Fig. 3

Confusion matrices from the two internal datasets with example lesions from the internal test dataset. a Confusion matrices comparing U-Net predicted and manually drawn ground truth lesions split up into subplots based on the largest axial diameter of the ground truth segmentation (or predicted segmentation for lesions not marked in the ground truth segmentation) lesions. b-d True positive (segmented in the correct category) examples. e False negative (undetected) examples. f False positive (manually not marked) examples. Numbers on the left in the given row indicate largest axial diameter range from which the lesion is sampled. Input channels in order: NCE – non-contrast-enhanced, pre-contrast T1; AP, PVP, TRA—arterial, portal venous, transitional phase contrast-enhanced T1; HBP – hepatobiliary phase contrast-enhanced T1; IP, OOP – in- and out-of-phase T1-weighted sequences; T2H – T2-weighted HASTE; T2B – T2-weighted BLADE; T2LTE – multiple types of T2-weighted images with longer time to echo; DWI-L, -M, -H – diffusion-weighted imaging with three increasing b-value ranges; ADC – apparent diffusion coefficient maps. BG: background; mm: millimeter; GT: ground truth: n: number of patients; LR-3, LR-4, LR-5, LR-M: LI-RADS categories