Information loss in features from cellprofiler

Does cell-profiler provide features that explain clusters and cell aggregation?
and how much information loss occurs after converting an image to set of features?

CellProfiler’s MeasureObjectNeighbors module provides a number of features to do with the clustering of cells; you could also do something like a distance transform in the Morph module and measure that.

The second question is a nearly philosophical one, and I’m afraid I can’t answer other than “it depends”! Certainly the more measurements you add, the more information will be captured, though we certainly know it reaches a point of diminishing returns where new features end up simply highly correlated with old ones. I’m not aware of any quantitative way to assess this but would be fascinated to hear others’ opinions!

CellProfiler includes a handful of entropy and loss measurements. I believe the majority are packaged in the MeasureTexture module. You can read more about them in the manual.

1 Like

By information loss, I meant that some features that can be extracted by deep models for example from a cluster or whole image are how much probable to be extracted by cell-profiler? (because in my view, cell-profiler provide more information about single-cells and pay less attention to their interactions)

Ultimately, your pipeline will extract whatever you tell it to, so a lot of it comes down to pipeline design. If you add lots of different MeasureObjectNeighbors parameters, look at distance transforms, possibly even look at objects at different scales (“cells” and “clumps”), it will have a lot of rich information about cell context. If you don’t add those things, it won’t tell you much if anything about cell contexts.