Camaraderie: Content-based Knowledge Transfer for Medical Image Labelling using Supervised Autoencoders in a Decentralized Setting

Deep neural networks for medical imaging require large high-quality labelled data, a huge bottleneck for resource poor settings. Given the privacy requirements of medical data, institutes are un-willing to share data, causing an hindrance in resource poor settings. In the present paper, (Camaraderie: Content-based Knowledge Transfer for Medical Image Labelling using Supervised Autoencoders in a Decentralized Setting) we propose to use Discrete Classifier Supervised Autoencoder (DC-SAE) to generate low-dimensional representations of a few annotated images at the Donor client and transfer both the DC-SAE's encoder part and the latent space representations to the Recipient client without sharing raw data. We then pass the unlabelled images of the Recipient Client through this encoder to obtain their latent space representation. In a supervised setting, using latent space representation of Donor client's labelled images, we accurately annotate images of Recipient client. Camaraderie demonstrates that DC-SAE outperforms Recipient end label accuracy beyond classical VAE based classification and anomaly detection based VAE. Thus, given a limited amount of labelled data in a decentralized privacy preserving scenario, one can transfer latent space representation across clients to annotate large number of unlabelled images with high accuracy.