Project Description
I am building a research-grade pipeline that focuses first and foremost on Diabetic Retinopathy (DR) grading, while keeping an eye on downstream extensions such as DME detection, retinal biomarker discovery, longitudinal modelling, cross-dataset generalisation and explainable retinal AI.
Data at hand
• Thousands of colour fundus photographs sourced from OLIVES, MMRDR and my own curated set.
• Structured patient medical history accompanying a subset of those images.
Core objectives
1. Train and validate a model that assigns the correct severity level to every DR image.
2. Surface early-stage indicators that could alert clinicians before conventional thresholds are reached.
3. Provide per-patient progression tracking so that sequential visits can be plotted and forecast.
Technical expectations
You are free to choose your preferred deep-learning stack (PyTorch, TensorFlow, JAX, etc.) as long as the final code is fully reproducible and runs on a single high-memory GPU. Class-activation maps, attention heatmaps or similarly intuitive visualisations must accompany each prediction to satisfy the explainability requirement. Where possible, strategies that encourage cross-dataset robustness—domain adaptation, stain normalisation, self-supervised pre-training—should be incorporated.
Deliverables
• Well-documented source code and environment file
• Trained weights and an inference script that accepts batches of images (optionally with tabular metadata) and outputs severity grade, early-warning flags and progression curve updates
• A concise report summarising method, experiments on OLIVES + MMRDR splits, confusion matrices and visual explanations
• Short video or PDF walkthrough showing how to reproduce results end-to-end
Acceptance criteria
The model must reach or exceed state-of-the-art F1 on internal test sets and maintain no worse than 5 % performance drop when applied to a hold-out external dataset. Explanatory visualisations should be clinically interpretable by an ophthalmologist.