Distinction between multi_class_dice_loss & avg_dice_loss

[gattia]

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

avg_dice_loss

• 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

[ad12]

A bunch of these were legacy names that we preserved because a few different projects were using them. I'd stick to using medsegpy.loss.DiceLoss, which defaults to the same functionality as avg_dice_loss but will be more stable long term

[gattia]

Perfect - for anyone else reading this, medsegpy.loss.DiceLoss is currently specified as: LOSS: ("avg_dice_no_reduce", "sigmoid") in the yaml here and here

More versions of DiceLoss (different reductions weightings, etc.) can be added in the code near those references.

I want to try working with softmax, so I will add a softmax option that works with DiceLoss and make a pull request eventually.

[ad12]

sounds good - i think long term we should have a way of configuring this dynamically. maybe in the config the user can specify the loss class they want to instantiate and with what parameters