Purchase this article with an account.
M. Wilson, P. Nguyen, B. Raman, S. Nemeth, P. Soliz, S. Wolf; Automatic Registration of Multi-modal Retinal Images, and FA and ICG Videos for Examinations Over a Six Month Period . Invest. Ophthalmol. Vis. Sci. 2003;44(13):1769.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose:To demonstrate an automatic technique to register (align) spatially and temporally multi-modal retinal digital images to enable inter-mode comparisons of lesion margins and to precisely monitor longitudinal changes. Methods: Automatic alignment of six different imaging modalities, including video and still images, has never been successfully implemented. The six imaging modalities include digital color photographs, autofluorescence, infrared, red free, and both indocyanine green (ICG) and fluorescein angiography (FA) videos with still frames. All images were acquired with the Heidelberg Retinal Angiograph (HRA) or a digital fundus camera at baseline, 6 weeks, 3 and 6 month intervals. All of the still and video images were brought into alignment with a universal base FA still image selected by an expert retinal grader and used as the point of reference for all other modalities. A correlation technique was developed that accounts for the widely varying intensity characteristics inherent with multi-modal images. Control points, which are automatically located, were determined by the correlation technique are then fed into a modified local affine transformation algorithm. Results: A technique which combines qualitative and quantitative analysis for validation was performed to assess the accuracy of the alignment. The qualitative analysis called for an expert grader to manually mark 25 points of similarity between the registered image and the universal base. The points were distributed throughout the field of view through use of a 5x5 grid overlay. The quantitative portion of the analysis calculated radial distance between the point pair positions for all points within the grid. The results determined that the accuracy was 1.5 pixel (on average) within a 7.2- millimeter radius of the fovea and 2.5 pixels (on average) outside of this para-macular region. Conclusions: Longitudinal retinal screenings of many ocular disorders can involve images acquired from many different imaging modalities over time. We have demonstrated a robust and accurate registration technique that aligns multi-modal images through time. Registration is critical for mining image data and employing it in future feature analysis and pattern recognition processes.
This PDF is available to Subscribers Only