Fig. 1 |. Schematics of the use of adversarial domain adaptive neural networks for medical image analysis.

a, Supervised learning networks for medical image analysis are limited to fully expert-annotated datasets for training and are generally unable to adapt to unseen distributions of data collected using different imaging systems used in different clinical settings. Clinical expert staff may not be able to reliably annotate medical images obtained through portable point-of-care optical systems that are usually of lower quality compared with bulky and expensive benchtop microscopes. However, adversarial learning networks can be used to utilize standardized annotated image datasets obtained from one distribution (source) to adapt themselves with unannotated data obtained from a different distribution (target) towards a substantially more generalized neural network. b, Schematic of the general framework of the adversarial domain adaptive medical neural networks (MD-nets). The base network layers can be replaced using any regular custom neural network architecture. Additional elements for pseudolabelling can be added to enable the network to achieve adaptation in the absence of source data.