Fig. 1: HAPPY workflow. | Nature Communications

Fig. 1: HAPPY workflow.

From: Mapping cell-to-tissue graphs across human placenta histology whole slide images using deep learning with HAPPY

Fig. 1

A hematoxylin and eosin (H&E) stained whole slide image (WSI) is first sectioned into overlapping 1600 × 1200 (177.44 × 133.08 μm) pixel images and passed to an object detection RetinaNet model which identifies the nuclei in these images. 200 × 200 (22.18 × 22.18 μm) pixel images centred on each nucleus are classified into one of 11 cell types by a ResNet-50 model. The 64-dimension embeddings from the cell classifier and their nucleus coordinates are used to build a cell graph across the whole slide image. The cell graph is input into a ClusterGCN graph neural network which classifies the tissue microstructure to which each cell belongs. Images a-d show characteristic tissue regions of the WSI: a chorionic plate, b stem and distal villi, c distal villi, d basal plate and anchoring villi.

Back to article page