My experiment is as follows:
Stained for DAPI, Actin and a tracker probe on a subpopulation of cells.
I have a single dataset, but would like to classify it twice and compare the two results. I have one set of rules which use the tracker stain to detect the sub population, and this has an accuracy of ~98% - perfect! I then remove that data by excluding those columns in my properties file, and reclassify with the same training set using the other fluorescent channel, which is a simple actin stain, with measure areashape, intensity, texture etc in CP. This is currently about 70% accurate and should improve with some better microscopy.
Now, I can score all images, and compare the total number of positive and negative cells with both classifier runs, and this tells me that there is a 1.5% reduction in positive cells in the whole data set. This does not take into account cells changing from positive to negative and vice versa though.
Is it possible to get an output of each individual cell and it’s classification, so I can calculate the real error between the two rule sets? I’ve tried adding dynamic groups with imagenumber and objectnumber but this doesn’t give me a table with unique cell classification data.
The only way I can think of doing this is exporting my rules and re-running cell profiler with two rule based filters on the end of the pipeline (one including the tracker and one excluding), but this would take 8 hours. Then I could use the filter results columns to compare the two rule sets
Any suggestions would be much appreciated
Also, is it possible to extract the data for the Cross-Validation Accuracy graph so I can re-plot it?