Radiomic features to predict microvascular invasion in hepatocellular carcinoma based on conventional MRI: preliminary findings / 中华放射学杂志
Chinese Journal of Radiology
; (12): 292-298, 2019.
Article
in Zh
| WPRIM
| ID: wpr-754923
Responsible library:
WPRO
ABSTRACT
Objective To identify the preoperative MRI findings for predicting microvascular invasion (MVI) using texture analysis (TA) on multiple MRI sequences. Methods Two hundred and fifty patients with HCC pathologically confirmed by surgery in Zhongshan Hospital from October 2015 to October 2016 were analyzed retrospectively. All patients underwent conventional MRI plain scan and dynamic contrast?enhanced examination within 2 weeks before operation. According to the ratio of 1∶1, the patients were divided into a training set (125 cases) and a test set (125 cases).The training set was used to establish a classifier to predict MVI of HCCs via the TA, and the test set was used to evaluate the performance of the classifier. An image analysis was performed using an in?house software contained a set of 2 415 features which were generated from all conventional axial sequences, including the T2WI, DWI, ADC map, and dynamic enhancement images.. A four?fold cross validation (FFCV) and sequential forward floating feature selection strategy (SFFS) were employed to select an optimal subset of features and a linear discriminant analysis (LDA) was employed to establish a classifier. The clinical laboratory examination, morphologic characteristics and quantitative analysis of conventional MR were used to compare the performance of predicting MVI with the classifier. A Chi?squared test or Fisher exact probablities test were used for categorical variables, and independent t test or Mann?Whitney U test were used for used for continuous variables. Factors with a P value less than 0.05 at univariate analyses were entered into the multivariate model to identify independent predictors. The Hosmer?Lemeshow test was performed to explain the goodness of fit of the multivariate logistic model. A receiver operating characteristic curve analysis was performed to evaluate the diagnostic performance. Results The classifier set up by the training set consists of 13 texture features. When conventional MRI texture features of test set were used to judge whether there was MVI or not, the AUC of all texture features of arterial phase (AP) was the highest (0.506 3). Univariate regression analysis showed that there were significant differences in pathological grade (P=0.026), AFP level (P=0.033), lesion edge shape (P=0.038), AP enhancement (P=0.038), and AP peritumoral enhancement (P=0.008). Multivariate binary logistic regression analysis showed that peritumoral enhancement and texture classifier assessed MVI with P values of 0.005 and 0.001,which were independent risk factors for MVI. The significance level of Hosmer Lemeshow test was 0.796, indicating the goodness of fit of acceptable models. The AUCs of single variable, combined variable (including of AFP level, irregular tumor margin, enhancement intensity in AP and peritumoral enhancement in AP) and texture classifier for MVI were 0.588 to 0.627, 0.798 and 0.733, respectively. When compared the AUC of the combination features (including of AFP level, irregular tumor margin, enhancement intensity in AP and peritumoral enhancement in AP) with the classifier to identify MVI of HCC in the test set, no significant difference was found(P=0.108 6). However, although the sensitivity of them were same as 70.73%, the specificity of the combination features was mildly higher than that of classifier (82.14% vs. 78.57%). Conclusions Combination features of AFP level, tumor margin, enhancement intensity in AP and peritumoral enhancement in AP can be used to predict MVI of HCCs. It is a new method of noninvasive evaluation of MVI before operation. The performance of the classifier made by TA was not superior to that of combination features based on clinic and conventional MR sequences.
Full text:
1
Index:
WPRIM
Type of study:
Diagnostic_studies
/
Prognostic_studies
Language:
Zh
Journal:
Chinese Journal of Radiology
Year:
2019
Type:
Article