Table 1.

Voxel-wise test results of the random forest classifier (RF), a Unet, the BRAVE-NET networks, referred to as proposed and BRAVE-NET networks trained with augmented training data, referred to as proposed-augmented.

Voxel-wise scores			Aggregated segments		Detailed segments
Model	Input	Segment washing	mF1	bAcc	mF1	bAcc
RF	Image + vessel	–	0.54	0.77	0.34	0.57
Unet	Image + vessel	–	0.83	0.83	0.71	0.73
Unet	Image + vessel	✓	0.86	0.85	0.75	0.77
Proposed	Image	–	0.50	0.47	0.40	0.40
Proposed	Image	✓	0.63	0.55	0.47	0.45
Proposed	Vessel	–	0.79	0.80	0.68	0.72
Proposed	Vessel	✓	0.84	0.84	0.74	0.76
Proposed	Image + vessel	–	0.84	0.84	0.73	0.76
Proposed	Image + vessel	✓	0.88	0.88	0.78	0.80
Proposed-augmented	Image + vessel	–	0.86	0.87	0.76	0.79
Proposed-augmented	Image + vessel	✓	0.89	0.90	0.80	0.83

Results are shown for models trained on aggregated and detailed vessel constellations and subsequent application of segment washing of predictions is indicated. Metrics shown are macro F1 score (mF1) and balanced class accuracy (bAcc). Performance evaluation of the proposed models with segment washing did not involve additional training or retrieval of predictions. Bold values indicate overall best performance with respect to the given metric.