Uncertainty-Guided Pixel-Level Contrastive Learning for Biomarker Segmentation in OCT Images

Optical coherence tomography (OCT) has been widely leveraged to assist doctors in clinical ophthalmic diagnosis, since it can show the hierarchical structure of the retina. The type and size of biomarkers are crucial in the classification and grading of diseases. Hence, automatic segmentation of biomarkers is important to quantitative analysis, which can reduce a heavy workload. In this paper, we propose a novel deep learning-based method for biomarker segmentation on OCT images. The contrastive learning is introduced to enhance the contextual relationship between pixels in the dataset instead of just in an image. In addition, uncertainty is used to weight the segmentation loss to prompt the network focus on the learning of hard pixels. At the same time, uncertainty is utilized to select hard pixels for guiding network to perform contrastive learning, which makes the segmentation result more accurate. The experiment results evaluated on a local dataset, demonstrate the effectiveness of the proposed biomarker segmentation framework.