Differentiation of malignant from benign focal liver lesions in triphase-enhanced CT using machine-learning-based radiomics.

Abstract

Objectives: Triphasic enhanced CT provides more information about blood supply. The aim was to establish a radiomics model of triphasic-enhanced CT to differentiate malignant from benign focal liver lesions (FLLs).

Methods: Patients with FLLs who underwent triphasic enhanced CT with histopathological results were retrospectively included. We extracted the radiomic features of each lesion in arterial phase (AP), portal vein phase (PVP), delayed phase (DP), slope of AP to PVP, and slope of PVP to DP. The features that best discriminated malignant from benign FLLs were selected using the Boruta algorithm and random forest algorithm and combined to create a radiomic signature. Three radiologists independently graded the Liver Imaging Reporting and Data System category.

Results: Of the 322 FLLs, the training, validation and test cohorts consisted of 160 (122 malignant, 76.3%), 83 (63 malignant, 75.9%), and 79 (63 malignant, 79.7%) lesions. The three observers classified 235, 169, and 220 as malignant, respectively. In the test cohort, the area under the curve of the radiomic signature in identifying malignant FLLs was 0.896 (0.850-0.973), lower than 0.935 (0.873-0.996) (P = .463) of the senior radiologist, but higher than 0.812 (0.713-0.910) (P = .228) and 0.747 (0.667-0.827) (P = .016) of the two less-experienced radiologists.

Conclusions: The radiomics-based model for triphasic enhanced CT images performed well in differentiating malignant from benign FLLs and may be a potential tool to screen for positive cases and avoid false negatives.

Advances in knowledge: The radiomics-based model for triphasic enhanced CT achieved high performance in differentiating malignant from benign FLLs and may help to screen for positive cases and avoid false negatives.