Automated differentiation of malignant and benign primary solid liver lesions on MRI: an externally validated radiomics model

Background & Aims: Distinguishing malignant from benign primary solid liver lesions is highly important for treatment planning. However, diagnosis on radiological imaging is challenging. In this study, we developed a radiomics model based on magnetic resonance imaging (MRI) to distinguish the most common malignant and benign primary solid liver lesions, and externally validated the model in two centers. Approach & Results: Datasets were retrospectively collected from three tertiary referral centers (A, B and C) including data from affiliated hospitals sent for revision. Patients with malignant (hepatocellular carcinoma and intrahepatic cholangiocarcinoma) and benign (hepatocellular adenoma and focal nodular hyperplasia) lesions were included. For each patient, only a T2-weighted MRI was included. A radiomics model was developed on dataset A using a combination of machine learning approaches, and internally evaluated on dataset A through cross-validation. Next, the model was externally validated on datasets B and C, and compared to scoring by two experienced abdominal radiologists on dataset C. In the resulting dataset, in total, 486 patients were included (A: 187, B: 98 and C: 201). Despite substantial MRI acquisition heterogeneity, the radiomics model developed on dataset A had a mean area under the receiver operating characteristic curve (AUC) of 0.78 in the internal validation on dataset A, and a similar AUC in the external validations (B: 0.74, C: 0.76). In dataset C, the two radiologists showed moderate agreement (Cohens {kappa}: 0.61) and achieved AUCs of 0.86 and 0.82, respectively. Conclusions: Our radiomics model using T2-weighted MRI only can non-invasively distinguish malignant from benign primary solid liver lesions. External validation indicated that our model is generalizable despite substantial differences in the acquisition protocols.