Retinopathy Detection & Progression Model
Summary
Freelancer Client is hiring: Retinopathy Detection & Progression Model.
Location: Remote
Temporal Lesion-Aware Dynamic Gated Multimodal Fusion Framework for DR and DME Analysis Using OLIVES and MMRDR Datasets
What you'll do:
• improve DR and DME classification performance,
Requirements:
• and demonstrate strong cross-dataset robustness.
Skills: Python, Statistics, Machine Learning (ML), Statistical Analysis, Data Science, Image Processing, Data Analysis, Deep Learning, Predictive Analytics, Convolutional Neural Network
Budget: $1500–$12500 USD
Source: Freelancer Client via Remote / Online. Apply on the source website.
Original
Temporal Lesion-Aware Dynamic Gated Multimodal Fusion Framework for DR and DME Analysis Using OLIVES and MMRDR Datasets
Framework Overview
The proposed framework introduces a Temporal Lesion-Aware Dynamic Gated Multimodal Fusion System for automated analysis of Diabetic Retinopathy (DR) and Diabetic Macular Edema (DME) using multimodal retinal imaging data. The framework combines fundus images, OCT scans, longitudinal retinal information, and optional clinical metadata to improve retinal disease classification, biomarker understanding, and temporal disease progression analysis.
Unlike conventional multimodal retinal systems that use static feature fusion, the proposed framework employs a:
Dynamic Gated Cross-Modal Fusion Mechanism that adaptively learns the importance of each retinal modality during disease prediction.
The framework utilizes:
• OLIVES dataset for multimodal temporal retinal learning,
• MMRDR dataset for cross-dataset robustness and generalization analysis.
Input Modalities
The framework uses:
1. Fundus Images
Provide:
• vascular abnormalities,
• hemorrhages,
• exudates,
• global retinal appearance.
2. OCT Images
Provide:
• retinal layer structure,
• intraretinal fluid (IRF),
• subretinal fluid (SRF),
• edema-related biomarkers.
3. Clinical Metadata
Includes:
• CST,
• BCVA,
• age,
• biomarker annotations.
Proposed Architecture
1. Modality-Specific Feature Extraction
Fundus Branch
A lightweight:
Swin Transformer Tiny is used for:
• retinal vascular representation learning,
• global retinal contextual understanding.
OCT Branch
A lightweight:
EfficientNet-B3 is used for:
• retinal biomarker extraction,
• OCT structural feature learning.
Clinical Branch
A lightweight:
MLP encoder is used for clinical feature embedding.
2. Lightweight Temporal Transformer
To model retinal progression over time, multimodal retinal embeddings are processed using a:
Lightweight Temporal Transformer
This module learns:
• disease progression patterns,
• biomarker evolution,
• temporal retinal dependencies across visits.
3. Lesion-Aware Attention Module
A lesion-aware attention mechanism dynamically prioritizes clinically important retinal regions such as:
• IRF,
• SRF,
• hemorrhages,
• exudates,
• retinal lesion areas.
This improves:
• biomarker-sensitive learning,
• lesion localization,
• explainability.
4. Dynamic Gated Cross-Modal Fusion
Instead of static concatenation, the framework uses:
Dynamic Gated Cross-Modal Fusion to adaptively learn:
• which modality is more important,
• how fundus and OCT features should interact,
• and how multimodal information should contribute to final prediction.
The gating mechanism dynamically adjusts modality importance based on retinal pathology characteristics.
This improves:
• multimodal consistency,
• adaptive retinal reasoning,
• and robust feature fusion.
Tasks Performed
The framework simultaneously performs:
Task
Description
DR Grading
Classification of DR severity (0–4)
DME Detection
Binary classification
Biomarker Prediction
Detection of retinal biomarkers
Temporal Progression Analysis
Longitudinal retinal progression learning
Explainability
To improve clinical interpretability, the framework incorporates:
• Grad-CAM++ and SHAP,
• temporal attention visualization,
• retinal heatmaps,
• biomarker localization overlays.
These outputs help ophthalmologists understand and validate model predictions.
Cross-Dataset Evaluation
The framework supports:
Train on OLIVES → Test on MMRDR to evaluate:
• domain robustness,
• multimodal transferability,
• and real-world clinical generalization.
Expected Outcomes
The proposed framework is expected to:
• improve DR and DME classification performance,
• capture temporal retinal progression,
• enhance biomarker-aware retinal learning,
• provide clinically interpretable attention maps,
• and demonstrate strong cross-dataset robustness.
If you need clarification on data format or evaluation protocol, just let me know and I will provide sample files.
Location & Details
Apply on source →About this listing
This remote opportunity was imported from Freelancer and is shown here for discovery. To apply, follow the link to the original posting.