Regarding classes in mitotic figure outputs

Hi, I have few doubts regarding the finaly output format: 1. I have noticed in the evalutaiton that false postivites are being considered, so what are you considering as false postitives? 2. If you're considering classes then what are appropriate classes labels to be submmited,? Like what are you considering to be a mitosis like treating mitosis as '1' or '0'? 3. And in the mitotic figures output there is minor and major keys, what do they represent? If they are representing the classes, then what are their respective class names? 4. And when we are returning the predictions form detections.py to process.py do we have to send only the mitotic bounding boxes are both mitotic as well as non-mitotic bounding boxes?

Thanking you in advance, Sasidhar K

You'd be best off searching the original MIDOG website for clarification on these issues. The post about the JSON format is here: https://imi.thi.de/midog/2021/08/18/a-note-on-the-json-output-format.

Non-mitotic figures are not included in the competition submission, only mitotic figures. False positives are any predictions your algorithm makes that are not sufficiently close to a ground truth mitotic figure.

Jack

Thanks Jack. Can you tell me weather the pred (i.e., class labels) will have any influence in treating the points for evaluation as I am treating '0' for mitosis in my detection? If yes, so that I can change the script accordingly.

Thank you Sasidhar

I'm not sure I quite understand your question, but to be clear any prediction you submit will be taken to be a mitotic figure, meaning you don't have to do any work regarding other classes if you don't want to

Jack

Ok. Thank you so much for your patience and response. I was worried as in the offical repo there I didn't see any line of code restricting the predictions only to mitosis. So I thought as my label is 0 for mitosis and it might be 1 in case of the midog repo and I am getting false negatives because of that. I got your point that it doesn't matter as we only have to submit mitotic figures so it wouldn't matter.

Thanks, Jack, for clarifying this so well.

In fact the notebook provided by us might be a bit misleading in this point, but we tried to be explicit there, too, that the whole challenge is only about mitosis detection. The secondary class (which is mitosis-like cells) is however extremely helpful for setting up a proper training scheme (see "Detection of Mitotic Figures on WSI" here: https://www.nature.com/articles/s41597-019-0290-4), because they act as hard negative examples and thus help your model to not show too many false positives.

Best,

Marc

Thank you Mark , I got your point. Will try to improve on resutls.

Regards, Sasidhar