Contrastive Learning Driven Self-Supervised Framework for Segmentation of Biomarker of Diabetic Macular Edema

Abstract

Optical coherence tomography (OCT) has been widely used to investigate the pathological changes due to Diabetic Macular Edema (DME). In this paper, we developed a two-stage self-supervised learning approach to extract DME biomarkers. The backbone of the proposed scheme is the RAG-Net model trained in the first stage to extract different DME biomarkers, such as intra-retinal fluid, sub-retinal fluid, and hard exudates, using the low-shot supervision from the Zhang dataset. In the second training stage, it learns to extract the same biomarkers across the Duke-II dataset in a self-supervised manner (i.e., without using any ground-truth annotations) via the triplet loss function. We validated the proposed approach across both datasets at the inference stage, where it achieved the mean IoU score of 0.7610, and 0.7232, outperforming the state-of-the-art by 1.734%, and 4.839% for extracting DME biomarkers irrespective of the scanner specifications, and vendor artifacts across the Zhang and Duke-II datasets, respectively. In addition, we employed an external BIOMISA dataset to evaluate the proposed model and obtained an IOU of 0.752. The suggested model has the potential to improve clinical practice by allowing more objective assessment of DME patients. In the future, we plan to expand the application of our self-supervised model to other ocular diseases, such as glaucoma and AMD, and to incorporate more imaging modalities, such as fundus photographs and OCT-Angiography (OCTA).

Keywords: Diabetic Macular Edema (DME), Hard exudates (HE), Intra-retinal fluid (IRF), Optical Coherence Tomography (OCT), Sub-retinal fluid (SRF)