How to create masking layer for image defects in ground truth volume (VAST)

Seung Lab

General rule:

1. If image defects are not severe and could be annotated with high confidence with pixel accuracy, there is no need to create a masking layer.
2. If image defects are fairly severe that it’s not easy to completely “imagine” through them with confidence (even with the aid of 3D context), then consider creating an additional annotation layer/masking layer for the defects (same as what is done for the folds in minnie.)
3. It’s always better to have rather “conservative” masking layer for the defects, i.e., “overcoloring” is better than “undercoloring.” (e.g. color when you see different grayness of precipitation from cytoplasm)
4. Regarding (2) & (3), be very careful not to make the following mistake: crudely “imagining” through and leave the uncertainty “unmasked.” If something remains to be uncertain (e.g. blank, incomplete coloring), they must be covered by additional masking layer.

Fig.1. For the pixel inaccurate parts, additional masking layer should be added if they are hard to imagine through with high confidence.

5. There should be no remaining uncertainty in the main annotation. Everything should be covered by either (1) main annotation or (2) additional mask.
6. For masking precipitation, all could be in one segment color.

Parking lot:

• Assigning a single color to each object is still okay. But we might want to assign different colors to different types of image defects. It’s too early to talk about that much specific strategy (Kisuk)