Table 3:

Phase classification performance of models on the VinDr-Multiphase dataset: XGBoost, ResNet3D 18-layer (r3d_18), Mixed Convolution Network 18-layer (mc3_18), R(2+1)D 18-layer (r2plus1d_18), and TotalSegmentator (ts_phase). Models are evaluated using AUC, Sensitivity, Specificity, PPV, F1 Score, and Accuracy.

	AUC	Sensitivity	Specificity	PPV	F1-score	Accuracy	p-value
Non-contrast
XGBoost	0.999	0.994	0.999	0.994	0.994	0.994	—
r3d_18	0.995	0.983	0.996	0.978	0.980	0.983	0.479
mc3_18	0.999	0.994	0.994	0.968	0.981	0.994	NaN
r2plus1d_18	0.997	0.983	0.991	0.952	0.967	0.983	0.479
ts_phase	0.986	0.972	1.000	1.000	0.986	0.995	0.113
Arterial
XGBoost	0.977	0.885	0.997	0.995	0.937	0.885	—
r3d_18	0.960	0.725	0.991	0.983	0.834	0.725	<0.001
mc3_18	0.973	0.845	0.977	0.962	0.900	0.845	0.011
r2plus1d_18	0.963	0.637	0.995	0.990	0.775	0.637	<0.001
ts_phase	0.877	0.961	0.794	0.767	0.853	0.863	<0.001
Venous
XGBoost	0.974	0.939	0.919	0.861	0.898	0.939	—
r3d_18	0.971	0.934	0.927	0.873	0.902	0.934	0.838
mc3_18	0.969	0.924	0.965	0.933	0.929	0.924	0.361
r2plus1d_18	0.967	0.927	0.907	0.841	0.882	0.927	0.475
ts_phase	0.913	0.871	0.956	0.913	0.891	0.926	<0.001
Delayed
XGBoost	0.937	0.780	0.964	0.666	0.718	0.780	—
r3d_18	0.945	0.911	0.900	0.462	0.613	0.911	0.003
mc3_18	0.953	0.862	0.932	0.546	0.669	0.862	0.061
r2plus1d_18	0.957	0.921	0.875	0.410	0.567	0.921	0.002
ts_phase	0.500	0.000	1.000	0.000	0.000	0.915	<0.001

P-values indicate the significance of accuracy differences compared to XGBoost (p<0.001 considered significant).