Self-supervised retinal thickness prediction enables deep learning from unlabelled data to boost classification of diabetic retinopathy

Access to large, annotated samples represents a considerable challenge for training accurate deep-learning models in medical imaging. While current leading-edge transfer learning from pre-trained models can help with cases lacking data, it limits design choices, and generally results in the use of unnecessarily large models. We propose a novel, self-supervised training scheme for obtaining high-quality, pre-trained networks from unlabeled, cross-modal medical imaging data, which will allow for creating accurate and efficient models. We demonstrate this by accurately predicting optical coherence tomography (OCT)-based retinal thickness measurements from simple infrared (IR) fundus images. Subsequently, learned representations outperformed advanced classifiers on a separate diabetic retinopathy classification task in a scenario of scarce training data. Our cross-modal, three-staged scheme effectively replaced 26,343 diabetic retinopathy annotations with 1,009 semantic segmentations on OCT and reached the same classification accuracy using only 25% of fundus images, without any drawbacks, since OCT is not required for predictions. We expect this concept will also apply to other multimodal clinical data-imaging, health records, and genomics data, and be applicable to corresponding sample-starved learning problems.