PROJECT TITLE :

Deep Neural Network Regression for Automated Retinal Layer Segmentation in Optical Coherence Tomography Images

ABSTRACT:

The quantification of layer information in early diagnosis of retinal disorders, the primary cause of permanent blindness, is made possible through the segmentation of retinal layers in optical coherence tomography (OCT) pictures. When it comes to protecting one's eyesight, segmentation is crucial. For this reason, experts in the field of ophthalmology are still required to physically segment the retinal layers. A feature-learning regression network without human bias is proposed in this paper for OCT picture segmentation. An image segment's intensity, gradient, and adaptive normalised intensity score (ANIS) are used as features for learning in the proposed deep neural network regression, which subsequently predicts a retinal boundary pixel. According to this computational complexity analysis, a regression problem can be transformed into one that does not require a large dataset and thereby minimises the computing complexity. It also works well on OCT pictures with low contrast, speckle noise, and blood vessels, but stays accurate and time efficient thanks to the help of ANIS. It only took 30 seconds of training for each boundary line to identify eight boundaries utilising 114 images in the method's evaluation. The processing time for each image was roughly 10.596 s. Dice similarity coefficients for testing accuracy produced a computed value of 0.966, which is consistent with the results of this study. There was an average of less than a one-pixel difference between human and automatic segmentation utilising the proposed approach.