ConKeD++ -- Improving descriptor learning for retinal image registration: A comprehensive study of contrastive losses

TL;DR

The paper tackles retinal color fundus image registration by enhancing descriptor learning in the ConKeD framework through a systematic study of contrastive losses. It adopts a two-network baseline (keypoint detection and dense descriptor learning) with a multiview, multi-positive/multi-negative training regime and analyzes four losses (SupCon, MP-InfoNCE, MP-N-Pair, FastAP) for descriptor learning, using cosine similarity and $L2$ normalization. Evaluations on FIRE and two new datasets (LongDRS, DeepDRiD) show that the Average Precision–based loss (AP) delivers the strongest and most robust registration performance across datasets, with FastAP also performing well and other losses showing variable results; NP is notably weaker. The findings highlight that ranking-oriented losses impart useful semantic structure to descriptors, enabling more reliable keypoint matching and registration, and demonstrate the method’s generalization beyond FIRE, providing a practical, data-efficient approach to retinal image registration with publicly released pairings for reproducibility.

Abstract

Self-supervised contrastive learning has emerged as one of the most successful deep learning paradigms. In this regard, it has seen extensive use in image registration and, more recently, in the particular field of medical image registration. In this work, we propose to test and extend and improve a state-of-the-art framework for color fundus image registration, ConKeD. Using the ConKeD framework we test multiple loss functions, adapting them to the framework and the application domain. Furthermore, we evaluate our models using the standarized benchmark dataset FIRE as well as several datasets that have never been used before for color fundus registration, for which we are releasing the pairing data as well as a standardized evaluation approach. Our work demonstrates state-of-the-art performance across all datasets and metrics demonstrating several advantages over current SOTA color fundus registration methods

Figures (4)

• Figure 1: Overview of the method for both training and inference.
• Figure 2: Correlation between the proposed metrics and the ground truth in the FIRE dataset. a) is DICE, b) is IoU, c) is IoM, d) is SM, e) is SSIM and f) is LPIPS
• Figure 3: Representative example images from the DeepDRiD dataset, registered using our method trained with AP loss. Each image corresponds to a registration pair from a different patient.
• Figure 4: Representative example images from the LongDRS dataset, registered using our method trained with AP loss. a, b and c are intra-visit registration while the rest are inter-visit registrations. Each image corresponds to a registration pair from a different patient.