Nikhil Thaker, Medical Director of Oncology Research at Capital Health (US), shared a post on LinkedIn:
“Excited to share our new publication: RadOncRAG: A Novel Retrieval-Augmented Generation Framework Improves Large Language Model Benchmark Performance in Radiation Oncology
In this study, we evaluated 15 leading large language models using 298 questions from the American College of Radiology (ACR) in-training exam.
What we found:
- RAG significantly boosts performance for non-reasoning models, with gains of +5.7% to +8.4%.
- GPT-4o with RadOncRAG reached 85.6% accuracy—14% higher than graduating PGY-5 residents.
- Reasoning models (GPT4-o1, o3-mini, DeepSeek-R1) already scored ~91–92% and saw no additional gain, suggesting their internal reasoning already compensates for knowledge gaps.
- Performance improved across clinical, biology, and physics domains—critical areas for safe radiation oncology care.
- RAG provides traceable, evidence-grounded answers, helping reduce hallucinations and improve explainability.
Why this matters:
Radiation oncology requires accurate, multidisciplinary, up-to-date knowledge. Our findings show that RAG can elevate cost-effective models to near–expert performance, democratizing advanced clinical decision support without requiring huge reasoning-model costs.
What we built:
The Iridium Model, a radiation-oncology–specific RAG pipeline with embeddings, semantic search, and curated corpora (guidelines, textbooks, protocols, Q&A banks).
Big picture:
RAG isn’t just a feature—it’s infrastructure for safer, more transparent AI in oncology. It helps LLMs cite evidence, reduce hallucinations, and support clinicians with real, verifiable data.
If you’re working in AI, oncology, or medical education, I’d love to connect. This is just the beginning.”
Title: RadOncRAG: A Novel Retrieval-Augmented Generation Framework Improves Large Language Model Benchmark Performance in Radiation Oncology
Authors: Nikhil Gautam Thaker, Navid Redjal, Adam Dicker, Arturo Loaiza-Bonilla, Trevor Royce, Vivek Subbiah, Vikash Deendyal, Jonathan R. Gabriel, Neena Shetty, Ajay Choudhri, Gautam H. Thaker
Read the Full Article.
More posts featuring Nikhil Thaker.