Implement Normalised Surface Distance (Normalised Surface Dice) in metrics. #100
Description
Is your feature request related to a problem? Please describe.
A major problem related to the Dice metric is that it is not robust against the error-prone reference annotations. Specifically in the context of histopathological images where the structures are small, the dice metric might provide misleading performance measures. However, there are better metrics that allow for a boundary tolerance specifically to address the issue of erroneous boundary annotations such as the normalised surface dice (NSD).
Check here. It would be nice to have an implementation of this in ahcore.
Even though, Dice score is generally robust against large anotomical structures and we already have an implementation of an "overall-dice" in our metrics, the probelms at the annotation boundaries aren't solved due to the strict penalisation for pixels at annotation boundaries.
Below are some observed problems while training our pipelines:
- Gaps between two polygons within the annotations which do not get assigned a class during one-hot-encoding
- Some edge artefacts that have happened due to human error while annotating.
Describe the solution you'd like
Integrate the normalised surface dice along with the metrics we currently have. Monai repository provides a ready-made implementation.
Describe alternatives you've considered
N/A
Additional context
A screenshot from the paper cited above.
