Use of super-resolution image reconstruction techniques in optical coherence tomography

Abstract

 

Purpose

 

To assess the use of super-resolution (SR) image reconstruction techniques to improve the quality of OCT images.

 

Methods

 

OCT axial resolution is limited by the characteristics of the light source, while transverse is limited by the eye optics (in retinal imaging). Additionally, speckle noise limits the image content. The high acquisition speed of current systems enables multiple B-scan acquisition from the same retinal location. Eye-instrument microfluctuations mean that there are small displacements between images in a set of multiple B-scans. Image registration followed by an average (R+Av) of the image set is commonly used to remove speckle. However, assuming the set of scans are taken across the same retinal cross-section, additional analysis of the image set provides further information regarding the tissue. A 3-step procedure was used, based on multi-frame SR techniques involving: registration (subpixel motion estimation), interpolation (high resolution image from a composite of low resolution {LR} images), and restoration (removes blur and noise) to provide a B-scan of increased resolution. The SR performance was evaluated visually and by calculating the contrast to noise ratio (CNR) [contrast between features {retinal layer} and background regions {vitreous}].

 

Results

 

To illustrate this approach, we captured a sequence of 8 retinal B-scans [see Fig 1]. The analysis is performed on 2 regions of interest. Fig 1A shows the so-called LR B-scans whereas Fig 1C shows the SR result. In this example the SR factor was set to 2.5. For comparison, R+Av results are included [Fig 1B]. The SR algorithm performed reliably, resulting in an image of increased resolution compared to the initial image, as can be seen by comparing Fig 1A and 1C. The image details and delineation between layers appears increased compared to the R+Av techniques Fig 1B (i.e. RPE/Bruch’s membrane complex). Fig 2 provides an edge profile across the central foveal A-scan to assess this improvement. The SR CNR metrics improved by 36% and 14% with respect to the LR and R+Av.

 

Conclusions

 

SR techniques can improve OCT visualization and have the potential to further utilize multi-frame information to enhance OCT imaging.

 

 

Fig 1. (A) LR B-scans [1 out of 8]. (B) R+Av of the 8 LR B-scans. (C) SR technique on the 8 LR B-scans [zoomed 0.4x for comparison purposes].

 

 

Fig 2. A-scan profile across the foveal retinal image for different images.

 

Keywords: 549 image processing • 551 imaging/image analysis: non-clinical