Allen Cell Discussion Forum

Deep Learning Framework for Allen Cell Segmenter

Is your iterative deep learning framework based on this paper?

I was also curious about how the deep learning framework was trained. Do you train using 3D cell image data that was annotated slice-by-slice? If so, do you need to annotate all the slices or can the model work using sparse annotation like in this paper?

The details of the iterative deep learning framework are explained in the above bioArxiv paper.

Regarding annotation, slice by slice painting is probably the worst option, but technically our framwork still supports that. Our recommended strategy is detailed in the paper. In brief, we start with one version of suboptimal automatic segmentation, creating the training data with such results via sorting/merging, and training a new model which will produce better results. If results are still not fully satisfactory, more of above curation/training iteration can be used. Inside our institute, we are using many deep learning based segmentation in different applications, but never manually paint on the target structute slice by slice.

Let me know if you are still not clear after checking out the bioAxiv paper.

That approach sounds good.

However, after using all these tools I am interested in comparing my result to my own personal manual annotation. Do you know how manual annotation of 3D images is typically done in practice? Is it done slice-by-slice?

as far as i know, slice by slice is the most common way. itk-snap or napari are good tools for this purpose.

1 Like