Differences of postprocessing allowed in the final submission

Dear Mart,

I would like to get the clarification about the differences allowed for the post-processing part between L1 and L2 in the final submission. I understand from our previous communications that post-processing used in the final submission could be different for L1 and L2 as long as the model weights are identical? Are there any specific restrictions? For example, the overlapping percentage of tiles, the method of merging tiles, the tile sizes, etc should be kept same for both leaderboards? Thanks!

best, Shuoyu

Dear Shuoyu

Indeed you are correct post-processing used in the final submission could be different for L1 and L2. For example, you might post-process the lymphocyte detections to be only predicted in the tumor stroma such that only stromal TILs are considered in L2. In principle, the idea of the challenge is to have the same algorithm evaluated both in L1 and L2. Therefore, we can imagine that the models for segmentation and detection should be capable of producing the same results in both leaderboards. Using different tile sizes does not change the results if the model is exactly the same for leaderboards 1 and 2. However, merging multiple prediction tiles to generate a segmentation output might not be seen as a form of post-processing as it would change the results in a similar way that an ensemble of models would change it

Best wishes, Mart