Predicting Diabetic Retinopathy in OCTA Images using a Deep Learning Framework

The demand for diabetic retinopathy (DR) screening is growing as the prevalence of diabetes mellitus (DM) rises. Optical coherence tomography angiography (OCTA) is a novel imaging technique that has shown promising results for the early diagnosis of diabetic retinopathy (DR). However, little researc...

Descripción completa

Detalles Bibliográficos
Autor: Zurita Martel, Yazmina
Tipo de recurso: tesis de maestría
Fecha de publicación:2023
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/385009
Acceso en línea:https://hdl.handle.net/2117/385009
Access Level:acceso abierto
Palabra clave:Diabetic retinopathy
Deep learning (Machine learning)
diabetic retinopathy
OCTA
deep learning
machine learning
artificial intelligence
diabetes
convolutional neural network
transfer learning
super-resolution
optical coherence tomography angiography
medical imaging
automated diagnosis
Retinopatia diabètica
Aprenentatge profund
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Descripción
Sumario:The demand for diabetic retinopathy (DR) screening is growing as the prevalence of diabetes mellitus (DM) rises. Optical coherence tomography angiography (OCTA) is a novel imaging technique that has shown promising results for the early diagnosis of diabetic retinopathy (DR). However, little research exists on automated frameworks for DR detection and grading based on OCTA data. In this study, we evaluate for the first time the effectiveness of a convolutional neural network (CNN) in classifying DM, DR and referable DR (RF-DR) using the largest dataset of labelled OCTA images currently available. The dataset is composed by 3×3 mm and 6×6 mm images of the superficial and deep capillary plexuses of 726 eyes. Three preprocessing schemes were implemented where either individual or concatenations of the original images or their super-resolved equivalents were used. The InceptionV3 architecture pretrained on the ImageNet dataset was fine-tuned for robust classification. After performing three times three-fold cross-validation, the proposed CNN classifier reported mean AUC values of 0.80, 0.81 and 0.92 for DM, DR and RF-DR diagnosis respectively. These results motivate further exploration of CNN-based classification of OCTA images to assist on DR diagnosis and alleviate the burden of mass screenings.