Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

EARLY DETECTION OF

GLAUCOMA FROM FUNDUS


IMAGES USING DEEP
LEARNING
Team Beta Blockers: Anamitra Das, Shiva Kumar Jadala, Mohit
Kukreja, Shikhar Panwar, and Sarah Parrett
Introduction
Symptoms:
- Peripheral vision loss
- Blind Spots
- Blindness
Current Methods
PROJECT
OVERVIEW
Assist in early glaucoma
detection by using deep
learning techniques for
optic cup and disc
segmentation and the
evaluation of the optic
cup and disc ratio.
Dataset
■ We have selected “Retinal Fundus
Glaucoma Challenge (REFUGE)" as
our dataset. This is a publicly available
dataset of retinal fundus images for
glaucoma classification, optic disc/cup
segmentation, and fovea
localization. This dataset is the largest
ethnicity-specific fundus images
database containing a total of 1200
images from two different sources: 400
from the Glaucoma Research
Community (GRC) and 800 from the
Singapore Eye Research Institute
(SERI). The annotations include the
optic disc and cup boundaries, as well
as the cup-to-disc ratio, which is an
important indicator of glaucoma.
METHODOLOGY
Step 1: Proper segmentation of the
optic disc (OD) and the optic cup
(OC).
■ Our model will be trained on the REFUGE dataset. The
segmentation model will take a color fundus image as an
input and output a grayscale image highlighting
accurate segmentations of the optic disc and cup. The
model will be based on the popular U-Net architecture for
medical image segmentation.
Step 2: Classification of Clinical
Glaucoma
■ A 2nd convolutional neural network will be used as an
image classifier, which inputs the grayscale output from
the previous model that outputs segmentations for the
optic disk and cup, and outputs a diagnosis for
glaucoma.
Step 3: Evaluation Of Deep Learning
Model

■ Using the REFUGE dataset will allow us to use the


model evaluation pipeline paired with it, so we can
benchmark our model’s performance against other state-
of-the-art models using similar approaches.
Step 4: Achieving SOTA Performance

■ We aim to achieve SOTA performance with our model


and make a robust pipeline for automated analysis of
retinal fundus images. By introducing the design of this
project, the proposal will advance the diagnoses of
glaucoma in the patient population
Results
• Model used: UNet (from MONAI)
• Currently only tested on 10% of dataset.
• Dice metric does not include background.
• Grayscale Labels split into 3 channels for each class
• Small ROI for validation sliding window.
■ Segmentation
Results
• Model used: DenseNet201 (From MONAI)
• Run for 10 epochs
• Resized to (128x128) - local GPU memory issues

■ Classification
DEMO!
THANK YOU
Questions?

You might also like