Support vector machine based MRI radiomics to identify primary hepatic carcinomas with different pathological types

Abstract

Objective To investigate the value of support vector machine based MRI-radiomics in the differential diagnosis of primary hepatic carcinomas (PHCs). Methods PHCs patients were retrospectively collected from July 2013 to February 2017 in the First Affiliated Hospital of Zhejiang University. All patients underwent unenhanced and enhanced MRI liver scan before surgery, and confirmed by pathology. A total of 294 PHCs patients (305 lesions), including 96 cases (97 lesions) of massive type cholangiocarcinoma (mCC), 107 (107 lesions) of hepatocellular carcinoma (HCC), and 91 (101 lesions) of mixed hepatocellular and cholangiocellular carcinomas (HCC-CC). All patients underwent unenhanced and dynamic enhanced MRI liver scan including arterial, portal venous and equilibrium phases. Two hundred and three lesions (65 mCC, 71 HCC, 67 HCC-CC) were assigned into the training set, the remaining 102 lesions (32 mCC, 36 HCC, 34 HCC-CC) into the validation set, according to a ratio of 2∶1. The entire lesions were delineated manually using a region of interest on equilibrium phase of enhanced MRI by using a home-made dedicated software (Analysis Kit, AK, General Electrics, US). Then the least absolute shrinkage and selection operator (LASSO) regression was used to select image features with a method of 10 fold cross-validation, and to reduce the dimensionality. The spearman method was used afterwards to condense the image features by removing redundant. A predictive model of diagnosing the different types of PHCs was established based on support vector machines(SVM),and the accuracy of applying the model in the data sets was used to evaluate the diagnostic efficacy of the model. Results A total of 280 quantitative imaging features were extracted in the training set. Thirty one imaging features were selected after LASSO regression and dimensionality reduction, and 21 features were remained after redundancy removing. The SVM showed the best generalization ability when the first 11 imaging features were used due to the Hughes effect. The 11 imaging features include 4 parameters of histogram, 2 of textures, 4 of gray-level co-occurrence matrix and 1 of gray-level run length matrix. A predictive model for PHCs was established after the study of the 11 imaging features and a regression analysis using SVM. The accuracy of the predictive model was 80.3% (163/203) in differentiating PHCs in the training set. The accuracy of the model was 75.5% (77/102) after applying it in the validation set. The diagnostic accuracy for HCC-CC, HCC and mCC was 85.3% (29/34), 77.8% (28/36) and 62.5% (20/32), respectively, in the validation set. For HCC-CC, 3 cases were misdiagnosed as mCC and 2 cases as HCC. For HCC, 3 cases were misdiagnosed as HCC-CC and 5 cases as mCC. For mCC, 9 cases were misdiagnosed as HCC-CC and 3 cases as HCC. The model showed the highest accuracy in predicting HCC-CC. Conclusion Radiomics method based on SVM may have a high accuracy in predicting different pathologic types of PHC, with the highest accuracy for HCC-CC. Key words: Liver neoplasms; Texture; Radiomics; Magnetic resonance imaging