False positive and false negative?

Hello Everyone,

I have a feeling this is a common problem. When I’m using the identify primary objects tool, because the image is black-and-white and there are sometimes shaded regions in the cell, the computer thinks that it is TWO cells and thus, I have a false positive number. If I increase the minimum allowed distance for local maxima, then this will cause some other cells to have a false negative (because they’d be too clumped together.

What should I do? Attached is my pipeline and a picture.



Still_Testing_DAPI_1.cp (10.2 KB)

That type of error is termed a segmentation error it is not a true false positive error. For example, an erroneous object count is only a true false positive error if it is scored by the classifier and the classifier calls a DAPI positive nucleus as DAPI negative which is not what is happening here. You just need to count more cells because segmentation errors are a fact of life when performing quantitative image analysis. The solution is to count many more cells and score them as positive and negative for your phenotype of interest (i.e. DAPI+your favorite protein). Setup the classifier to fetch 100 positive cells and 100 negative cells after you train the classifier to yield an acceptable per-object scoring accuracy. If the classifier retrieves 99 TP and 1 FP from the positive class it would have a false negative rate of 1%. Do the same for the negative class to calculate the false positive rate. If you don’t want to include segmentation errors in the analysis you could try to filter the segmentation errors out by bracketing the nuclear area around what you think is acceptable. However, I would not worry too much about it since counting more cells tends to average out the small number of segmentation errors that could be in your image analysis. Is your image out of focus because the nuclei look a little blurry? If I were you I would try to image all experiments at 0.5 um/pixel or higher resolution because it helps the segmentation routine to have more pixels in the nuclear compartment. Also, make sure the integration time for the detector yields mean nuclear DAPI staining intensities around 0.7-0.9 of the full pixel well capacity since that can help with segmentation also.

Derek

Hi tbrri,

Derek is pretty much spot on with his assessment, which is particularly relevant if you are using a classifier of some type (whether CellProfiler Analyst or otherwise). However, if you are not, and you want to correct segmentation errors in CP alone, it may be worth tweaking the smoothing filter size, keeping the maxima suppression distance constant. However, it’s hard to tell how much this would help based on the example output you posted since you seem to do quite a bit of pre-processing prior to detection; the smoothing filter size may have the same net effect as the pre-processing you’re performing.

Also, you may want to try Laplacian of Gaussian (LoG) for the declumping method since it sometimes does better at maintaining blob segmentation than the other methods. However, the associated LoG settings are not intuitive if you need to adjust them.

Regards,
-Mark