Purchase this article with an account.
Sandeep Bhat, Chaithanya Ramachandra, Muneeswar G Nittala, Srinivas R Sadda, Kaushal Solanki; Advanced Image Enhancement and Analysis Tools for Image Quality Assessment. Invest. Ophthalmol. Vis. Sci. 2014;55(13):4834.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Image quality assessment is the first step to retinal image analysis for screening or biomarker computation. The cameras used, illumination, the technicians, and anatomy/race of patients can all be different for fundus images, leading to varying blur, color saturation levels, dynamic range, and sensor noise. This makes it difficult for same algorithms to process these images. Our image enhancement step “neutralizes” image appearance, allowing the images to be processed by the same algorithm using identical parameters. Then quality assessment step gauges image suitability for further analysis.
Retinal fundus mask is first estimated using thresholding and morphological filtering. Noise is reduced using edge-preserving bilateral filter. Then a novel median filter based normalization technique is applied, that uses local background estimation to locally enhance the image at every pixel as shown in Fig 1. The normalized images also show improved visibility of vessels and lesions. The appearance “neutralized” images are processed using a Hessian-based interest region and “vesselness” map detection method, and then used to obtain the following quality descriptors on original/enhanced images: sum-modified Laplacian (for focus/blur), saturation measure, Michelson contrast, image entropy (for texture), local binary patterns (for texture), color measure, local noise metric. The descriptors are concatenated, subjected to dimensionality reduction using PCA, and used to train a support vector regressor that produces a continuous score to indicate image quality (see Fig 2).
Image enhancement produces good normalized results for images taken on different cameras (Canon, Topcon, Zeiss, iPhone Ocular Cellscope), under varying imaging conditions (see Fig 1). On a USC/DEI clinical dataset of 125 images (73 adequate, 31 fair, 21 poor quality), regression had mean squared error of only 0.22. For a good (adequate/fair) vs bad (poor) quality classifier, we obtained AUROC of 0.91.
Proposed enhancement method “neutralizes” image appearance and allows same algorithms to process the images. The quality step reliably assesses suitability of image for further analysis. Together they enable improved screening or biomarker computation on retinal fundus images.
This PDF is available to Subscribers Only