Hepatocellular adenoma subtyping by qualitative MRI features and machine learning algorithm of integrated qualitative
X Liu1, O Espin-Garcia2, F Khalvati3
1Joint Department of Medical Imaging, University Health Network, Sinai Health System, University of Toronto, 585 University Ave., Toronto, ON, M5G 2N2, Canada.
Aim:
To evaluate hepatocellular adenoma (HCA) subtyping using qualitative magnetic resonance imaging (MRI) features and feasibility of differentiating HCA subtypes using machine learning (ML) of qualitative and quantitative MRI features with histopathology as the reference standard.
Materials And Methods:
This retrospective study included 39 histopathologically subtyped HCAs (13 hepatocyte nuclear factor (HNF)-1-alpha mutated [HHCA], 11 inflammatory [IHCA], one beta-catenin-mutated [BHCA], and 14 unclassified [UHCA]) in 36 patients. HCA subtyping by two blinded radiologists using the proposed schema of qualitative MRI features and using the random forest algorithm was compared against histopathology. For quantitative features, 1,409 radiomic features were extracted after segmentation and reduced to 10 principle components. Support vector machine and logistic regression was applied to assess HCA subtyping.
Results:
Qualitative MRI features with proposed flow chart yielded diagnostic accuracies of 87%, 82%, and 74% for HHCA, IHCA, and UHCA respectively. The ML algorithm based on qualitative MRI features showed AUCs (area under the receiver operating characteristic curve [ROC] curve) of 0.846, 0.642, and 0.766 for diagnosing HHCA, IHCA, and UHCA, respectively. Quantitative radiomic features from portal venous and hepatic venous phase MRI demonstrated AUCs of 0.83 and 0.82, with a sensitivity of 72% and a specificity of 85% in predicting HHCA subtype.
Conclusions:
The proposed schema of integrated qualitative MRI features with ML algorithm provided high accuracy for HCA subtyping while quantitative radiomic features provide value for diagnosis of HHCA. The key qualitative MRI features for differentiating HCA subtypes were concordant between the radiologists and the ML algorithm. These approaches appear promising to better inform clinical management for patients with HCA.
