Multi-view Deep Learning for Diabetic Retinopathy Detection
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2022-10-17
Department
Major/Subject
Data Science
Mcode
SCI3115
Degree programme
Master's Programme in ICT Innovation
Language
en
Pages
45+1
Series
Abstract
Diabetic retinopathy is one of the complications of diabetes, which is the main cause of vision loss. Recent development in deep convolutional neural networks (DCNN) has achieved high performance in diabetic retinopathy detection, which could serve to assist professionals to make early diagnoses and treatment for the patients based on their fundus images. However, the existing works have mostly focused on training the model with only single-view fundus images, ignoring the interaction as well as possible correlation across fundus images taken from different views of the retina of a patient, and thus might lead to a sub-optimal outcome. In this thesis, we propose a DCNN-based architecture to predict the degree of diabetic retinopathy from multi-view fundus images. Specifically, we utilize EfficientNet-B0 as the basic image feature extractor, and then fuse the features per patient's eye with three different fusion methods to get the summarized representation. We preprocess three benchmark datasets by dividing each fundus image into multiple views, and then randomly drop some views to simulate missing retinal images. Due to the ordinal labels, the quadradic weighted kappa score (kq) is used to measure the performance. Motivated by the clinical situation, the baseline model is defined with the maximum of predictions of single-view images per patient's eye. We make experiments on both simulated multi-view benchmark datasets and a clinical dataset with baseline model and our three fusion methods. Specifically, with an average of 30% images dropped inside fundus images set per patient, on the largest benchmark we used the mean fusion, max fusion, attention fusion methods achieved kq values of 81.92, 81.31, 81.10, respectively, while the baseline model had the kq value of 80.22. On the clinical dataset, the mean and max fusion had the performance values of 74.42 and 73.47 respectively compared to the baseline model's kq value of 71.66. However, the attention fusion method failed to converge on the clinical dataset.Description
Supervisor
Marttinen, PekkaThesis advisor
Kaski, KimmoSahlsten, Jaakko
Keywords
diabetic retinopathy detection, deep convolutional neural networks, supervised learning, multi-view deep learning