Accuracy module?

cellprofiler

#1

Hello,

Are you planning to prepare a module to measure accuracy of results (comparing them with correct solution). Let’s say confusion matrices, precision-recall plots, etc I am assuming there is nothing like that in the current version, right?

Thanks

Joaquin.


#2

Hi Joaquin,

Currently we have no modules that do what you have mentioned. Could you explain more what it is you are trying to do? What do you want to compare for accuracy?

Mike


#3

For example. let’s say I have “labeled” a small part of the dataset. I know, for example, the number of mitotic cells per image, the centroids and the number of centrosomes per cell. And, of course, I would like to play with the configurations until I have the best possible results in terms of number of cells and number of centrosomes per cell detected.

Although I guess it could be a long-term aim, do you think it is worth it to start developing a module like that?

Best.

Joaquin


#4

Hi Joaquin,

I think, in general, you can determine the accuracy of your analysis by looking at the identification windows. If your count is incorrect, visual analysis would quickly tell you what is wrong, as opposed to a number which didn’t match what you were expecting.

Mike


#5

For more quantitative comparisons, you could also try using Data Tools > Plot Measurement to choose two measurements, one being the CellProfiler measured results, and the other being your own “ground truth”, validated results. But first, you will need to add your ground truth results to the output file: to add to an existing output file, use DataTools > AddData and to add to output files automatically when you are running pipelines, add the module LoadText so your own data will automatically be integrated into the output file.

In fact, the entire process can be automated if you also add the DisplayMeasurement module to the end of your pipeline, which has the same functionality of the PlotMeasurement data tool. Let us know if this works well for you!

Anne


#6

Thank you,
this is exactly what I was searching for.

best regards
Ingo