This document outlines a process for detecting diabetic retinopathy through image processing and ensemble learning. Fundus images are first enhanced through techniques like extracting the green channel and histogram equalization. Blood vessels and exudates are then extracted as features. Multiple classifiers are trained and their predictions are aggregated to classify images as having retinopathy or not. The model achieves an accuracy of 74.42% on a test set. Areas for future work include improving feature extraction and optimizing the features and models used.

1. Diabetic Retinopathy Detection
December 10, 2015

2. Page 2
Outline
Modeling Process
Image Processing
Feature Extraction
- Blood vessels
- Exudates
Ensemble Learning
Future steps

3. Page 3
Modeling Process
Retinopathy grades 0 and 1 are considered
as class 0 (No-Retinopathy) and
grades 2 and 3 as class 1 (Retinopathy)
Input
Data
Image-
Texture
Based
Features
Blood Vessels
and Exudates
Ensemble Learning
Classify each image as retinopathy or not-retinopathy
using multiple classifying algorithms
Image Processing
Number of features used : 58

4. Page 4
Image Processing
Enhance Fundus image
- Extract green channel
- Grayscale conversion
- Image processing techniques
Original
Image to
Enhanced
Image

5. Page 5
Blood Vessels Extraction

6. Page 6
Exudates Extraction
Original Color Image
Extracted Green
Channel
Contrast
Enhancement -
Adaptive Histogram
Equalization
Labeled Exudates
Binary Mask using
thresholding
Smoothing –
Gaussian Filtering

7. Page 7
Ensemble Learning
 Construct a set of classifiers for input images
 Predict retinopathy grade of previously unseen images by aggregating
predictions made by multiple classifiers (e.g. by using weighted voting)

8. Page 8
Classification Results
Metrics Formula
Value
(%)
Accuracy Correctly Identified Cases / Total Number of Cases 74.42
Sensitivity Correctly Identified Retinopathy Cases / Total Retinopathy Cases 77.60
Specificity Correctly Identified No-Retinopathy Cases / Total No-Retinopathy Cases 70.59
Positive Predictive
Value
Correctly Identified Retinopathy Cases / Total Identified Retinopathy Cases 76.24
Negative Predictive
Value
Correctly No- Identified Retinopathy Cases / Total No- Identified Retinopathy Cases 72.57

9. Page 9
Classification Results : ROC Curve
Area Under ROC
Curve
82.48%
The bigger the area under ROC
curve, the better!

10. Page 10
Correctly Classified Images
No Retinopathy
(95.69 % confidence)
Truth : Retinopathy Grade 0 (No-Retinopathy)
Retinopathy
(95.70 % confidence)
Truth : Retinopathy Grade 3 (Retinopathy)

11. Page 11
Misclassified Images
No Retinopathy
(83.75 % confidence)
Truth : Retinopathy Grade 2 (Retinopathy)
Retinopathy
(95.50 % confidence)
Truth : Retinopathy Grade 0 (No-Retinopathy)

12. Page 12
Future Work
• Improvements in blood vessels extraction, exudates extraction, etc.
• Features based on hemorrhages, macula, optical disk, etc.
Image processing & Feature computation
• Optimize features used for classification
Feature selection
• Model tuning
• Different techniques of ensemble learning
Model selection & Ensemble Learning
• Trying the model on different dataset / Addition of new images as input
• Similar analysis for multiclass problem
• Change in threshold for categorizing as retinopathy
Further Analysis

13. Page 13
Thank You