Overview

Mediscan was created to explore how edge-hosted, privacy-first AI models can aid clinicians in preliminary medical image review without transmitting sensitive patient data over external networks.

Architecture & Key Features

  • Local Model Execution: Employs Ollama running quantized Llama 3.2 vision models on premises.
  • REST API Pipeline: Lightweight Python/Node backend that processes DICOM and standard image formats.
  • Zero-Cloud Data Leaks: All inference occurs locally on clinical hardware.

Challenges Solved


1. Low-latency processing on commodity GPU hardware.
2. Standardized structured output formatted for EHR integration.
3. Clinical explainability, highlighting image attention regions.