Is it possible to use 2D slices?

Patch the Planet
1

Hello all,
I was wondering, if it is possible to train a CNN model using 2D slices. Also can anyone please give some information about data that you are using (2D or 3D data) to build the model. I have a low end machine. (16GB Ram +RTX 4060)

2

Hi @jarupularakesh1482. Thanks for the question. You are free to use either 2D slices or the full 3D volume for training. Just remember that the model output must be a 3D volume with a shape of the missing block of data.

Happy coding!

Onward Team

3

@team, I have a question regarding the current evaluation code. It currently uses only 3 slices (75, 150, 225). Will the final evaluation also use only these 3 slices, or will it include the full volume (75x300x1259)?

Thanks,

4

Hi @daisuke0530. The evaluation code will select 3 2D slices in the inline direction (x) and 3 slices in the xline direction (y). This will be done for each prediction volume in the test dataset. SSIM will be calculated for each of these 2D slices in the test dataset. The highest and lowest SSIM values will be dropped. The remaining values will be averaged for a final score.

Please review this post to understand the data shapes.

Onward Team

5

@team Thanks for the answer. I believe I understand the current evaluation metric. My question is about whether this metric is the same for the final prediction. I’m asking this because our predictions are actually for the (75, 300, 1259) volume, but only about 2% is used in the evaluation. I assume this is to save on evaluation costs. If that’s the case, I think using the full volume for the final prediction makes sense.

6

Hi @daisuke0530

The metric for the final submission will be the same, but the percentiles will be modified in order to prevent overfitting on the current 6 slices.

Onward Team

7

So, does this mean there is no hidden holdout test set for the final submission, but rather just adjustments to the percentiles for the current 15 test files?

8

Hi @daisuke0530

There is indeed a holdout test set for the final submissions in this challenge.

Onward Team

9

Can I ask a further explicit question, does this competition end up having a hold out test set that is completely different from the test set currently offered? In more detail, is this part of the hold out test set data that we have never seen before? Rather than just replacing unknown mask positions or slice percentiles?

10

Hi @salen. Yes, the holdout data has been generated specifically for this challenge, and has not been seen in any other challenges hosted by Onward.

Onward Team