Contrastive learning to deal with multiple contrasts

[jcohenadad]

Contrastive Predicted Coding uses a contrastive loss to learn a representation which can be predicted with an autoregressive model [Oord 2018]. With contrastive multiview coding, a network must produce similar features for images of different modalities if they correspond to the same object [Tian 2019]. Contrastive learning applied to image segmentation was investigated in [Chaitanya 2020]. More recent work from Peng et al, using mutual information and regularization across decoder/encoder.

These approaches might be beneficial if our goal is to have the model behave robustly to input image contrasts that are not in the image contrasts used to train our contrast-agnostic segmentation.

Maybe worth investigating?