For one task we can have several test datasets. Because of this, there can be several identical or near identical modalities. The current aggression method computes the dice/hd95 for each dataset_modality (ex: chaos t1w dice = 0.9, Amos t1w dice = 0.8, Amos CT dice = 0.5 => average = 0,733). This leads to a overrepresentation of the repeated modalities (here t1w).
To mitigate this, we'll instead average all same modality results, regardless of the dataset provenance, and then average the per modality results. Ex: chaos t1w dice = 0.9, Amos t1w dice = 0.8, Amos CT dice = 0.5 => t1w dice = 0.85, t2w dice = 0.5 => final average = 0,675.
For one task we can have several test datasets. Because of this, there can be several identical or near identical modalities. The current aggression method computes the dice/hd95 for each dataset_modality (ex: chaos t1w dice = 0.9, Amos t1w dice = 0.8, Amos CT dice = 0.5 => average = 0,733). This leads to a overrepresentation of the repeated modalities (here t1w).
To mitigate this, we'll instead average all same modality results, regardless of the dataset provenance, and then average the per modality results. Ex: chaos t1w dice = 0.9, Amos t1w dice = 0.8, Amos CT dice = 0.5 => t1w dice = 0.85, t2w dice = 0.5 => final average = 0,675.