Abstract
Purpose :
Diagnosing cataractous retinal images is error-prone due to the severe haze-effect produced by the scattering of cataract layers that decreases the image contrast. To support clinical diagnosing, we propose an image processing algorithm termed multilevel-stimulated denoising (MUTED) that can significantly increase the visual quality and suppresses the haze effect of cataractous retinal images.
Methods :
We introduce the double pass fundus reflection model and a multilevel stimulated denoising algorithm. The transmission matrix of the cataract layer is expressed as the superposition of denoised raw images of different levels weighted by stimulated functions. An intensity-based cost function was designed, and the gradient descent with adaptive momentum estimation is used to update the parameters resulting in the final refined transmission matrix of the cataract layer. We tested our methods on a total of 194 images from both public and proprietary databases and compared the performance of our method with other state-of-art enhancement methods. We used fog-ware density estimation (FADE) to evaluate the dehazing quality, and use multiscale contrast to evaluate the contrast of the image. In addition, we collected retinal images before and after cataract surgery to see whether our proposed method leads to unexpected artificial structures and test its reliability.
Results :
Figure 1 shows experimental results for visual assessment. MUTED effectively improved the clarity of retinal images as structures like blood vessels that were hidden behind the cataract in the raw image could be observed after enhancement. In addition, by comparing retinal structures presented in pre- and post-cataract surgery images, we found that MUTED did not lead to unexpected artificial structures and maintained the correct information from the images. Table 1, comparing of contrast and FADE scores of the raw and of two methods to enhance images. The MUTED has the highest contrast and the lowest FADE values denoting good dehazing results.
Conclusions :
The MUTED is able to reveal the information of cataractous retinal images and significantly increases the visual quality of cataractous retinal images. Both visual and objective assessment using fog-ware density estimation show the superiority of our method.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.