Fig 1.
(A) Distribution of LDs of Abnormal Liver Density. (B) Distribution of areas of abnormal liver density. (C) Distribution of LDs after bootstrap resampling with a mean value of 100 samples repeated 1,000 times (r = 0.9980, p < 0.0001). (D) Distribution of areas after bootstrap resampling with a mean value of 100 samples repeated 1,000 times (r = 0.9988, p < 0.0001).
Table 1.
Properties of training and testing data.
Fig 2.
(A) Abnormal Hyperdensities Without Labeling ROI. (B) Hyperdensities with hand-marked ROIs. (C) Hyperdensities with AI-predicted ROIs.
Fig 3.
(A) Abnormal Hypodensities Without Labeling ROI. (B) Hypodensities with hand-marked ROIs. (C) Hypodensities with AI-predicted ROIs.
Fig 4.
(A) AI-predicted Rectangle ROIs of Heterogeneous Densities with Concurrent Outputs of LD and Area. (B) AI-labeled LDs (7.4 and 5.2 cm) of abnormal densities through ellipse fitting. (C) AI-labeled areas of abnormal densities through contour delineation (area = 35.5 and 15.6 cm2). (D) Overlapping hand-drawn GT ROIs (light blue) and Mask R-CNN-predicted ROIs (pink) of abnormal densities (Dice coefficients = 0.9667 and 0.8922).
Fig 5.
Diagram of mask R-CNN modeling and inference building.
Fig 6.
Process of mask R-CNN modeling and inference building.
Fig 7.
Receiver operating characteristic curve of lesion detection inference for all tumor densities.
Fig 8.
Receiver operating characteristic curve of lesion detection inference for different densities.
Table 2.
TPR (%), FPR (%), and dice coefficients of different confidence THs and mask THs.
Fig 9.
Analysis of 1781 false positive images.
(IVC: inferior vena cava; HV: hepatic vein; PV: portal vein).