The article introduces ConKeD, a novel deep learning approach for Retinal Image Registration (RIR). Unlike existing methods, ConKeD uses a multi-positive multi-negative contrastive learning strategy to utilize additional information from available training samples. This approach enables high-quality descriptor learning from limited training data. The descriptors are combined with domain-specific keypoints, such as blood vessel bifurcations and crossovers, detected using a deep neural network. Experimental results show that this strategy outperforms the widely used triplet loss technique and the single-positive multi-negative alternative. ConKeD offers several advantages, including avoiding pre-processing, requiring fewer training samples, and fewer detected keypoints.
Publication date: 11 Jan 2024
Project Page: https://arxiv.org/abs/2401.05901
Paper: https://arxiv.org/pdf/2401.05901
