Automated Detection of Diabetic Retinopathy using Deep Learning

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Diabetic retinopathy is a major cause of vision loss in working adults.
• Early detection is crucial for effective treatment and prognosis.
• Automated detection systems can aid in screening and diagnosis.

Purpose of the Study:

• To evaluate the efficacy of convolutional neural networks (CNNs) for diabetic retinopathy staging using color fundus images.
• To investigate methods for improving CNN performance, particularly in detecting subtle disease features.
• To assess the impact of preprocessing techniques and transfer learning on classification accuracy.

Main Methods:

• Development and application of CNN models for diabetic retinopathy classification.
• Utilized color fundus images for training and validation.
• Explored preprocessing techniques like contrast limited adaptive histogram equalization.
• Implemented transfer learning using pretrained GoogLeNet and AlexNet models.
• Ensured dataset fidelity through expert verification of class labels.

Main Results:

• CNN models achieved validation sensitivity of 95%, comparable to existing literature.
• Multinomial classification models struggled with mild disease detection, often misclassifying it as normal.
• Preprocessing and dataset verification improved recognition of subtle diabetic retinopathy features.
• Transfer learning boosted peak test set accuracies to 74.5% (2-class), 68.8% (3-class), and 57.2% (4-class).

Conclusions:

• CNNs are effective for diabetic retinopathy staging, but subtle feature detection remains a challenge.
• Preprocessing and transfer learning significantly enhance the performance of CNNs for this task.
• Further research is needed to address limitations in detecting early-stage diabetic retinopathy.