Self-Supervised Deep Learning —The Next Frontier

The most common way to train a deep learning model for medical image classification purposes, including for ophthalmic images, involves supervised learning in which training data are manually labeled by trained human graders. Then, transfer learning may be applied to a pretrained “off-the-shelf” model backbone, such as VGG and ResNet, and model fine-tuning is performed with the labeled ophthalmic data. This common workflow is limited by the time-consuming and labor-intensive nature of training data annotation. One promising approach to address this limitation is self-sup ervised learning (SSL), which is the focus of the article by Gholami et al in JAMA Ophthalmology. As the name suggests, SSL is a technique that obviates the need for human annotation of the training data.
Source: JAMA Ophthalmology - Category: Opthalmology Source Type: research