COVD- and COVD-SLL Results (Deprecated)¶

In addition to the M2U-Net architecture, we also evaluated the larger DRIU network and a variation of it that contains batch normalization (DRIU+BN) on COVD- (Combined Vessel Dataset from all training data minus target test set) and SSL (Semi-Supervised Learning). Perhaps surprisingly, for the majority of combinations, the performance of the DRIU variants are roughly equal or worse to the ones obtained with the much smaller M2U-Net. We anticipate that one reason for this could be overparameterization of large VGG-16 models that are pretrained on ImageNet.

F1 Scores¶

The following table describes recommended batch sizes for 24Gb of RAM GPU card, for semi-supervised learning of COVD- systems. Use it like this:

# change <model> and <dataset> by one of items bellow
$ bob binseg train -vv --ssl <model> <dataset> --batch-size=<see-table> --device="cuda:0"

Models / Datasets	`drive-ssl`	`stare-ssl`	`chasedb1-ssl`	`iostar-vessel-ssl`	`hrf-ssl`
`driu-ssl` / `driu-bn-ssl`	4	4	2	1	1
`m2unet-ssl`	4	4	2	2	2

Comparison of F1 Scores (micro-level and standard deviation) of DRIU and M2U-Net on COVD- and COVD-SSL. Standard deviation across test-images in brackets.

F1 score	`DRIU`/`DRIU@SSL`	`DRIU+BN`/`DRIU+BN@SSL`	`M2U-Net`/`M2U-Net@SSL`
`COVD-DRIVE`	0.788 (0.018)	0.797 (0.019)	0.789 (0.018)
`COVD-DRIVE+SSL`	0.785 (0.018)	0.783 (0.019)	0.791 (0.014)
`COVD-STARE`	0.778 (0.117)	0.778 (0.122)	0.812 (0.046)
`COVD-STARE+SSL`	0.788 (0.102)	0.811 (0.074)	0.820 (0.044)
`COVD-CHASEDB1`	0.796 (0.027)	0.791 (0.025)	0.788 (0.024)
`COVD-CHASEDB1+SSL`	0.796 (0.024)	0.798 (0.025)	0.799 (0.026)
`COVD-HRF`	0.799 (0.044)	0.800 (0.045)	0.802 (0.045)
`COVD-HRF+SSL`	0.799 (0.044)	0.784 (0.048)	0.797 (0.044)
`COVD-IOSTAR-VESSEL`	0.791 (0.021)	0.777 (0.032)	0.793 (0.015)
`COVD-IOSTAR-VESSEL+SSL`	0.797 (0.017)	0.811 (0.074)	0.785 (0.018)