How can I create a binary contour from a probability map ?

I use a CNN to perform a pixel classification with two classes : mitochondria and background. As an output, I get the score of each pixel, that is an image where each pixel has a value between 0 and 1. (see the attached file). The lowest score (darker) corresponds to the pixels that are at the edges between the mitochondria and the background.

I am trying to convert this output to a binary image which characterizes the contour of the mitochondria, where each contour is closed. I have tried global and local threshold without success: if the threshold value is too high, I get black pixels within a mitochondria that divide them into two different elements, and if the threshold is too low, I lost the closure of the contour.

Do you have any idea on how process this probability map to get this binary image ? Is there any function that can compute the "lowest intensity path" and give a binary image ?