Performance Metrics

[AlexInJar]

May I ask how is the true positive in cell detection instance calculated in the metric. Is there perhaps a IoU threshold which you have applied that determined whether we have a true positive? Also, may I ask if the deep cell toolbox package supports using the AUC score and other metrics related to instance segmentation?

Thanks

[ngreenwald]

The true positive is determined using a matching algorithm. You can take a look at the code to see how its implemented. There's also a section in the methods section of the Mesmer paper that discusses it.

We don't report AUC, but there is a F1, precision, and recall reported as a default. However, the predictions use the same format as most other segmentation algorithms (integer cell id masks), so you could feed them into whatever evaluation code you like to generate additional metrics.