Context-Aware Intelligence | AI Augmented + Streamlined Reporting | High Yield Checklist Manifesto Workflow
RADFLOW is the AI-powered cognitive co-pilot for elite radiologists who refuse to compromise on precision, speed, or excellence.
Built on flexible multi-provider AI architecture (supporting Google Gemini, OpenAI, Anthropic Claude, OpenRouter, Perplexity, and more), RADFLOW deploys a sophisticated multi-agent AI system that works in parallel to provide:
- Real-time clinical intelligence across 8 simultaneous guidance streams
- Adaptive differential diagnosis that evolves with your findings
- Context-aware orchestration routing each cognitive task to specialized AI models
- Evidence-grounded recommendations with automatic web search when needed
- Interactive AI consultation for complex case discussions
This isn't just software. It's high-performance AI meeting high-stakes medicine—seamlessly integrated into a distraction-free workspace engineered to maintain your flow state, where your best diagnostic work happens.
Your attending radiologist's knowledge, instantly available. RADFLOW's groundbreaking parallel AI architecture simultaneously generates 8 comprehensive clinical guidance sections plus a separate appropriateness evaluation in real-time:
- Most Likely Diagnoses - Weighted differentials based on imaging findings
- Top Facts - High-yield pearls for rapid decision-making
- What to Look For - Systematic search strategy for the specific case
- Pitfalls & Mimics - Critical don't-miss warnings and lookalikes
- Search Pattern - Optimized workflow checklists
- Pertinent Negatives - Key findings to document for completeness
- Classic Signs - Pathognomonic features to confirm diagnoses
- Bottom Line - Synthesized clinical recommendation
Plus: ACR Appropriateness Criteria evaluation with intelligent web grounding—when the AI needs more context, it automatically searches authoritative sources to provide evidence-based guidance.
Native, low-latency speech recognition engineered for radiology workflows. Speak naturally, report effortlessly. Your voice becomes structured findings instantly.
Experience AI that evolves with your case. RADFLOW generates weighted differential diagnoses in real-time, then intelligently refines them as you add findings—removing outdated possibilities and suggesting new considerations. Get ranked differentials with clinical rationale, imaging features, and next-step recommendations. It's like having a diagnostic reasoning consultant at your fingertips.
A comprehensive AI-powered review system ensures excellence:
- Interactive Q&A - Chat with the AI about complex cases, ask for clarifications, explore alternative diagnoses
- Guideline-Aware Review - Integrated knowledge base of major radiology guidelines (Fleischer, Bosniak, ACR, LI-RADS) ensures follow-up recommendations meet best practices
- Language Refinement - AI polishes your prose while preserving your medical precision
- Quality Assurance - Final checklist review before sign-off
Every report, clinically sound and eloquently presented.
A revolutionary ambient workspace powered by Imagen (when using Google Gemini provider, dynamically selects the best available version—3.0, 4.0, or newer). Each session generates unique, AI-created medical concept art in the style of master impressionists—dynamic visualizations of neurological landscapes, cardiovascular systems, and cellular structures. Beautiful enough to inspire, subtle enough to prevent distraction during marathon reading sessions.
Note: Image generation requires a Google Gemini provider with Imagen model support.
RADFLOW represents a paradigm shift in how AI assists radiological interpretation. Unlike simple chatbots, RADFLOW implements a multi-agent orchestration system with specialized intelligence at every stage:
RADFLOW supports multiple AI providers, giving you the flexibility to choose the best model for each task:
- Google Gemini - Advanced multi-modal models with grounding capabilities and Imagen for background generation
- OpenAI - GPT-4 and other models for high-quality text generation
- Anthropic Claude - Claude 3 models with extended context windows
- OpenRouter - Access to multiple models through a unified API
- Perplexity - Specialized models with built-in search capabilities
- Custom Providers - Configure custom base URLs for self-hosted or proxy endpoints
Each provider can be configured with its own model assignments, allowing you to optimize for cost, speed, or quality on a per-task basis.
When you input a case, RADFLOW's Real-World Rundown launches 8 independent AI tasks that process simultaneously (plus a separate appropriateness evaluation):
- Each task is optimized for a specific cognitive function (differential generation, pitfall detection, search pattern creation, etc.)
- Responses arrive asynchronously as they complete, keeping you in flow
- Total guidance generation: ~15-30 seconds for comprehensive clinical intelligence that would take humans hours to research
The AI Orchestrator analyzes each request and routes it to the optimal model:
- JSON-structured tasks (categorization, differentials) → High-precision schema-validated models
- Streaming tasks (report drafting, impression synthesis) → Models optimized for natural language generation
- Grounding-enabled tasks (appropriateness evaluation) → Models with web search capabilities when local knowledge is insufficient (currently Google Gemini only)
- Image generation (ambient backgrounds) → Imagen models (Google Gemini only) with carefully crafted impressionist prompts
Unlike isolated AI calls, RADFLOW maintains semantic context across the entire workflow:
- Clinical history informs appropriateness evaluation
- Appropriateness insights guide the Real-World Rundown sections
- Dictated findings automatically trigger differential refinement
- Selected differentials integrate seamlessly into impression synthesis
- Every AI interaction builds on previous understanding
The differential diagnosis system demonstrates true reactive intelligence:
- Initial generation based on clinical history and study type
- Continuous monitoring of your dictated findings
- Automatic refinement when findings change—removing outdated differentials, adding newly relevant ones
- Ranking and weighting based on supporting/refuting evidence
- Clinical rationale for each diagnosis with key imaging features
This is AI that thinks alongside you, not just responds to prompts.
Before experiencing RADFLOW, ensure you have:
- Node.js (v18 or higher) — Download
- AI Provider API Key — At least one of:
- Google Gemini API Key (Default, recommended for full feature support)
- OpenAI API Key
- Anthropic API Key
- OpenRouter API Key
- Perplexity API Key
⚡ Power Requires Fuel
RADFLOW's intelligence requires API access to AI models. Google Gemini is recommended as the default provider for full feature support including grounding and image generation. Other providers can be added and configured in Settings.
git clone https://github.com/yourusername/radflow.git
cd radflownpm installOption A: Create a .env.local file in the root directory:
# Google Gemini (Default - Recommended for full features)
GEMINI_API_KEY=your_gemini_api_key_here
# Or use environment variable API_KEY for any provider
API_KEY=your_api_key_hereOption B: Enter your API key directly in the application's Settings panel after launch.
Option C: Add multiple providers in the Settings panel for flexibility:
- Configure different providers for different tasks
- Switch between providers based on your needs
- Use custom base URLs for proxies or self-hosted endpoints
npm run devAccess your workspace at http://localhost:5173
RADFLOW's multi-stage AI workflow orchestrates sophisticated intelligence at every step:
| Stage | What Happens | AI Power at Work |
|---|---|---|
| 1. Input | Paste clinical history or use voice dictation | Smart Categorization AI extracts structured data: demographics, allergies, medications, labs, prior imaging—organized instantly |
| 2. Guidance | Real-World Rundown activates | 8 parallel AI tasks + appropriateness evaluation simultaneously generate: clinical context, differential diagnoses, search patterns, pitfalls, classic signs, and clinical synthesis |
| 3. Dictate Findings | Voice or text entry of imaging observations | AI monitors your findings and auto-generates/refines differential diagnoses in real-time as you work |
| 4. Build Impression | Select relevant differentials | Impression Synthesis AI crafts a cohesive, evidence-based impression integrating your findings and selected diagnoses |
| 5. Refine & Review | Interactive AI consultation | Chat with the AI about the case, apply guideline-aware recommendations, polish language with AI assistance |
| 6. Finalize | Quality assurance check | Final Review AI validates completeness, consistency, and adherence to best practices before sign-off |
RADFLOW's AI Orchestrator intelligently routes each task to specialized models optimized for that specific cognitive function—categorization, clinical reasoning, guideline retrieval, language refinement. Context flows seamlessly between stages, creating a truly intelligent co-pilot that understands your case holistically.
| 🎯 Precision Multi-stage AI workflow with context preservation across all phases |
⚡ Speed 8 parallel AI tasks + native dictation = Maximum throughput |
🧘 Flow State Ambient design eliminates cognitive switching costs |
| 🔬 Evidence-Based ACR criteria, web grounding, and curated guideline knowledge base |
🤖 Intelligent AI orchestrator routes tasks to specialized models for optimal performance |
🎨 Inspired Dynamic impressionist medical art reduces fatigue during long sessions |
| 💬 Interactive Chat with AI about complex cases, get second opinions on demand |
🔄 Adaptive Differentials auto-refine as findings evolve—your AI thinks with you |
🚀 Modern Multi-provider support: Google, OpenAI, Anthropic, OpenRouter, Perplexity |
- DICOM viewer integration
- Multi-modal reporting (CT, MRI, US, X-ray templates)
- Team collaboration features
- Custom template library
- Advanced analytics dashboard
RADFLOW is built for radiologists, by those who understand radiology. We welcome contributions from the community.
Please read our Contributing Guidelines before submitting pull requests.
This project is available for personal use and research only. Commercial use requires express written permission from the creator. See the LICENSE file for details.
If RADFLOW enhances your practice, consider:
- ⭐ Starring this repository
- 📢 Sharing with colleagues
- 🐛 Reporting bugs or suggesting features
Engineered for radiologists who demand excellence.
Made with precision and purpose | Powered by Multiple AI Providers