Matching 3D OCT Retina Images into Super-Resolution Dataset Agnieszka Stankiewicz 1 , Tomasz Marciniak 1 , Adam Dąbrowski 1 1 Division of Signal Processing and Electronic Systems Chair of Control and Systems Engineering Faculty of Computing Poznan University of Technology Poznan, Poland tomasz.marciniak@put.poznan.pl Marcin Stopa 2,3 , Elżbieta Marciniak 2 , Andrzej Michalski 2 2 Clinical Eye Unit and Pediatric Ophthalmology Service, Heliodor Swiecicki University Hospital, Poznan University of Medical Sciences Poznan, Poland 3 Department of Optometry and Biology of Visual System, Poznan University of Medical Sciences Poznan, Poland stopa@ump.edu.pl Abstract—Optical coherence tomography (OCT) is the current very fast and accurate modality for noninvasive assessment of 3D retinal structure. Due to large amount of data acquired with this technique the resolution of 3D scans is limited. In this paper we present a new method for improving resolution of 3D macula scans while maintaining short acquisition time and robustness with respect to motion artifacts. Our approach is based on multi- frame super-resolution method applied to several 3D standard resolution OCT scans. Presented experiments where performed on volumetric data acquired from adult patients with the use of Avanti RTvue device. Each OCT cross-section (B-scan) was subjected to image denoising and retinal layers segmentation. The generated 3D super-resolution scans have significantly improved quality of the vertical cross-sections. Keywords – OCT, super-resolution, multi-frame, retina image segmentation I. INTRODUCTION Algorithms for biometric analysis of retina images are very valuable to clinicians, as they lead to expansion of diagnostic procedures. Implementation of non-invasive diagnostic methods, such as optical coherence tomography (OCT) in modern visualization systems allows to investigate the causes of ever more frequent lesions in the eye [1]. Proper detection of retinal layers and existing pathologies is critical for diagnosis and further treatment [2]. Factors that significantly impede segmentation process are low quality of acquired images (due to heavily noisy data or uneven tissue reflectivity). Currently used algorithms in spite of high efficiency cannot cope with this type of artifacts. Volumetric OCT scan consists of a set of measurements called A-scans representing retinal tissue in depth in a single retina point. Any vector of A-scans creates a cross-section (B- scan) in one direction, and a collection of B-scans is referred to as the 3D OCT examination. Unfortunately, a typical 3D examination obtained with the fast scanner working in the horizontal direction will cause lower data resolution in the vertical direction. It is not possible to retrieve data between the cross-sections even with the use of interpolation algorithms. The calculated information would be inexact and not sufficient for the detailed, three-dimensional analysis of investigated pathologies or vessels modeling. Innovative approach presented in this article involves combination of multiple 3D scans into one super-resolution scan. The proposed solution addresses this problem by approximating detailed informative content in spaces between subsequent OCT cross-sections. The data obtained from multiple examinations of macular area takes into account displacements between the scans. This will counter for the poor resolution of the retinal thickness maps produced by OCT, which is the biggest disadvantage of OCT and still remains to be solved in forthcoming OCT models [3]. II. SUPERRESOLUTION IN OCT A. State-of-the-art Biomedical imaging devices have limited achievable resolution due to both theoretical and practical constraints. Table 1 presents standard resolutions of the 3D OCT examinations for selected devices. TABLE I. STANDARD RESOLUTIONS OF 3D EXAMINATION FOR TOPCON [4], AVANTI [5] AND COPERNICUS [6] DEVICES Device Topcon Triton Avanti RTvue Copernicus Data resolution [px] 256×512×992 141×385×640 100×800×1010 Volume [mm] 6×6×3 7×7×2 8×8×2 Typical 3D OCT scan acquired with the speed of 70 000 A- scans / second with the resolution of 141×385 A-scans takes approximately 0.94 second [5]. Such examination, representing a volume of 7×7×2 mm, has unequal data resolution (141×385×640 pixels, respectively) in respect to the measured tissue. An example of horizontal and vertical cross-section from such scan is illustrated in Fig. 1 (fast scanner working in the horizontal direction). As can be seen, B-scan corresponding to the vertical direction in the eye is deficient (presents discontinuities, and extrapolated data samples). This work was prepared within the PRELUDIUM CADOCT-Project number 2014/15/N/ST6/00710 founded by National Science Centre Poland.