search

microsoft / Recursive-Cascaded-Networks

[ICCV 2019] Recursive Cascaded Networks for Unsupervised Medical Image Registration
https://arxiv.org/abs/1907.12353
MIT License
361 stars 87 forks source link

Questions about Datasets #45

Closed CGC1031 closed 3 years ago

CGC1031 commented 3 years ago

These are some questions about your datasets that I believe will answer many people's questions about datasets. Thank you for your work and I am looking forward to your reply. I am a loyal fan of your work.

  1. Q1:

SLIVER Figure 1 is the test file of SILVER. Could you please tell me whether yan_x10 and yan_x11 are from the same person's liver part or the liver part of different people? In other words, does yan_x represent liver part of the same person at different times or images of different people's livers?

  1. Q2：

L lITS Similar to the first question, in figure2，I want to figure out lits/0, lits/1, lits2...Are they images of different people's livers? You wrote in your paper that the LITs contained 131scans, does it mean that they came from 131 images of livers of the different person, rather than 131 images of the same person at different moments.

  1. Q3:

LSPIG Through reading your paper, I understand that the LSPIG data set contains liver images from 17 different pigs, which are registered in pairs before and during surgery. Do I understand you correctly?

  1. Q4：

Thank you for reading here, the last question. For the training set MSD, it contains four kinds of liver lesion images, such as Hepatic animal and Pancreas Tumours. When you select two images for training, do you select two images from the same lesion image for concatenate ? Or will one of the two different lesion areas be selected and concatenate into the network? In other words, I would like to know how you select a pair of registration images in the MSD dataset for training.

zsyzzsoft commented 3 years ago

Q1&Q2: These scans are from different people. Q3: Correct. Q4: All pairs (can be from different kinds) are randomly selected for training.

CGC1031 commented 3 years ago

Thank you for replying so quickly. During the training stage, datasets randomly selected from different kinds, I think it's understandable. Because in fact, the training process is the process of extracting features. During the testing stage, for LSPIG data, registration for preoperative and intraoperative registration of the same person will have a good application in surgical navigation. It's OK. However, I have been puzzled by the choice of liver images from different people for the registration of the SILVER and LITS datasets. Because everyone's physical condition is different. Is there any application for scene registration by using images of two different individuals?

zsyzzsoft commented 3 years ago

I can say there are applications, though not as many as intra-subject registration, e.g. when you want to compare someone's scan with a healthy one or a standard one.

CGC1031 commented 3 years ago

Thank you for your prompt reply. I have understood.

CGC1031 commented 3 years ago

Hello! There is another question. I can get all the 2D slices after registration, but the visual effect is not good. How can I get the 3D image in the article of VTN? How do you do 3D reconstruction? Looking forward to your suggestions 3d

zsyzzsoft commented 3 years ago

These are generated by some code, but I don't have the code now.

CGC1031 commented 3 years ago

OK，I tried some methods, but they didn't look good. Thank you very much.