Fig. 4

Feature generation methods. a Trained classifiers can be converted into feature generators by specifying a neural network layer and calculating activations at the given layer. b Several pretrained models can be used as feature generators, including non-pathology pretrained models (e.g. ImageNet), or pathology-specific pretrained models (e.g. CTransPath). c Self-supervised contrastive learning (SimCLR) can be used to train a feature generator without requiring ground-truth labels for classification. d Features can be calculated from a model trained using self-supervised learning