Predicting Diabetic Retinopathy Progression with Time-Series Databases and Deep Learning

Overview

This repository contains work in progress on predicting diabetic retinopathy progression using segmentation and model blending techniques. The project is moving toward a paper tentatively titled:

“Predicting Diabetic Retinopathy Progression with Time-Series Databases and Deep Learning.”

---

Our current Kaggle Notebook being used

• Notebook: Link

---

Teams

Segmentation Team

• Focused on analyzing medical images to isolate important regions such as lesions, optic discs, and other retinal structures.
• This step improves the precision of downstream models by providing clean, labeled inputs for training and evaluation.

Model Blending Team

• Involves integrating the predictions of multiple deep learning architectures to enhance overall accuracy and robustness.
• By combining results from different models and datasets, the blended approach aims to achieve more reliable progression prediction and classification outcomes.

---

Datasets

• DeepDRiD – GitHub
• TJDR – Pixel-level lesion annotations for 561 fundus images.
• APTOS 2019 Blindness Detection – Kaggle

---

Relevant Papers:

1. Gulshan et al., 2016 – Deep Learning for Diabetic Retinopathy Detection
2. Dai et al., 2024 – A Deep Learning System for Predicting Time to Progression of Diabetic Retinopathy
3. Qiao et al., 2025 – Forecasting the Diabetic Retinopathy Progression Using Deep Learning
4. DeepDRiD: Automated Machine Learning for Diabetic Retinopathy Diagnosis and Grading
5. APTOS 2019 Blindness Detection – Kaggle 1st Place Solution Summary
6. Nature Scientific Data 2025 – Fundus Image Segmentation Reference

---

References

• Nature Scientific Data 2025 – Fundus Image Segmentation
• APTOS 2019 1st Place Solution Summary