, and 47 arterial phase MRI radiomics features remained, and the ICC values ​​of the remaining features ranged from 0.840 to 1.000; the removed features included: AUClow with an ICC value of 0.638, sphericity with an ICC value of 0.290, compactness2 with an ICC value of 0.513, and clustershade with an ICC value of 0.655; At the same time, the features with poor stability were removed from the hepatobiliary phase MRI imaging dataset through inter-observer repeatability calculation, and 50 hepatobiliary phase MRI imaging features remained, and the ICC value of the features ranged from 0.757 to 1.000; among them, the removed features included: sphericity with an ICC value of 0.

428.

2. The imaging feature acquisition method based on gadoxetate disodium enhanced MRI for predicting the histological grade of hepatocellular carcinoma according to claim 1, characterized in that: The pathological information described in step 1 is obtained from the electronic pathology system, and the pathological information includes: gender, age, serum AFP level, serum carcinoembryonic antigen level, serum carbohydrate antigen 19-9 level, ALT, AST, TNM stage, and ES grade.

Images