The Value Of Radiomics In Evaluating Small-Hepatocellular Carcinoma And Dysplastic Nodules

Objective:1.Assess the diagnostic efficacy of radiomics models using unenhanced magnetic resonance scan and enhanced images to distinguish small-Hepatocellular Carcinoma（s-HCC）and Dysplastic Nodule（DN）.2.To evaluate the diagnostic performance of different classifiers in distinguishing s-HCC from DN on the same MRI sequence.Materials and Methods:1.Image information acquisition:68 patients who underwent liver resection and liver transplantation at the Second Affiliated Hospital of Dalian Medical University from May 2019 to December 2020 were collected.Unenhanced scans（T₁WI,T₂WI）and enhancement were obtained within one month before surgery（Arterial phase,portal phase,balance phase,delay phase）MR images.The enroll criteria were as follows:（1）All patients were pathologically confirmed as patients with liver cancer or liver cirrhosis;（2）Patients with complete imaging and clinical data;（3）Patients with nodule diameters less than 2 cm measured on MR images.The criteria of exclusion were as follows:（1）The MR image caused by ascites,respiratory movement,etc.,which caused serious artifacts in the lesion area and affected the diagnosis;（2）Nodules which diameter larger than 2cm measured on MR images;（3）The nodules which pathological gross specimen and the image corresponded poorly.Among them,59 cases were s-HCC and 32 cases were DN.The imaging data and clinical information of all enrolled patients were recorded in detail.2.Pathological information acquisition:The method of imaging and pathological comparison is adopted.After the operation,the gross pathological specimen is sliced into thickness of 1 cm according to the image scanning level,and the target nodule is found according to the Couinaud segmentation method,and then dehydrated,paraffin embedded,and hematoxylin Hematoxylin-Eosin（HE）staining,etc..Finally,undergoing corresponding histological and immunological pathological examinations to identify the lesion.The nodules with pathological characterization of s-HCC and DN were selected into this study.3.Image segmentation and feature extraction:The magnetic resonance images of all patients in the group were transferred to the medical standard-Darwin intelligent scientific research platform in DICOM format,and a junior radiologist with two years of professional experience manually performed the contours of the lesions in each phase Delineate layer by layer to form a volume of interest（VOI）,which is then checked and confirmed by a senior radiologist with more than ten years of professional experience.The radiomics platform extracts four categories of image features（first-order statistical features,shape-based features,texture features,and high-order statistical features）with high throughput.4.Dimensionality reduction and the establishment of radiomics model:Using Least Absolute Shrinkage and Selection Operator（LASSO）algorithm to reduce the dependence and redundancy of the extracted image features.According to the 4-foldcross-validation method adopted for screening features.In this study,every three groups are used as training groups,and the rest are used as test groups.On the one hand,it can avoid data disasters caused by over-fitting,and on the other hand,it can improve the stability of the models.The cases in the training group and the test group are randomly allocated to ensure that there is no statistical difference in the pathological type distribution of nodules.The optimal features selected from the MR plain scan（T₁WI,T₂WI）and enhanced images of each phase are used for supervised learning with Logistic Regression（LR）and Support Vector Machines（SVM）classifiers respectively.5.Statistical analysis:Calculate the AUC,recall and f1-score of the MR image learning results and test results of the LR and SVM classifiers in each phase.Compare the stability of the model through the f1-Score score value.Result:1.There were 91 lesions were acquired in 68 patients,including 59 s-HCC and 32DN.Divided into training group（68）and test group（23）at a ratio of 3:1.2.The AUC of the LR radiomics model based on the T₁WI is 0.96（95%CI:0.90-1）and 0.83（95%CI:0.61-1）in the training group and test group,respectively.The AUC of the SVM radiomics model based on the T₁WI was 0.97（95%CI:0.89-1）and 0.81（95%CI:0.57-1）in the training group and test group,respectively.3.The AUC of the LR radiomics model based on the T₂WI is 0.90（95%CI:0.80-1）and 0.87（95%CI:0.64-1）in the training group and test group,respectively.The AUC of the SVM radiomics model based on the T₂WI was 0.89（95%CI:0.78-1）and 0.88（95%CI:0.69-1）in the training group and test group,respectively.4.Based on the T₁WI enhanced arterial phase as the sequence with the highest efficiency in each phase of MR enhancement,the AUC of LR and SVM classifiers to identify s-HCC and DN are all greater than 0.96,and the f1-Score value is greater than0.70.Conclusion:1.When using MR plain scan and enhanced images to diagnose s-HCC and DN,enhanced scan arterial phase images have the highest diagnostic efficiency.2.When two classifiers of LR and SVM are used to distinguish s-HCC and DN,the diagnostic performance of the two classifiers is similar,but the radiomics prediction model constructed by SVM is more stable.3.The gray-level co-occurrence matrix and gray-level run-length matrix processed by wavelet transform are used to diagnose s-HCC and DN the best image characteristics.