A novel methodology based on Contrast-Adaptive Threshold Model for accurate optic disc detection in retinal images of premature infants.

Abstract

Identifying the exact location of the optic disc in retinal images is an important task while performing the retinal image analysis. Localization of the optic disc generally fails to detect its exact location due to unclear boundaries and low contrast images, especially in retinal images of infants where the process of retrieving the images should be very quick in low light conditions. The objective of this research paper is to detect the location of optic disc using a segmentation algorithm titled "Contrast-Adaptive Threshold Model" in infant retinal images. The novelty of this approach lies in its two-step strategy: it initially utilizes Adaptive Gamma Correction Color Preserving Framework for image enhancement, followed by the application of CATM. This approach performs the processing of red channel, in which the optic disc region is extracted from the red channel by eliminating non-optic disc pixel values. Furthermore, the method computes an adaptive threshold based on standard deviation of the Gaussian filter and enhanced red channel image ( ${\text{I}}_{\mathrm{RE}}$ ), resulting in the precise localization of optic disc region. The results on ROP dataset of 1103 retina images achieved a dice score of 0.8285, accuracy of 0.9894, precision of 0.9958, recall of 0.9875, and specificity of 0.9999. The experimental evaluation of retinal image dataset of ROP infants, consisting of low-contrast, sub-optimal illumination, and false reflections, represents a significant improvement in optic disc localization, thereby contributing valuable support for the early and reliable diagnosis of ROP.