MedVisionAI

Screen Earlier. Explain Better. Reach Further.

Point-of-Care Explainable AI System for Diabetic Retinopathy Screening & Clinical Triage

---

📖 The Story

01. The Human Stakes

Diabetic retinopathy is preventable blindness. Screening for it usually isn’t.

Over 3.9 million people worldwide suffer from irreversible vision loss caused by Diabetic Retinopathy (DR). When detected early, 90%+ of severe vision loss is entirely preventable.

Yet, conventional screening requires dilated fundus examinations performed by specialized ophthalmologists using expensive, hospital-grade fundus cameras. Because community health centers and primary care clinics rarely have access to these resources, early microvascular lesions—microaneurysms, hemorrhages, and hard exudates—routinely go undetected until irreversible retinal damage has already occurred.

       [ Status Quo: Overburdened Specialty Pipeline ]
  Primary Clinic ──(Long Waitlist)──> Ophthalmology Queue ──> Delayed Diagnosis ──> Vision Loss

       [ MedVisionAI: Frontline Point-of-Care Triage ]
  Primary Clinic ──(60s AI + Grad-CAM)──> Stratified Triage ──> Timely Specialist Intervention

---

02. The Trust Problem

In clinical medicine, a black-box percentage is not enough.

A 94% confidence score alone does not give an attending clinician enough justification to initiate a specialist referral. Did the neural network identify authentic microvascular pathology near the macula, or was it influenced by optical dust or lighting artifacts at the edge of the photograph?

MedVisionAI solves the trust gap by coupling deep learning inference with layer-level Gradient-weighted Class Activation Mapping (Grad-CAM). Healthcare workers and consulting clinicians can visually verify whether the model’s focal attention correlates with authentic physiological lesions before taking clinical action.

---

✨ Core Capabilities

flowchart LR
    A[📷 Fundus Photo] --> B[🧠 EfficientNet-B0]
    B --> C[🔥 Grad-CAM Heatmap]
    B --> D[📊 DR Stage 0-4 & Confidence]
    C --> E[📑 Clinical PDF Report]
    D --> E
    E --> F[👨‍⚕️ Clinician Workstation]
    E --> G[👤 Patient Health Portal]
    G --> H[💬 Report-Grounded AI Assistant]

Loading

🩺 1. Clinical Workstation (/worker/dashboard)

• Instant Neural Inference: Evaluates digital fundus photographs in seconds using a fine-tuned PyTorch EfficientNet-B0 architecture across 5 International Clinical Diabetic Retinopathy (ICDR) severity levels:
  • Stage 0: No Diabetic Retinopathy
  • Stage 1: Mild Non-Proliferative DR
  • Stage 2: Moderate Non-Proliferative DR
  • Stage 3: Severe Non-Proliferative DR
  • Stage 4: Proliferative DR
• Visual Explainability (Grad-CAM): High-resolution thermal activation overlays showing exact pixel regions influencing the prediction.
• Comprehensive Patient Onboarding: Register patients with rich biodata (diabetes type, year of diagnosis, existing ocular conditions, DOB).
• Patient Directory & Status Tracking: Live patient management with search, account status tracking (Active vs Pending Activation), and one-click structured Excel export.
• Automated Clinical PDF Generation: Instant generation of audit-ready diagnostic screening summaries.

👤 2. Patient Health Portal (/patient/dashboard)

• In-Page Report Viewer: Interactive modal PDF reader allowing patients to review their retinal screening summaries directly on their device without forced downloads.
• Timeline & Risk Breakdown: Simple, color-coded diagnostic summaries designed for patient clarity.
• Report-Grounded AI Assistant: An intelligent conversational agent backed by ChromaDB vector search that is strictly bounded to the patient's own diagnostic report, explaining medical findings without hallucinations.
• Frictionless Self-Activation (/set-password): Email-independent account activation using Email + Date of Birth verification, enabling patients to set secure credentials on first login without requiring external email delivery infrastructure.

🛡️ 3. Security & Session Governance

• Role-Based Access Control (RBAC): Strict separation between DOCTOR, WORKER, SPECIALIST, ADMIN, and PATIENT roles.
• Autofill Credential Isolation: Separate DOM inputs (clinician_* vs patient_*) and dynamic lifecycle keying to prevent browser password managers from cross-contaminating doctor credentials into patient fields.
• Role-Specific Session Lifetimes:
  • Clinicians: 5-minute strict inactivity security timeout for clinical compliance.
  • Patients: Persistent 30-day session with "Remember me" capabilities.

---

🏗️ Architecture & Technology Stack

Layer	Technologies	Purpose
Frontend UI / UX	React 18, TypeScript, Vite, Tailwind CSS v4, Lucide Icons, Canvas Confetti	High-performance, responsive clinical interface with dual role-tailored dashboards.
Backend API	Python 3.10+, FastAPI, Pydantic, SQLAlchemy, Uvicorn	Async REST API, authentication middleware, and background task processing.
Deep Learning & XAI	PyTorch, Torchvision, Captum, PIL, OpenCV	Deep neural network inference (EfficientNet-B0) and Grad-CAM attention localization.
Report Generation	ReportLab, PyPDF2	Vectorized clinical PDF report synthesis with embedded fundus images and heatmaps.
Vector Search & RAG	ChromaDB, Embeddings API / Local Transformers	Scoped document embeddings for report-grounded conversational assistant.
Database	SQLite (Dev) / PostgreSQL (Prod)	Relational storage for patient records, screenings, accounts, and clinical notes.

---

🔄 Standardized Clinical Screening Pipeline

The platform enforces a four-stage screening protocol structured for clinical fidelity, operational speed, and full auditability:

  [01. Ingestion]        High-resolution retinal fundus photograph captured at point of care and verified for optical quality.
        │
        ▼
  [02. Inference]        EfficientNet-B0 screens retinal fields for microaneurysms, hemorrhages, and ICDR severity classification.
        │
        ▼
  [03. Explainability]   Grad-CAM heatmaps highlight pixel-level neural attention regions driving the model's triage decision.
        │
        ▼
  [04. Stratification]   Confidence-scored clinical PDF report synthesized for specialist review, doctor notes, and referral.

---

🔑 Demonstration Portals & Access Accounts

For evaluation and clinical review across the dual-role architecture:

Role	Email Identifier	Access Portal	Primary Capabilities
Clinician / Doctor	dr.screening.medvision@gmail.com	Clinical Workspace	Full fundus analysis, Grad-CAM inspection, patient registry, and diagnostic PDF sign-off.
Healthcare Worker	worker.demo@medvision.ai	Clinical Workspace	Point-of-care patient intake, fundus photo upload, and preliminary triage queue access.
Patient (Active)	priya.sharma.eye@gmail.com	Patient Health Portal	Interactive in-page PDF report viewing, risk timeline, and grounded medical AI assistant.
Patient (Pending)	Registered by Clinician	Account Self-Activation	Verified identity confirmation via Email + DOB with self-service password setup.

---

📁 Repository Structure

MedVisionAI/
├── backend/
│   ├── app/
│   │   ├── api/
│   │   │   ├── endpoints/
│   │   │   │   ├── auth.py         # Authentication & token governance
│   │   │   │   ├── screening.py    # Retinal inference & Grad-CAM pipeline
│   │   │   │   ├── patients.py     # Patient registration & directory API
│   │   │   │   ├── assistant.py    # Report-grounded RAG chatbot
│   │   │   │   └── reports.py      # PDF report generation
│   │   │   └── deps.py
│   │   ├── models/                 # SQLAlchemy ORM models
│   │   ├── schemas/                # Pydantic validation schemas
│   │   ├── services/               # Inference, XAI, RAG, & PDF services
│   │   └── main.py                 # FastAPI application root
│   ├── init_db.py                  # Database seed & initialization
│   └── requirements.txt
│
├── frontend/
│   ├── public/
│   │   └── logo.png                # Official MedVisionAI brand logo
│   ├── src/
│   │   ├── components/             # Reusable UI components & animations
│   │   ├── context/                # AuthContext & state providers
│   │   ├── pages/
│   │   │   ├── Landing.tsx         # Story-driven landing & interactive showcase
│   │   │   ├── Login.tsx           # Dual-role authenticated portal entry
│   │   │   ├── SetPassword.tsx     # DOB identity verification & activation
│   │   │   ├── WorkerDashboard.tsx # Doctor diagnostic workstation
│   │   │   ├── PatientDashboard.tsx# Patient results & PDF viewer
│   │   │   └── PatientList.tsx     # Directory & Excel export
│   │   └── services/               # Axios API client
│   └── package.json
└── README.md

---

⚖️ Clinical & Regulatory Disclaimer

IMPORTANT NOTICE:
MedVisionAI is an AI-assisted screening tool intended for preliminary assessment support only. It is not an autonomous medical device and does not provide a definitive medical diagnosis. All artificial intelligence predictions, severity classifications, and Grad-CAM visual heatmaps must be reviewed and validated by a licensed healthcare professional or qualified ophthalmologist before initiating any treatment or clinical intervention.

---

_{Developed for frontline healthcare workers and ophthalmologists worldwide. © 2026 MedVisionAI.}