search

liuquande / FedDG-ELCFS

[CVPR'21] FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space
229 stars 33 forks source link

Question about data pre-processing "center crop" #8

Open xxliang99 opened 3 years ago

xxliang99 commented 3 years ago

Dear Quande,

I read the previous issues and knew that data pre-processing had been finished before converted and saved as .npy files. I noticed that in the paper it is stated "For pre-processing, we center crop a 800 × 800 disc region for these data uniformly, then resize the cropped region to 384×384 as network input". Does it mean that only the 800800 region around the disc are saved to be the input, or the central 800800 region of the whole image? I only found the resize procedure in the provided prepare_dataset.py, so I am confused about the center-crop.

Further more, in images from domain 2 of Fundus, there are 2 ROIs in one single image and are placed on each sides of the images. If center-crop is performed on each disc region, it would generate 2 ROIs for each image; If it is performed on the whole image, then the cropped image would not contain any ROI. Does it mean that 2 input samples would be generated and saved from one single raw image in domain 2?

I would be very grateful if any clarification on "center crop" could be made. @liuquande

Thank you for your time! I was deeply inspired by your great contribution.

liuquande commented 3 years ago

Dear Vivian,

We downloaded the fundus data from Fundus, which actually has contained the pre-processed (center-cropped) samples. We directly use the pre-processed data for experiments, in which the center-cropped ROI region should be obtained from a pre-trained UNet as described in their paper.

Yes, in domain 2, the data with two views has been clipped and saved as two samples.

H-CODE6 commented 1 year ago

Hello, there was a problem in reproducing this paper recently. I don't understand the process that the author ran publicly. After seeing your successful training, would you please tell me the operation steps? Thank you very much.