Robust Pairwise Registration for Images of Indocyanine-Green Angiographic Sequences

Motivated by the need of information integration from different image modalities for treatment of ocular diseases, this paper introduces an algorithm that registers image pairs from a complete IndoCyanine green angiography (ICG),containing Infra-Red (IR) and ICG images, for diagnosis of diseases in the choroidal layer, such as exudative senile macular degeneration. Challenges of the work include low image quality due to the presence of the Pigmented Retinal Epithelium and substantial appearance differences between images of different phases due to the circulation of the dye. Improved upon our previous work, Edge-Driven DBICP, with the focus on the image properties of an ICG sequence, our algorithm extracts features for registration from images with enhanced vessels to reduce the effect of the noise. For registration, Lowe keypoint matches for initialization are rank ordered by both distance and saliency measures and transformations are refined, going from local and low-order to global and higher-order model, in succession. Our dataset consists of 62 randomly-selected, pathological ICG sequences, each on average having up to two IR images and 12 ICG images. Our method successfully registered 83.4% of IR-ICG pairs, 89.7% of ICG-ICG pairs, and 86.37% of all pairs, which are about 49.3%, 25.5%, and 30.7% improvement, respectively, over Edge-Driven DBICP.