Retinal Image Deconvolution: Revealing Hidden Structures and Pathology

Abstract

Abstract: : Purpose: To demonstrate the value of individualized image deconvolution for removing image blur caused by defocus and other anterior segment aberrations. Methods: Image deconvolution has been used in astronomical image restoration for decades. Most of the classical blind deconvolution techniques suffer from high computational requirements for setting parameters needed to estimate the point spread function of the imaging system. This computationally intense process limits blind deconvolution application in real–time processing of images. We used a recently–developed efficient blind deconvolution technique that is fast, direct, and easy to automate to remove the residual ocular aberrations from the AO–corrected high resolution retinal images. The technique can be embedded in the image instrument for real–time image capture and restoration. A qualitative study is performed to evaluate the impact of image quality improvement achieved by the blind deconvolution technique. Images before and after deconvolution is applied are compared. The image comparisons were performed by two highly experienced individuals, a retina specialist and an ophthalmic analyst. A blind, forced–choice pair–wise comparison was performed, where the observers was asked to select the best image (significantly better or better). Results: Four control subjects and four subjects with diagnosed retinal disease (2 diabetic retinopathy and 2 age–related macular degeneration) were imaged with a high resolution (4Meg pixel) fundus camera. A blind deconvolution was applied. In all 8 cases the ophthalmologist and the ophthalmic technician preferred the deconvolved image, giving higher contrast and resolution as the reason for their preferences. Quantitative metrics, such as calculated contrast and spatial frequency correlated with the visual assessment.Conclusions: This project demonstrated that computer–enhanced images gave greater detail of anatomical and pathological structures.

Keywords: image processing • diabetic retinopathy • imaging/image analysis: clinical