Evaluation code for TILs detection task

Dear organizers,

I wonder if you can share the evaluation code you are using for the challenge with the participants. I'm sure that would help lots of us with model selection. Especially, the evaluation crtierion for the TIL detection task (FROC) seems tricky to be replicated. I would really aprpeciate it if you can help with this.

Best regards, Mostafa

Here a snippet outlining the froc computation: https://surfdrive.surf.nl/files/index.php/s/RdtIAyuUQrj8fgI

Hi Witali,

Thanks for sharing the code. I appreciate it.

Hi @crunch

Can you please explain what these two fixed parameters are in the shred code: area_mm probs

Considering that we want to locally evaluate our predictions for the detection task dataset, what should be set for these two parameters? I assume these should not change from image to image, right?

Hi,

area_mm is the overall tissue area in mm² - its a normalizing factor and doesn't change the score ranking.

probs are the unique probabilities of all detections, computed from the list of all predictions per slide:

def get_unique_probs(preds_list): """ preds: [[x,y,prob]]""" probs = [] for pred in preds_list: if len(pred)>0: probs.extend(list(pred[:,2])) if len(probs)==0: return probs probs = np.sort(np.unique(probs)) return probs