Performance dependency of retinal image quality assessment algorithms on image resolution: analyses and solutions

Abstract

Retinal image quality assessment (RIQA) is the first step performed in retinal image processing systems necessary to assure that the processed images are suitable for analysis and medical diagnosis. RIQA algorithms created for controlled environments can result in degraded performance for cross-dataset experiments in which the train and test images have different resolutions. The effect of image resolution on the performance of four different RIQA algorithms, chosen to include generic, segmentation, and transform-based quality features, is studied using datasets of various resolutions. Analyses showed that for cross-dataset classifications, the performance of some RIQA algorithms was reduced by up to 50% in cases where the train and test dataset image resolutions were significantly different. A statistical analysis was conducted to study how the retinal image quality features are affected by image resolution which resulted in their categorization into resolution-dependent and resolution-independent features. Feature scaling was then introduced to overcome the transform-based RIQA algorithm's cross-dataset performance degradation resulting in a 100% performance enhancement. Based on this study, the investigation and enhancement of the cross-dataset performance of RIQA algorithms are recommended as a standard part of their design in order to assure their performance reliability in processing images of various resolutions.

http://ift.tt/2qWBu9y