model

• SA-Unet

Requirement

• torch--1.10.1
• wandb:Similar to tensorboard, record the result curve and save it to the cloud

Precision

Dataset	SE	SP	ACC	AUC	F1
DRIVE	0.8234	0.9840	0.9708	0.9872	0.8221
CHASEDB1	0.8352	0.9885	0.9774	0.9917	------

Be Careful

• The original author did not use mask, Instead, it crop the image to 592*592, which causes most of its precision to be high, and if you use mask, this network will only achieve average results
• Data enhancement: the author code generates 3 * 3 * 20=180 pictures, and the author writes 256. I don't know what the situation is.
• Please use the data enhancement script to augment your own dataset, put it in the aug folder, and then train the network.