01Final Year ProjectLive research demo
RetinoScope
A multimodal retinal-disease research prototype combining Fundus photography and OCT radiomics.
Seven fusion approaches were benchmarked across seven retinal categories, with interpretable radiomics and attention diagnostics built into the evaluation.
Attention-gated fusion pipelineFundus + OCT / 278 radiomics features ↗
- 88.92%
- Best geometric late-fusion accuracy
- 80.65%
- Attention mid-fusion balanced accuracy
- 8,550
- Class-paired hold-out samples
- Representation
- Each modality becomes 139 interpretable radiomics features spanning texture, frequency, and first-order statistics.
- Model design
- Classical XGBoost late fusion was compared with a sample-dependent attention gate over the combined 278-feature representation.
- Project scope
- Data pipeline, model benchmarking, interpretability analysis, and a deployed Streamlit research interface.
Research boundary: the hold-out pairs are class-matched rather than same-patient pairs, so results may be optimistic. This is a research prototype, not a clinical diagnostic product.
