Figure 7
From: Weakly-supervised learning for lung carcinoma classification using deep learning
Overview of the two training methods: fully-supervised learning and weakly supervised learning. In the fully-supervised learning all the tiles from the WSIs were used for training and their labels were assigned directly based on the labels of the annotation regions they belonged to. In the weakly-supervised method, we iteratively alternated between inference and training. During inference, the model weights were frozen, and it was applied in a sliding window fashion on each WSI. The top k tiles with the highest probabilities were then selected from each WSI. During training the selected tiles were then used to train the model.
