Abstract
Purpose :
The quality of fundus photos can be reduced by lack of patient cooperation, pupil diameter, and media opacities. In medical imaging, computer vision filters have been proposed as a tool to pre-process and enhance image quality for the development of new deep learning (DL) algorithms. This study aims to evaluate the impact of image enhancement techniques to improve quality of fundus photographs from glaucoma eyes.
Methods :
An open-source dataset with 28,792 retinal images (EyeQ) with a three-level grading for quality (i.e., 'Good,' 'Usable,' or 'Reject') was used to train a DL model classifier based on the ResNet50 to yield quality assessment of fundus photos. The DL model was applied to four open-source datasets with fundus images from eyes with glaucomatous optic neuropathy: ORIGA, G1020, REFUGE, and CHAKSU. Contrast-Limited Adaptive Histogram Equalization (CLAHE) filter, an image processing technique used to enhance the contrast of images while preventing noise amplification, was applied to the images and the quality was reassessed by the DL algorithm. The number of images classified as ‘good’ quality before and after applying the filters was used to compare the performance of the filters.
Results :
In the original combined dataset of 4,215 images, 2,474 (58.7%) were classified as ‘good’ quality by the DL model. After applying the CLAHE filter, an additional 434images achieved a ‘good’ quality classification (total 2,908 images, 69%). Other versions of the CLAHE filters were tested: the CLAHE Modified (modified to adjust the ‘lightness’, 'red-green', or 'blue-yellow' channels depending on the color dominance), and the CLAHE-WIENER (combined with a signal processing filter used for noise reduction), each of them increasing the number of ‘good’ quality images to 61%, and 68%, respectively.
Conclusions :
In this study, we were able to use CLAHE filters, previously applied successfully for quality enhancement of chest radiography and mammography, in glaucoma-specific datasets, improving the quality of the images. This tool could be used to increase the number of good images for the development of new algorithms for image classification in glaucoma.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.