Automatic segmentation of outer retinal layers with Deep  learning in ophtalmology

João Duarte do Nascimento Dias Afonso

Key information

Authors:

João Duarte do Nascimento Dias Afonso (João Duarte do Nascimento Dias Afonso)

Supervisors:

Daniel Simões Lopes (Daniel Simões Lopes); Diogo Gonçalo Reis Cabral

Published in

November 29, 2024

Abstract

With the development of machine learning models and their application in medical fields like ophthalmology, biomedical imaging has significantly evolved, enabling rapid and reproducible identification and monitoring of various biomarkers. These advancements in ophthalmology can aid in distinguishing similar hereditary diseases, detecting early structural changes, and quickly assessing the integrity and thickness of different retinal layers, thereby alleviating the financial burden on clinics and hospitals. This thesis aimed to develop an automatic segmentation model for the outer retinal bands in spectral-domain optical coherence tomography(SD-OCT) images. Additionally, it sought to evaluate the reproducibility and performance of the developed algorithms by comparing their results with measurements performed by two retina specialists. Manual intra- and inter-observer variability was assessed for the following structures: the second hyperreflective band(Band 2), the combined third and fourth hyperreflective bands(Band 3+4), and the photoreceptor outer segments(POS). A total of 63 B-scans from 21 eyes were included in the study. Intraclass correlation coefficient(ICC) values for Band 2, Band 3+4, and POS were 0.737, 0.830, and 0.878, respectively, while interclass correlation coefficients were 0.104, 0.810, and 0.752. All deep convolutional neural network(dCNN) models achieved Dice scores exceeding 85%, with DRUNET delivering the best performance: 89.41% for Band 2, 93.25% for Band 3+4, and 91.42% for POS. The high ICC values reflect strong consistency in manual segmentation, though interclass correlation revealed some variability. DRUNET exhibited the most robust performance among the tested models, demonstrating its potential to reduce interclass variability and improve segmentation accuracy for these minute structures.