Image registration with image masks

Hi organizers,

Just want to confirm with you about the usage of the image masks.

For task 1 and task 3, we noticed that if we solely input the mask pairs into a simple CNN, the performance, regarding the Dice score, is already surprisingly high. Then one may first adopt a segmentation model to segment these input images and then directly and solely compute the displacements on the masks. By doing so, obviously, the performance is dependent on the segmentation accuracy rather than registration itself.

It's like image registration without images.

A similar problem also exists in task 2, that is one can first compute landmarks on images and then directly estimate the displacements from the landmarks.

We believe the organizers should make this clear that whether we are allowed to directly compute displacements without original images involved. Are we allowed to stack the image and its mask as a whole as input as well?

Best regards xi

According to a paper published by one of the challenge organizers (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7085718/pdf/sensors-20-01392.pdf), the default approach would be to train a segmentation network and use the output for registration. This is also what the challenge organizers repplied when I asked a similar question.

Hi Xi,

Thanks for your question. Please be aware that the segmentation masks are not available for the test scans but only for the training and validation cases.

If you like to use a segmentation mask as an (additional) input for your registration network, you have to generate them with your algorithm. However, a pure segmentation-based registration might have some drawbacks, because it probably will focus too much on the alignment of these specific organs. We will further evaluate the registration accuracy besides the Dice Score of those organs.

Best, Alessa