Fig. 7: Workflow overview of the traditional patch-based method, MIL method, and our proposed whole-slide training method. | Nature Communications

Fig. 7: Workflow overview of the traditional patch-based method, MIL method, and our proposed whole-slide training method.

From: An annotation-free whole-slide training approach to pathological classification of lung cancer types using deep learning

Fig. 7

a Patch-based method requires a substantial number of patch-level annotations by human experts. Given pairs of a patch image and its corresponding label in manually annotated regions, a deep neural network is trained in a strong supervision manner. b MIL leverages slide-level annotations by exhaustive inferences patches and paired the top-k patches most likely to be cancerous to their slide-level tags. These weakly paired patches and temporary tags are then used to update a neural network. c Our proposed whole-slide training method feeds an entire slide image with its corresponding slide-level tag into a neural network to fulfill end-to-end slide-level training in a strong supervision manner.

Back to article page