PolarBean
Could Deepslice used for sagittal slice registration?
Beginner with Deepslice and ABBA, the performance of Deepslice is amazing for coronal sections, but seems no option for sagittal slices. I am just wondering if there are any solutions for sagittal slices. Thanks a lot for the project, it really helps!
Hey thanks for your question. At the moment DeepSlice only does coronal. You need to use QuickNII or ABBA to do it manually. I hope to make this a feature of a future DeepSlice version.
Hey thanks for your question. At the moment DeepSlice only does coronal. You need to use QuickNII or ABBA to do it manually. I hope to make this a feature of a future DeepSlice version.
Us sagittal-bros are eagerly awaiting this :)
@PolarBean just curious what your approach would be to supporting sagittal slices. More training data, or are there other considerations? It seems that due to the different orientation of sagittal slices compared to the coronal slices in the training dataset, that a lot of additional training data would be needed to get the model to learn the mapping from image -> CCF for sagittal slices.
Maybe you could construct "virtual" sagittal slices from the existing coronal training data without needing to collect new data that had actually been sliced sagittally?
Thanks for working on this!
Hey @aamster, exactly as you say all that's needed is sagittal training data. The Allen actually provide quite a lot of sagittal data, so it would be possible to replicate the coronal training process for sagittal. The original coronal training process was similar to what you're saying. We had a two stage process with first training on real coronal sections and then training on "virtual" slices. I think you need both to get good results. We tried a more synthetic training process with less real data for creating a rat model and the results were not nearly as good as the mouse model.
I have quite a few ideas for how we can train DeepSlice 2, they mainly focus on data augmentation and more realistic virtual slice generation.
I am not sure of the timeline for this as I now finishing my PhD and applying for postdocs so it really depends on where I end up. If it is possible to work on DeepSlice 2 I am very keen for collaboration on it.
Thanks @PolarBean! for context I am at Allen, and we have started exploring ways to support automated registration of other modalities than DeepSlice was trained on, including epifluourescence, sagittal slices, etc. Will definitely share our work with you and keep you posted!
Oh so cool. Why don't we meet to discuss this as I would like to contribute. My email is [email protected]