Deep Convolutional Neural Networks for Diabetic
Retinopathy Classification
Chunyan Lian, Yixiong Liang*, Rui Kang, Yao Xiang
School of Information Science and Engineering
Central South University
Changsha, China
yxliang@csu.edu.cn
ABSTRACT
Diabetic retinopathy (DR) is one of the leading cause of blindness,
but the classification of DR requires experienced ophthalmologist
to distinguish the presence of various small features, which is time-
consuming and difficult. Convolution neural network (CNN),
which enables learning hierarchical and discriminative features
without experiences of clinicians, is an alternative method to
address the above issue. In this paper, we investigate four factors
of employing deep CNN to DR classification problem, including
network architectures, preprocessing, class imbalance and fine-
tuning. The best performance we achieved is an accuracy of 79%
opposed to an accuracy of 75% for Harry’s scheme [14].
CCS Concepts
Computing methodologies → Artificial intelligence →
Computer vision →Computer vision tasks →Biometrics.
Keywords
component; Diabetic Retinopathy, Convolutional Neural Networks,
Deep Learning, Medical Image Classification, Digital Fundus
Images
1. INTRODUCTION
Diabetic Retinopathy (DR), a main complication caused by
diabetes, is one of the most common and severe eye diseases
causing blindness in the world. It is reported that DR is the leading
cause of blindness among adults for age 20-64 [1]. Early diagnosis
is crucial to slow down the progression of DR and prevent the
emergence of blindness [2-3]. Normally, diagnosing a DR patient
is performed manually by ophthalmologist, which is a time-
consuming work and also tends to make mistakes. Therefore,
automation of the retinal exam for DR is sorely needed.
Many automated DR classification algorithms using fundus image
have been studied over the past decade [4-6]. They are mainly
based on the ophthalmologist’s experiences and manually designed
features. In these algorithms, morphological features [4], blood
vessels, Microaneurysm (MA), and hard exudation HEM features
[5] are extracted to classify DR. These hand-craft features are
conducted on small data sets and the results of these algorithms
tend to over fit, therefore, the generalization of these models are
poor in actual scenes.
Convolution neural network (CNN), has substantially
outperformed the state-of-the-art in computer vision tasks such as
classification, detection and segmentation [7-9]. It has also
achieved remarkable performance in medical applications [10-12].
Some attempts have been done to employ CNN for DR
classification in large dataset [13-14]. These algorithms make
significant progress but there is still much margin for improvement.
In this paper, we extensively evaluate four factors of employing
deep convolutional neural networks, which are CNN architectures,
preprocessing, class imbalance and fine-tuning for DR
classification. We explore different CNN architectures, describe
the influence of class imbalance and preprocessing on performance,
and verify the impact of fine-tuning from pre-trained natural image
in medical image classification. In addition, we achieve remarkable
performance that outperforms Harry’s scheme [14] in terms of
accuracy.
2. PRELIMINARIES
2.1 Dataset
Figure 1. Some examples of fundus images with different DR stages in the
Kaggle dataset[15]. (a): normal; (b): mild NPDR; (c): moderate NPDR;
(d):severe NPDR; (e): PDR;
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. Copyrights
for components of this work owned by others than ACM must be
honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior
specific permission and/or a fee. Request permissions from
Permissions@acm.org.
ICAIP'18, June 16–18, 2018, Chengdu, China
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-6460-7/18/06…$15.00
DOI: https://doi.org/10.1145/3239576.3239589