Abstract
Purpose :
Diabetic retinopathy, an eye complication caused by diabetes, is a leading cause for blindness in the U.S. In the initial stages of diabetic retinopathy, patients are generally asymptomatic. With treatment in the early stage of non-proliferative diabetic retinopathy, the risk of blindness is reduced by 95%.
Current research has a broad scope and aims to analyze all stages of diabetic retinopathy, using methods generalized for all stages of diabetic retinopathy. This project addresses the research gaps with the hypothesis that stage optimized image processing can improve deep learning based diagnosis of early stage diabetic retinopathy.
Methods :
1. Trained 4 proven deep learning CNN models using adaptive step size: Resnet50, VGG16, Inception v3, EfficientB0.
2. Evaluated the image categorization accuracy, sensitivity, specificity, loss, and confusion matrix for each model to generate baseline metrics.
3. Evaluated the effects on the same metrics of image augmentation, varying image resolution, and various image processing methods on accuracy, sensitivity, and specificity for each of DR stages 0-4
Results :
Resnet50 provided the best baseline results and was used for studying various image processing methods. Using image augmentation, 512 by 512 pixel size, and CLAHE2 applied to grayscale; system accuracy, sensitivity, and specificity were 83.5%, 83.7%, and 97.9% respectively. Missed diagnosis of early stage DR was reduced ~80% from 6.58% to 1.32% using the same optimization techniques. Through stage specific analysis, LAB/CLAHE optimized results for Stages 0, 2, and 4, median filtering optimized results for Stage 1, and CLAHE applied to the green component with grayscale optimized results for Stage 3.
Conclusions :
The image processing method that provided the best accuracy, sensitivity, and specificity was identified for each stage of DR. Additionally, an image processing method that provided the best metrics for all stages and minimized missed diagnosis of stage 1 as normal was identified.
This is a 2021 Imaging in the Eye Conference abstract.