RADSIM: AI Powered Flight Simulator for Radiology Education

The Problem

Radiology residents need to interpret 10,000+ cases to reach proficiency, but training is limited by access to diverse, real patient cases. Traditional case libraries are static, expensive ($300-1000/year), and offer no adaptive learning. This creates a bottleneck in medical education where residents graduate with insufficient exposure to rare but critical pathologies.

The Solution

RADSIM is an open source, AI powered radiology training simulator that generates unlimited synthetic medical imaging cases. Think "flight simulator for radiologists" where residents practice diagnostic interpretation in a safe environment with AI generated cases that adapt to their skill level.

Demo: LinkedIn Post - Hackathon Build

Why GPU Resources Are Critical

I need GPU compute for three core components:

1. Medical Image Generation (Primary GPU Use)

• Fine tuning diffusion models (Stable Diffusion variants) on medical imaging datasets
• Generating realistic synthetic X-rays, CT scans, and MRIs
• Processing: ~1000 synthetic images initially, then 100-200/week for new cases
• Models: RoentGen based architectures, medical specific fine tuning

2. Medical LLM Fine tuning

• Fine tuning LLaMA or Mistral models on radiology reports and educational content
• Training on 50K+ radiology reports for realistic case generation
• DPO (Direct Preference Optimization) for educational feedback quality
• Already built proof of concept (see my LLM-From-Scratch repo - 29 stars)

3. Real Time Inference at Scale

• Serving LLM for case generation and adaptive feedback
• Image generation for on demand case creation
• Target: Supporting 100-1000 concurrent medical students globally

Current Progress & Validation

• Built medical LLM from scratch (instruction tuned with LoRA + DPO)
• Founder of BirthSense AI (NYU Prototyping Fund recipient)
• AI Research Fellow at In Vivo Group (AI/ML in life sciences)
• Multiple hackathon wins in healthcare AI
• Africa Oxford Initiative Health Innovation Programme 2025

Impact Potential

• 20,000+ radiology residents globally could train with unlimited, diverse cases
• Particularly impactful in low resource settings (Africa, Asia) where access to diverse pathology is limited
• Open source commitment: Will release trained models, synthetic datasets, and platform architecture to community
• Addresses global radiologist shortage: WHO estimates shortage of 2.3M healthcare workers, including radiologists

GPU Resource Needs

Training Phase (Months 1-3)

• Hardware: A100/H100 for medical image generation fine tuning and LLM training
• Duration: ~200-400 GPU hours

Inference Phase (Ongoing)

• Hardware: T4/A10 for serving models to users
• Duration: ~50-100 GPU hours/month initially

Dataset

Using publicly available medical imaging datasets (MIMIC-CXR, NIH ChestX-ray14, etc.)

Timeline

Open Science Commitment

All work will be:

• Open source on GitHub (MIT license)
• Models published on Hugging Face Hub
• Synthetic medical datasets made available to research community
• Educational materials and training pipeline documented

Why Hugging Face?

Your infrastructure is perfect for this: models hosted on Hub, Spaces for deployment, and GPU resources for training. This aligns with HF's mission of democratizing AI, making advanced medical education accessible globally.

Links

• GitHub: github.com/samadon1
• LinkedIn: linkedin.com/in/samdon
• Medical LLM Project: LLM From Scratch (29 stars)

Request Summary

Training: A100 or H100 access for 200-400 hours over 3 months (medical image generation fine tuning)

Inference: T4 or A10G for continuous serving (3-6 months initially)

This project bridges AI innovation and global health equity using cutting edge AI to train the next generation of radiologists, particularly in underserved regions.