This project investigates techniques for improving microaneurysm segmentation in diabetic retinopathy fundus imaging, where lesions are extremely small and often low-contrast relative to surrounding tissue. We experiment on an adapted version of a previously established segmentation model for diabetic retinopathy, LA-Net, into HydraLA-Net.
Read the Paper HERE!
Visit our Deployed App HERE!
Jessica Yuan - Technical Project Manager
Michael Liu - Experimental Design, Model Architecture & Modifications, Model Training (WATGPU via SSH), Loss Functions, Dataset Curation, Preprocessing & Augmentations, Documentation (README.md), Consulted with UWaterloo School of Optometry, Technical Diagrams & Visualizations
Andrew Yang - Experimental Design, Model Training (WATGPU via SSH), Model Architecture,Literature Review and Research on Related Works, Dataset Curation, Research Paper Writing
Sidharth Shah - App Deployment, Research Paper Writing
Christopher Risi - Technical Support
William Chiu - Research Paper Writing, Loss Function Development & Analysis
Tom Almog
Apisan Kaneshan
This repository contains work in progress on the semantic segmentation of microaneurysms, hemorrhages, soft exudates, and hard exudates (lesions resulting from Diabetic Retinopathy) from fundus images. In addition to building the full segmentation pipeline, the project also conducts experimentation on techniques for enchancing the detection of microaneurysms.
Our research primarily focuses on:
- Contrast enhancement strategies to improve lesion visibility, including channel-aware selective preprocessing (CASP) and local contrast normalization (CLAHE).
- Training-time techniques to improve sensitivity to small structures, including loss functions designed to emphasize microaneurysm recall.
- Dataset Curation
- Preprocessing and Augmentation Pipelines
- Loss Function Development
- Data Visualizations
- HydraLA-Net Implementation in PyTorch
- Training Scripts
- Part 1: Baseline Training
- Part 2: Preprocessing Variation Analysis
- Part 3: Class Imbalance Aware Loss Function Analaysis
- Part 4: Individual Dataset Analysis
- Get results on testing sets
- Publish research paper!
The segmentation architecture used in this project is based directly on the original research paper that introduced it.
- LA-Net Paper: Lesion-Aware Network for Diabetic Retinopathy Diagnosis
The GitHub repository containing the experiment conducted in the paper above can be found at: LANet-DR GitHub Repo
Additionally, a full dynamic implementation of our adapted HydraLA-Net can be found in this repository.
Three datasets are chosen for the project. All datasets contain fundus images and segmentation masks for microaneuryms, hemorrhages, soft exudates, and hard exudates. The IDRiD and DDR datasets contain the masks in the binary form. The TJDR dataset contains the masks as a image with color mappings.
- IDRiD: Indian Diabetic Retinopathy Dataset from Kaggle
- DDR: Diabetic Retinopathy Lesion Segmentation and Lesion Detection Dataset from GitHub
- TJDR: TJDR: A High-Quality Diabetic Retinopathy Pixel-Level Annotation Dataset from GitHub
Dataset Pixel Distributions by Class
