Precision recall curve

Hi, I have a question regarding the Precision-Recall curves generated with picai_eval: very often they are not monotonic and go up and down.: https://ibb.co/K5H4VbG

Is it normal? Does it have something to do with the way the detection maps are genrated from softmax predictions?

Thank you!

Hi Youssef.

PR curves aren't necessarily monotonically increasing (unlike ROC and FROC curves). In fact, most commonly they never are. Your plotted curve looks as expected, just that you seem to be using a step function for plotting, while we typically use the plot or scatter functions.

You can refer to this thread for a clear and concise explanation on why PR curves aren't always monotonically increasing. There are more extensive explanations all over the internet too, of course.

Hope this helps.

Hi Youssef and Anindo,

I would say the normal way to plot a Precision-Recall curve is with the step function, as in the screenshot. At least, that is what scikit-learn implements in their PrecisionRecallDisplay, and what we have recommended in picai_eval so far.

Kind regards, Joeran

Thanks a lot! Youssef