Distinction between multi_class_dice_loss & avg_dice_loss #27
Description
I was looking through the code and thinking I should probably use multi_class_dice_loss with softmax after a quick look but then it wasn't an option - so, I added it. But, then I wondered why it wasn't an option so I did more comparing of multi_class_dice_loss and avg_dice_loss. Below I summarize what I'm interpreting the main difference between them to be.
My general take is that I'd be inclined to use avg_dice_loss because it would be less likely to wash out a few bad results on segments with only a few pixels (e.g., if one image in a batch has thick healthy patellar cartilage and the other has barely any, multi_class would be skewed more towards the healthy, I think?). I'm curious on the original rationale/distinction between them and if I might be missing something.
My take is:
multi_class_dice_loss
- Flattens the batch of images into shape [batch_size* product(image_dims), n_classes].
- Use this to calculate a dice loss per class which effectively treats all images in the batch as one image.
- Shape after dice loss = [n_classes,]
- Average the dice loss from the multiple classes to get a single value
- Reshape the batch of images into shape [batch_size, product(image_dims), n_classes].
- Calcualate dice loss per image & per class
- Shape after dice loss = [batch_size, n_classes]
- Average the dice losses