Team NeuroVista
SIH 2026 — Problem Statement 26038
Seeing Beyond. Thinking Ahead.
NeuroVista-DR is an AI-assisted diabetic retinopathy (DR) screening system designed for resource-constrained and rural healthcare environments.
The system analyzes retinal fundus images and aims to provide:
- Retinal image quality assessment
- Diabetic retinopathy severity grading
- Referable DR identification
- Visual explanation using Grad-CAM
- Clear screening results for frontline users
- A pathway for ophthalmologist review
The long-term system will expand toward lesion-level evidence, calibrated confidence, clinical validation, and workflow simulation using Simulink.
Diabetic retinopathy is a major complication of diabetes that can lead to vision loss if not detected and managed appropriately.
Rural screening programs face challenges including:
- Limited availability of ophthalmologists
- Variable fundus image quality
- Limited access to specialist review
- Large screening volumes
- Lack of transparency in many AI-based systems
SIH Problem Statement 26038 asks for a MATLAB-based retinal image analysis pipeline addressing image quality, retinal structures, DR severity grading, explainability, and screening workflow simulation.
NeuroVista-DR is designed around a quality-gated, explainable screening pipeline.
Fundus Image
↓
Image Quality Assessment
↓
Preprocessing / Enhancement
↓
DR Classification
↓
ICDR Severity Grade (0–4)
↓
Referable DR Decision
↓
Grad-CAM Explanation
↓
Screening Result
↓
Future: Ophthalmologist Review
The system is intended as AI-assisted screening/decision support, not as a replacement for an ophthalmologist.
The first internal-hackathon prototype focuses on a working end-to-end vertical slice:
- Fundus image upload
- Image quality gate
- Image preprocessing
- DR severity classification
- ICDR 0–4 grading
- Referable DR decision
- Grad-CAM visualization
- Web-based screening interface
- Backend API integration
- Lesion segmentation
- Lesion-level evidence
- Optic disc/fovea localization
- Calibrated confidence
- Annotated reports
- Ophthalmologist review workflow
- External dataset validation
- Rural/offline optimizations
- Simulink workflow simulation
NEUROVISTA-DR
│
▼
FUNDUS IMAGE
│
▼
IMAGE QUALITY GATE
│
┌────────┴────────┐
│ │
GOOD POOR
│ │
│ Recapture Feedback
│
▼
PREPROCESSING
│
▼
DR CLASSIFICATION
│
▼
ICDR 0–4
│
┌─────┴─────┐
▼ ▼
REFERABLE? GRAD-CAM
│ │
└─────┬─────┘
▼
SCREENING RESULT
See docs/architecture.md for details.
The prototype uses transfer learning for retinal image classification.
EfficientNet-B0
The model will classify images into five International Clinical DR severity levels:
| Grade | Severity |
|---|---|
| 0 | No DR |
| 1 | Mild NPDR |
| 2 | Moderate NPDR |
| 3 | Severe NPDR |
| 4 | Proliferative DR |
For screening purposes:
Referable DR = Grade ≥ 2
Model performance will be reported using measured evaluation results rather than assumed or target values.
The prototype uses Grad-CAM to visualize image regions contributing to the model's prediction.
Grad-CAM should be interpreted as a model-attention visualization, not as definitive proof that a highlighted region represents a specific retinal lesion.
Future versions will strengthen explainability by combining model attention with lesion-level evidence.
The project will evaluate:
- Accuracy
- Precision
- Recall / Sensitivity
- Specificity
- F1-score
- Confusion matrix
- ROC-AUC
- Calibration where applicable
The SIH problem statement specifies a target of:
- >90% sensitivity
- >85% specificity
for referable DR.
These are project targets, not achieved results.
NeuroVista-DR/
│
├── assets/
│ ├── demo/
│ ├── diagrams/
│ └── screenshots/
│
├── data/
│ ├── raw/
│ └── processed/
│
├── docs/
│
├── models/
│ ├── grading/
│ ├── quality/
│ └── segmentation/
│
├── notebooks/
│
├── simulink/
│ ├── configs/
│ ├── models/
│ └── results/
│
├── src/
│ ├── backend/
│ ├── frontend/
│ └── matlab/
│
├── tests/
│
├── .gitignore
├── environment.yml
├── requirements.txt
└── README.md
- MATLAB
- MATLAB Image Processing Toolbox
- MATLAB Deep Learning Toolbox
- Computer Vision Toolbox
- Statistics and Machine Learning Toolbox
- Python
- FastAPI
- MATLAB Engine integration where required
- React
- Vite
- MATLAB Simulink
- SimEvents where required
- Git
- GitHub
- VS Code
The project will use publicly available retinal datasets for research and development, subject to their respective licenses and usage conditions.
A major Indian-domain dataset under consideration is the Indian Diabetic Retinopathy Image Dataset (IDRiD), which provides DR grading and lesion-related annotations.
Dataset selection, splitting, and licensing details are documented separately.
See:
docs/
└── problem-statement.md
NeuroVista-DR is a research and hackathon prototype.
It is intended to demonstrate an AI-assisted screening and decision-support workflow.
It is not a clinically validated diagnostic system and must not be used as a substitute for professional medical evaluation.
Current Phase: M0 — Project Foundation
Development will proceed through:
M0 — Project Foundation
↓
M1 — Core AI Pipeline
↓
M2 — Explainability + Application
↓
M3 — SIH Integration & Demo
Seeing Beyond. Thinking Ahead.
The project is being developed for SIH 2026 Problem Statement 26038.