Possible inconsistency between Medical Decathlon Task01 BrainTumour labels and ConvertToMultiChannelBasedOnBratsClassesd class

[coding17Star]

Hi everyone,

While following the BraTS segmentation tutorial, I noticed something that confused me regarding the label definitions used in ConvertToMultiChannelBasedOnBratsClassesd.

The transform assumes:

• Label 1 = edema
• Label 2 = enhancing tumor (ET)
• Label 3 = necrotic/non-enhancing tumor core

and constructs ET as:

result.append(d[key] == 2)

However, in the Medical Segmentation Decathlon Task01 BrainTumour dataset (http://medicaldecathlon.com/), the dataset.json file describes the labels as:

• 0 = background
• 1 = edema
• 2 = non-enhancing tumor
• 3 = enhancing tumor

To better understand this, I visualized labels 2 and 3 on the T1gd modality for several cases. Example :

Visually, label 3 appears to correspond to the contrast-enhancing rim, while label 2 appears to correspond to the inner non-enhancing/necrotic region, which seems consistent with the dataset.json description.

I could not find any label remapping step before ConvertToMultiChannelBasedOnBratsClassesd in the tutorial pipeline.

Based on the dataset description and the visual inspection above, I would have expected ET to correspond to label 3 rather than label 2.

Am I missing a remapping step somewhere, or is the tutorial assuming a different BraTS label convention than the Medical Decathlon Task01 dataset?

Thanks!