Retinal Vessel Segmentation for Retinography

		driu	hed	m2unet	unet	lwnet
Dataset	2nd. Annot.	15M	14.7M	550k	25.8M	68k
drive	0.788 (0.021)	0.821 (0.014)	0.813 (0.016)	0.802 (0.014)	0.825 (0.015)	0.828
stare	0.759 (0.028)	0.828 (0.039)	0.815 (0.047)	0.818 (0.035)	0.828 (0.050)	0.839
chasedb1	0.768 (0.023)	0.812 (0.018)	0.806 (0.020)	0.798 (0.018)	0.807 (0.017)	0.820
hrf (1168x1648)		0.808 (0.038)	0.803 (0.040)	0.796 (0.048)	0.811 (0.039)	0.814
hrf (2336x3296)		0.722 (0.073)	0.703 (0.090)	0.713 (0.143)	0.756 (0.051)	0.744
iostar-vessel		0.825 (0.020)	0.827 (0.020)	0.820 (0.018)	0.818 (0.020)	0.832

Notes

• HRF models were trained using half the full resolution (1168x1648)

• The following table describes recommended batch sizes for 24Gb of RAM GPU card:

  Models / Datasets	drive	stare	chasedb1	iostar-vessel	hrf
  unet	4	2	2	2	1
  hed	8	4	4	4	1
  driu / driu-bn	8	5	4	4	1
  m2unet	16	6	6	6	1