This document describes a computer-based image processing approach to quantify acellular capillaries in retinal images of control and diabetic mice in order to assess diabetic retinopathy. The approach uses Python programming and open source packages to preprocess retinal images using techniques like background separation, blurring and segmentation. It then applies a medial axis transform to skeletonize blood vessels and identify branch points. Post-processing filters are applied to remove noise before identifying and counting acellular capillaries based on branch width and length thresholds. The results showed the program was able to automatically count acellular capillaries.

1. Computer-based Quantification of Acellular Capillaries to
Assess Experimental Diabetic Retinopathy
Hemmady, Anish1,2;Tuceryan, Mihran1; Bhatwadekar, Ashay D2
1) Department of Computer Science, IUPUI
2)Department of Ophthalmology, Indiana University
Introduction
Diabetic retinopathy (DR) is a
disease of small blood vessels in the
retina. The increase in the number
of acellular capillaries is used as a
marker to assess the severity of DR.
The traditional approach for
identifying acellular capillaries is
manual counting of the capillaries
either directly under the microscope
or using the captured images.
However, these methods are
cumbersome and often involve
inconsistencies among researchers.
The purpose of this study is to
reduce discrepancies in
enumeration of acellular capillaries
using computer-based image
processing algorithms.
Frameworks
• The retinas of control and
diabetic mice were processed
using trypsin digestion.
• The high resolution png format
images of retinal quadrants
were prepared from the trypsin
digested retina.
• The computer programming
was performed using the Python
language along with open source
packages such as
• OpenCv,
• Python Imaging Library
(PIL)
• NumPy (Numerical Python)
• SciPy
Preprocessing Steps
 Initially background separation was performed on input
image.
 These images were then corrected for a Gaussian Blur
and a Median blur to remove noise
(input image) (background separated image)
Methods and Results
Histogram based Image Segmentation
 Image segmentation on corrected images was performed
using a histogram analysis through which a binary image
was generated.
 The segmentation threshold for binary image was
determined using histogram peak-and-valley detection
method.
(input image) (segmented image)
Medial Axis Transform
Medial axis Transform uses the binary image as input
image. The MAT skeletonizes the blood vessels and is used
to detect branches and branch-points in those blood vessels.
As part of the MAT computation, the distances from the
skeleton to the vessel boundaries are encoded.
Skeleton from Medial Axis Transform
Post Process Filtering
 Post process filtering was undertaken to remove resulting
imperfections in Segmented Image.
 Using this approach the image undergoes Morphological
operations of Opening/closing which removes the noise by
expanding and shrinking image pixels.
 This removes the holes in images and smooths the
boundaries of foreground regions.
Branch Point Identification
 Branches were identified by segmenting the MAT into
continuous curves thus segmenting the vessels and
identifying candidate acellular capillaries. Branch points
were identified first by finding pixels with more than
three neighboring pixels that are separated by a blank.
Each branch is then constructed by starting at a branch
point and adding pixels until another branch point is
reached.
 This process is repeated until all branches were built. It is
a recursive process
Identification of
Acellular Capillaries
• A filter was applied to the branches
of the image to identify which ones
are acellular capillaries.
• Each branch was checked to see if it
fits the criteria of a branch with
more than 2 neighboring foreground
pixels. Two thresholds are used: one
threshold is for width and other is
for length. Very short branches are
filtered out as noise. The narrow
branches were kept as an acellular
capillary.Counting of Acellular
Capillaries was performed using
Connected Component Labelling on
the identified branches.
(Final Output) (Count displayed)
Results
The program was automatically able to
count the number of acellular
capillaries.
Sponsors
Thanks to International Retinal Research
Foundation Inc.(IRRF) for their support.