SunX: AI Radiography & Clinical Decision Support
An AI radiography platform integrating a DenseNet-121 vision model for chest X-ray pathology detection with a pgvector RAG engine to surface evidence-based clinical treatment guidelines.
Rationale
Why SunX? Most medical AI tools stop at pathology classification, leaving clinicians with raw probability numbers. SunX bridges visual AI and clinical action by retrieving evidence-based treatment plans directly from peer-reviewed clinical guidelines (IDSA/ATS, AHA).
The Clinical Guardrails: SunX uses RAG over verified medical literature rather than relying on LLM parametric memory, ensuring all treatment suggestions include traceable source citations for physician review.
Tech Stack
Core Engineering Challenge
The Clinical Challenge: Raw pathology detection scores leave physicians to look up treatment protocols manually.
The Solution
Integrated DenseNet-121 visual findings with a pgvector LangChain RAG engine, surfacing AHA/ACC & IDSA guideline-backed treatment recommendations with source citations.
Key Highlights
- ▹Built a DICOM image processing and HIPAA-aware PHI anonymization pipeline using `pydicom` and `MONAI` for automated medical image ETL.
- ▹Deployed a fine-tuned DenseNet-121 PyTorch vision model on NVIDIA Triton Inference Server to classify 14 pathology classes (Pneumonia, Cardiomegaly, Effusion, etc.).
- ▹Engineered a clinical RAG pipeline using LangChain and `pgvector` over 1536-dimensional embeddings, retrieving treatment recommendations grounded in AHA/ACC and IDSA medical guidelines.
- ▹Integrated Cornerstone.js WebGL DICOM viewer into a Next.js dashboard for zero-latency client-side pan, zoom, and window-level manipulation.
- ▹Orchestrated async inference tasks with Celery and Redis, fully containerized via Docker Compose with Nginx reverse proxying.
Architecture Details
SunX combines high-throughput vision model inference with vector-based medical retrieval.
1. Visual Inference Pipeline
- Incoming DICOM files are anonymized and preprocessed via MONAI.
- Tensors are dispatched to NVIDIA Triton Inference Server, running DenseNet-121 to output 14 multi-label pathology probability scores and bounding boxes.
2. Clinical Decision Support (RAG)
- Detected findings trigger a LangChain workflow that queries PostgreSQL + pgvector.
- The RAG engine retrieves matching clinical practice guidelines (IDSA/ATS) and formats a structured treatment recommendation.
3. WebGL Imaging Interface
- The Next.js frontend embeds Cornerstone.js, rendering DICOM pixel arrays directly on the GPU via WebGL for instantaneous window-leveling and annotation.
Interactive System Design
DICOM Image Ingestion
Engineering Insight
Receives raw medical imaging DICOM files securely, enforcing HIPAA compliance and preliminary header validation before queueing pipeline tasks.
Platform Showcase