Unsupervised Domain Adaptation with Contrastive Learning for OCT Segmentation

AbstractAccurate segmentation of retinal fluids in 3D Optical Coherence Tomography images is key for diagnosis and personalized treatment eye diseases. While deep learning has been successful at this task, trained supervised models often fail that do not resemble labeled examples, e.g. acquired using different devices. We hereby propose a novel semi-supervised framework volumetric from new unlabeled domains. jointly use contrastive learning, also introducing pairing scheme leverages similarity between nearby slices 3D. In addition, we channel-wise aggregation as an alternative to conventional spatial-pooling feature map projection. evaluate our methods domain adaptation (labeled) source (unlabeled) target domain, each containing with acquisition the method achieves Dice coefficient 13.8% higher than SimCLR (a state-of-the-art framework), leads results comparable upper bound training domain. model improves by 5.4% Dice, successfully leveraging information many images. KeywordsSegmentation volumesSemi-supervised