🧠 AlzDetect: Next-Generation Neurological Diagnostics

An Enterprise-Grade Vision Transformer (ViT) Architecture for Early Alzheimer's Detection

View Live Deployment (alzdetect.live) • Diagnostic Scanner • Technical Documentation • Clinical Validation

---

📖 Executive Summary

AlzDetect is a production-ready medical technology platform engineered to assist neurologists in the early-stage detection of Alzheimer's disease. Moving beyond conventional diagnostics, the platform leverages state-of-the-art Vision Transformers (ViT-B/32) to extract granular, high-dimensional features from patient MRI sequences.

Designed with enterprise architecture in mind, the system integrates a mathematically rigorous Explainable AI (XAI) engine, real-time clinical protocol generation via LLMs, and stringent HIPAA-aligned data security protocols utilizing Supabase Row Level Security (RLS) and End-to-End Encryption methodologies.

---

🏗️ System Architecture

The application is structured as a decoupled microservices environment, ensuring strict separation of concerns, massive horizontal scalability, and adherence to medical data handling standards.

graph TD
    Client[Next.js 15 Client] -->|OAuth & Session| Auth[Supabase Auth]
    Client -->|HTTPS / REST| Fast[FastAPI Compute Cluster]
    Fast -->|Base64 Tensors| ML[Vision Transformer Engine]
    ML -->|Rollout Algorithm| XAI[Attention Map Generator]
    Client -->|XAI Outputs| Storage[Supabase PostgreSQL RLS]
    ML -->|Clinical Signals| OpenAI[GPT-4o Protocol Engine]

Loading

Core Tiers:

1. The Presentation Layer (Client): Engineered with Next.js 15 (React 18), featuring a custom glassmorphic UI system, fluid hardware-accelerated animations, and responsive clinical drill-down views.
2. The Compute Layer (Backend): A high-concurrency Python ASGI server (FastAPI) deployed via Uvicorn. Handles heavy tensor matrix transformations and AI inference routing.
3. The Intelligence Layer (Model): A dynamically loaded Vision Transformer employing a custom manual-rollout script to bypass Keras 3 schema restrictions, enabling pixel-perfect Attention Map visualizations.
4. The Persistence Layer (Database): Supabase (PostgreSQL) handling robust authentication, user-scoped data segregation via RLS, and blob storage for diagnostic visual histories.

---

🔬 Explainable AI & Clinical Validation

A core tenet of modern medical software is interpretability. The AlzDetect architecture implements a rigorous Attention Map Rollout.

Instead of treating the AI as a "black box," the system manually intercepts the neural network's internal Multi-Head Self-Attention layers block-by-block. It projects these highly activated weights backward to the input space, generating a high-contrast heatmap overlay on the original MRI. This mathematically demonstrates exactly which anatomical regions (e.g., Hippocampus, Ventricles) the network analyzed to output its prediction.

Empirical Benchmark Summary (ViT-B/32 vs. Baselines)

Model / Evaluation Configuration	Partition / Cohort	Test Accuracy	Balanced Acc	Macro F1	Macro ROC-AUC
Fine-Tuned ViT-B/32 (Primary Locked Model)	Primary Held-Out Test (n = 6,400)	98.84%	99.31%	99.24%	1.0000
Fine-Tuned ViT-B/32 (External Evaluation)	Secondary Test Set (n = 1,279)	97.73%	98.77%	98.59%	0.9994
Controlled 3-Seed Full Fine-Tuning (Mean ± SD)	Seeds 42, 123, 2024 (n = 6,400)	98.47% ± 0.53%	98.89%	98.84% ± 0.43%	0.9995
Baseline Custom CNN	Primary Held-Out Test (n = 6,400)	64.20%	72.00%	68.75%	0.8640
Baseline Unadapted ViT	Primary Held-Out Test (n = 6,400)	54.00%	61.25%	44.50%	0.7810

• Quantitative Attention Faithfulness (n = 80 Test MRIs, 2,000 Bootstrap Resamples): Statistically significant positive deletion AUC advantage (+0.12455, 95% CI [0.08904, 0.16481]) and insertion AUC advantage (+0.15387, 95% CI [0.10233, 0.20219]) over random patch perturbation.
• Live Production Platform: https://www.alzdetect.live/

---

🛡️ Enterprise Security & Compliance

• Role-Based Access Control (RBAC): Enforced at the database level via Supabase.
• Data Segregation: PostgreSQL Row Level Security (RLS) guarantees that individual clinical sessions and diagnostic logs are completely isolated and inaccessible across accounts.
• Stateless Inference: The diagnostic compute cluster immediately purges MRI buffers post-inference. History is permanently archived only if the user explicitly saves the validated sequence to their secure Vault.

---

🚀 Local Deployment Guide

Prerequisites

• Node.js (v18+)
• Python (v3.12, strict requirement for legacy tf-keras compatibility)
• pip / pnpm

1. The Client (Next.js)

cd client
pnpm install
# Ensure .env contains NEXT_PUBLIC_SUPABASE_URL and ANON_KEY
pnpm dev

Client will mount at http://localhost:3000

2. The Compute Engine (FastAPI)

cd server
python -m venv .venv_prod
source .venv_prod/Scripts/activate # Windows
pip install -r requirements.txt
# TF_USE_LEGACY_KERAS=1 is strictly required in .env
python -m uvicorn src.server:app --host 0.0.0.0 --port 8000 --reload

Engine will mount at http://localhost:8000

---

👤 Lead Architect

Bassey Riman
Lead Architect & AI Engineer
Integrating cutting-edge machine learning into scalable, beautifully-engineered consumer health interfaces.

_{Built for the future of Neurological Diagnostic Care.}