An Implementation of Diabetic Retinopathy Grading
System on Android Device Using Fully Automated
Algorithm
Duc Ngoc Tran
Computer Engineer Faculty
University of Information Technology
Ho Chi Minh, Vietnam
ductn@uit.edu.vn
Abstract— Diabetic retinopathy (DR) is a diabetic eye disease
which causes damage to the blood vessels in the retina. Most of
studies are applied in MATLAB environment that is designed for
research only. For solving these problems above, a new prototype
of Android based DR grading system will be introduced in this
study. The proposed design architecture emphasizes and
enhances the computing power usage efficiency of DR grading
system in ARM platform running Android. The novelty lies in
the idea of creating MATLAB-like environment using IT++
library to ease the conversion from MATLAB source code to
C++. Another contribution is applying the multithreading
technique to image processing module to decrease the execution
time in Android OS. This approach have improved the system in
terms of execution time with 14.5 seconds.
Keywords—Diabetic retinopathy, Android, multithreading
I. INTRODUCTION (HEADING 1)
Diabetic retinopathy (DR) is the common diabetic eye
disease that eventually develops to some degree in almost all
long-standing diabetic patients. DR severity is divided into five
levels including no DR, mild non-proliferative diabetic
retinopathy (NPDR), moderate NPDR, severe NPDR and
proliferative diabetic retinopathy (PDR) [1].
In previous research, CISIR in UTP has developed
successfully the automatic computerized DR Grading system
running in MATLAB [2]. This current system has achieved
many advantages such as high accuracy, quick diagnosis, low
cost and suitable for diabetic patient. However, the previous
computerized DR system requires a personal computer to
operate, so it is cumbrous machine, which is less mobility than
other portable machine like popular modern tablet or
smartphone. Another disadvantage is the requirement of the
environment to use this system that is MATLAB. This
environment is only designed to use in a laboratory in which
the researchers can easily develop and modify the algorithm in
research and development stage (R&D stage). Thus, this
system is not suitable for mass screening of a large number of
diabetic patients nationwide.
Based on several reasons above, a portable, mobile and
standalone DR system is very necessary. In this study, a new
XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
prototype of standalone and portable computerized DR grading
system using fully-automated DR algorithm will be introduced.
II. LITERATURE REVIEW
A. Diabetic Retinopathy Diagnosis System
Over the last decade, there was a rapid development of
automated DR pathology detection systems. These automated
diagnostic systems are based on several types of lesions of the
eye to detect certain pathologies such as microaneurysms,
hemorrhages, exudates, cotton wool spots and distances
between the exudates and foveas [3]. Another approach for
automated grading system is the analysis of Foveal Avascular
Zone (FAZ). Foveal avascular zone is a circular area of retina
which is responsible for central and color vision and devoid of
capillaries in the macular region [4]. In diabetic retinopathy,
the loss of capillaries in the perifoveal capillary network make
an enlargement of the FAZ [5]–[7]. A group of researcher at
the Centre for Intelligent Signal and Imaging Research (CISIR)
in Universiti Teknologi Petronas studied a method for
measurement of FAZ based on digital color fundus images. In
these studies, the increase of FAZ is strongly correlated to the
progression of DR from normal, mild NPDR, moderate NPDR,
and severe NPDR. This promising approach will be widely
used in the future due to its advantages such as ease of
operation, low cost and quick diagnosis.
This approach of automated DR grading system includes
three main processes, i.e. contrast enhancement of retinal
vessels, segmentation of retinal vessels and DR grading based
on analysis of FAZ [2] as shown in Figure . In the first process,
Retinex and Independence Component Analysis (ICA) are
used to conduct the contrast enhancement. The output of this
process is fundus image, which has high contrast retinal blood
vessels. The second process is used to segment the retinal
blood vessels in the fundus image. First, the blood vessel
detection algorithm named Chaudhuri is utilized by applying a
series of digital filters. The segmentation of retinal blood vessel
is generated by automatic region growing method. In the last
process, the blood vessel endpoints at perifoveal capillary are
detected using segmentation result in previous process. The
foveal avascular zone area is determined by detecting all
nearest points to the center of macula. The size of FAZ is the
size of a specific area, which is formed by connecting these
detected points encircling the perimeter of macula. Radius of
FAZ, which is defined as half of the longest line that
connecting to detected points of FAZ, is also calculated to give
another parameter for DR diagnosis.
Fig. 1. Flowchart of fully automated DR algorithm
B. Development Platform
To implement a mobile system for the DR Grading system,
the market include many available embedded high
performance platforms including ARM-based MCU, MIPSbased MCU, Power Architecture-based MCU or x64/x86based MCU. On these platforms, ARM has a fast growth speed
and combine many good characteristics for an effective
embedded system such as low power consumption, low cost
and high performance. ARM architecture is chosen to develop
the portable DR grading system in this project.
To build a good system, the DR grading system need a
potential OS to make a good interaction with the user. Using
Linux kernel with many libraries and APIs written in C,
Android provides an open-source and convenient platform for
implementation and maintain. As a Google’s product, Android
has a major role in the direction of the embedded industry. By
using Android OS in this project, a vast of benefits as above
can be achieved for this product and continuous improvement
ability for the next version of portable DR grading system.
C. Project Library Implementation
The DR algorithm requires many subroutines for most
basic algorithms in numerical linear algebra such as matrix
factorization, systems of linear equations, matrix
multiplication. Blount used JLAPACK, a subset of the
LAPACK library in Java, library to implement high
performance numerical software in Java [8]. When used with
the native BLAS library, JLAPACK performs comparably with
the Fortran version of LAPACK using the native BLAS
library. They found that the Java version is about two or three
times worse than the Fortran version on the larger problem
sizes for both the factorization and the triangular solve. The
interpreted Java implementation was unusably slow. In
addition, the performance when the native BLAS library was
used increased significantly. This demonstrates that the native
library provides better performance. Mouton made a study of
the existing linear algebra libraries that can be used from C++
[9]. This study showed that LAPACK library is essential
library for linear algebra libraries. It provides high performance
linear algebra computation for many scientific computing
softwares.
To implement the algorithm in MATLAB to Android OS
efficiently, a library of mathematical and signal processing is
very necessary. IT++ is a C++ library of mathematical, signal
processing and communication classes and functions. Its main
use is in simulation of communication systems and for
performing research in the area of communications. The kernel
of the library consists of generic vector and matrix classes, and
a set of accompanying routines. Such a kernel makes IT++
similar to MATLAB [10]. Cristea has presented an extension
of the IT++ library useful for the study of turbo receivers [11].
In this research, the advantage of using a compiled language
(C++) for complex algorithms implementation was clearly
demonstrated. The IT++ library, with its MATLAB-like kernel,
is in their opinion a good replacement for MATLAB and other
interpreted languages used by scientists and engineers as
simulations tools.
III. IMPLEMENTATION
The Figure 2 shows the design of portable DR grading
system. In this design, the computing component comprises a
mathematical library module and an IT++ library module. The
mathematical library module implementing LAPACK (Linear
Algebra PACKage) and BLAS (Basic Linear Algebra
Subroutine) algorithms are placed at the lowest layer in the
design as essential libraries for other modules in the upper
layers. Since it is a standard mathematical library, the results
provided by the lower layer are reliable. The mathematical
library module provides a set of routines such as matrix
multiplication, solving linear equations and matrix
decomposition for modules placed at the upper layers which
are actual application modules such as Retinex, Independent
component analysis (ICA), Region growing and Chaudhuri’s
algorithm of the grading component. It will use all the essential
classes and functionalities, which are provided by lower layers,
to implement the algorithm. Thus, if there is any change in
design or upgrading, only the upper layer modules need to be
modified, the lower layer module remains stable. This will save
a lot of time for debugging the lower layer. Based on LAPACK
and BLAS, IT++ library module is capable of providing many
useful signal processing functions. IT++ library module is an
open source C++ library of mathematical, signal processing
and communication classes and functions. Its main use is in
simulation of communication systems and for performing
research in the area of communications. The IT++ library
component implements a Fast ICA package and a Kurtosis
algorithm. The kernel of the IT++ library component consists
of generic vector and matrix classes, and a set of
accompanying routines. Such a kernel makes IT++ library
component similar to MATLAB. Thus, a MATLAB-like
environment with IT++ is created to ease the conversion from
MATLAB source code to C++. The IT++ library component of
the present design, which employs the native C code, achieves
much better performance than using the same algorithm
running on Java code. Although IT++ is not designed for
Android platform, it can run in Android with modification in
source code related to Android.mk file and IT++ source code.
Fast ICA and Chaudhuri modules are time critical function,
which accounts total 86% of execution time in DR grading
system. Thus, these modules will be investigated to give
appropriate solutions for optimization.
Fig. 3. Android based DR grading system profile
B. Retinex Optimization
In DR Grading system algorithm, three channels of the
image is undergone the Retinex processing sequentially. Three
results after this processing are used as inputs to ICA module.
Because the processing of each channel is independent with
each other’s, the multithreading technique can be applied in
this case to improve performance by utilize multi-core
architecture in modern ARM processor as Figure 4 below.
Fig. 2. Android based DR grading system design
This template, modified in MS Word 2007 and saved as a
“Word 97-2003 Document” for the PC, provides authors with
most of the formatting specifications needed for preparing
electronic versions of their papers. All standard paper
components have been specified for three reasons: (1) ease of
use when formatting individual papers, (2) automatic
compliance to electronic requirements that facilitate the
concurrent or later production of electronic products, and (3)
conformity of style throughout a conference proceedings.
Margins, column widths, line spacing, and type styles are builtin; examples of the type styles are provided throughout this
document and are identified in italic type, within parentheses,
following the example. Some components, such as multileveled equations, graphics, and tables are not prescribed,
although the various table text styles are provided. The
formatter will need to create these components, incorporating
the applicable criteria that follow.
IV. SYSTEM OPTIMIZATION
A. Android based DR Grading System Analysis
Figure 3 shows the profile of Android based DR Grading
system for ARM architecture. It demonstrated that Retinex,
Fig. 4. Multithreading optimization in Retinex
Ten images from FINDeRS database are chosen randomly
to evaluate the performance of Retinex module. Xiaomi Mi 5
with Snapdragon 820 quadcore 1.8GHz is used to as the
experiment platform. In general, multithreading technique can
improve performance in Retinex with speedup factor 1.91x
C. Blood Vessel Detection (Chaudhuri’s Filter) Optimization
In the previous study, an optimization of blood vessel
detection in retina images has been studied [12]. This
algorithm is integrated to DR Grading system to gain more
performance.
The performance of native code achieves about 92.77%
faster Java code after applying multithreading method. The
multithreading technique can bring 1.9x performance gain for
native code.
be further analyzed using relationship plots of DR-Android
against DR-MATLAB data as shown in Figure 5.
V. RESULT & DISCUSSION
The experiments have been performed to validate the
design and optimization of Android based DR Grading System.
Firstly, resource consumption of the system is benchmarked in
Xiaomi Mi5 which is configured to run on Android 6.0 and
Snapdragon 820 quadcore 1.8Ghz.
A. Resource Consumption
A total of 100 images samples in FINDeRS database have
been undergone to the Android based DR Application to obtain
the CPU and Memory usage of each image. The mean CPU
and Memory usage of 100 images is the average CPU and
Memory usage of the DR Application on the Mi5 phone.
Fig. 5. FAZ radius obtained by the DR-Android against DR-MATLAB
C. Execution time comparison
TABLE I.
CPU Usage
Memory Usage
AVERAGE CPU AND MEMORY USAGE OF THE DR
APPLICATION ON THE MI5
Resource Consumption
67.52 %
148479 KB
The Table I shows the average CPU and memory
consumption of Android based DR Grading running on
PandaBoard. During the experiment to measure the CPU and
memory usage, the application can use 99% CPU resource in
Chaudhuri’s filters module and the memory usage is always
below 150 MB. This result proved that the multithreading
technique, which has been applied in Chaudhuri’s filter,
Retinex module and Fast ICA module module, utilized all
available resource and increase the efficiency of CPU
utilization.
B. Diagnosis Accuracy Comparison
From the FINDeRS database, 100 fundus images are
chosen to undergo to the DR Grading system in MATLAB
(DR-MATLAB) and Android (DR-Android) to collect results
in both system
STATISTICS OF ERROR OF FAZ RADIUS FOR THE ACCURACY
ANALYSIS OF THE FULLY AUTOMATED DR SYSTEM IN MI5
In the following experiment, several test cases are
conducted to measure execution time in three devices including
Xiaomi Mi5 and PC (MATLAB). The PC, which is used in this
experiment, has Core i7 3.0 GHz with 8 cores, 8 GB RAM.
TABLE III.
EXECUTION TIME COMPARISON BETWEEN TWO DEVICES
Android
Xiaomi Mi5
No. of Fundus images
Average execution time
(second)
PC (MATLAB)
100
100
14.5
11.2
Although PC with MATLAB still can complete the
algorithm in 28.35 seconds which is faster than Mi5, it is not a
significant difference with the performance gap between PC
and Mi5.
The regression analysis of Figure 5.2 shows strong linear
relationship of DR-Android system with DR-MATLAB. It is
indicated by their most perfectly linear regression line.
TABLE II.
DR-MATLAB –
DR-Android
No. of fundus image
samples
Mean absolute error
(pixels)
Std. deviation (pixels)
Lower bound (pixels)
Upper bound (pixels)
100
2.63
2.29
0.32
7.91
From 100 observations, the maximum error produced by
the DR system in Android when using FAZ radius is less than
8 pixels with a significantly small standard deviation (less than
3 pixels). It indicates that the DR system in Android has very
limited difference with the DR system in MATLAB. This can
VI. CONCLUSION
Android based DR Grading System, a portable standalone
diagnosis devices for DR, has been development successfully.
The evaluation is confirmed that DR Grading in Android is
very similar with the DR in MATLAB with significant small
the mean absolute error (FAZ radius: 2.63). The design and
optimized techniques has been utilized the hardware resources
effectively to reduce the execution time. The techniques
includes native code, multithreading. Thus, the final version of
Android based DR grading system had a lot of improvement in
terms of execution time with 14.5 seconds.
ACKNOWLEDGMENT
The research is funded by Computer Engineering Faculty in
project D2016-03, University of Information Technology –
Vietnam National University.
[7]
[8]
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
“American Academy of Ophthalmology. Preferred Practice Patterns :
Diabetic Retinopathy PPP - September 2008,” Am. Acad.
Ophthalmol., 2008.
A. F. M. Hani et al., “Toward a Fully Automated DR Grading
System,” vol. 37, Springer Berlin Heidelberg, 2012, pp. 663–666.
O. Faust, “Algorithms for the Automated Detection of Diabetic
Retinopathy Using Digital Fundus Images: A Revi,” J. Med. Syst.,
2010.
G. Richard, G. Soubrane, and L. Yanuzzi, Fluorescein and ICG
Angiography. New York: Thieme Medical Publisher Inc, 1998.
J. Conrath et al., “Semi-automated detection of the foveal avascular
zone in fluorescein angiograms in diabetes mellitus,” Clin.
Experiment. Ophthalmol., vol. 34, no. 2, pp. 119–123, 2006.
G. H. Bresnick, R. Condit, S. Syrjala, M. Palta, A. Groo, and K.
Korth, “Abnormalities of the Foveal Avascular Zone in Diabetic
[9]
[10]
[11]
[12]
Retinopathy,” Arch. Ophthalmol., vol. 102, no. 9, pp. 1286–1293,
Sep. 1984.
J. Conrath, R. Giorgi, D. Raccah, and B. Ridings, “Foveal avascular
zone in diabetic retinopathy: quantitative vs qualitative assessment,”
Eye, vol. 19, no. 3, pp. 322–326, Jul. 2004.
B. Blount and S. Chatterjee, “An Evaluation of Java for Numerical
Computing,” in Computing in Object-Oriented Parallel
Environments, vol. 1505, D. Caromel, R. Oldehoeft, and M. Tholburn,
Eds. Springer Berlin / Heidelberg, 1998, pp. 501–502.
C. Mouton, “A study of the existing linear algebra libraries that you
can use from C++ (Une ’etude des biblioth\`eques d’alg\`ebre lin’eaire
utilisables en C++),” Comput. Res. Repos., vol. abs/1103.3021, Mar.
2011.
Free Software Foundation, “IT++ - Scientific Library.” [Online].
Available: http://itpp.sourceforge.net/current/index.html. [Accessed:
09-Sep-2012].
B. Cristea, “Turbo receivers with IT++,” in Proceedings of the 2nd
International Conference on Simulation Tools and Techniques, ICST,
Brussels, Belgium, Belgium, 2009, p. 17:1–17:7.
D. N. Tran, F. A. Hussin, and M. Z. Yusoff, “Optimization of blood
vessel detection in retina images using multithreading and native code
for portable devices,” in 2013 IEEE 9th International Colloquium on
Signal Processing and its Applications, 2013, pp. 194–198.