Virtual H&E staining and automatic segmentation of liver biopsies by means of a convolutional neural network coupled to a phasor-based algorithm.

Nowadays, pathologists provide their diagnosis by analysing stained biopsies and qualitatively evaluating the morphology and distribution of relevant biological structures. However, staining procedures are time-consuming, specimen-destructive and expensive. In an effort to manage some of these drawbacks, here, we propose a convolutional neural network (CNN) able to Haematoxylin Eosin (H&E) virtually stain entire murine hepatocarcinoma biopsies. The CNN, trained with 13000 patches ($50 \times 50 \mu m^{2})$, was validated on 4000 regions by pixelwise comparing the image colour content between the virtual and the real staining procedures ($RRMSE < 5$%). Then, phasor-based algorithms analysing colour and texture contents were applied to the virtually stained images to detect tumorous tissue and to segment relevant biological structures (accuracy $> 90%$ compared to the expert manual analysis). These results demonstrate the pipeline effectiveness to virtually stain and to segment entire biopsies. This Digital Pathology tool paves the way toward faster and quantitative diagnosis, reducing the costs sustained by the National Health System.