Closed wdeback closed 6 years ago
Great of have feedback so soon after releasing the paper (how did you even find it!?).
Thanks for the suggestion, I will definitely check it out. I worked on this implementation for a while and was disappointed when it did not work for the use case I wanted, so I just decided to release the code and the paper on arxiv.
Is this use-case something you work on? If so, perhaps we could collaborate and do some more impactful work.
Let me know what you think.
Is this use-case something you work on? If so, perhaps we could collaborate and do some more impactful work.
Not my own use-case, but friends are working on segmentation of 3D light sheet image data of brain cytoarchitecture. They are always facing issues with annotation and may be very interested. I'll ask.
Ok thank you, please let them know I am open to collaborating.
A suggestion rather than an issue.
In your paper (https://arxiv.org/pdf/1807.07464.pdf), you observed no improvement of using the CRF-RNN. However, this is based on using full voxel-wise annotated labels.
As explored in the following paper, a CRF/RNN layer could be very useful to allow for weak (scribble-based) annotated data:
Learning to Segment Medical Images with Scribble-Supervision Alone Yigit B. Can, Krishna Chaitanya, Basil Mustafa, Lisa M. Koch, Ender Konukoglu, Christian F. Baumgartner https://arxiv.org/abs/1807.04668
I think weak annotation is especially important in 3D where accurate voxel-wise annotation is extremely difficult and time-consuming. This may be where your 3D CRF/RNN implementation could offer a key advantage.