From Theory to Practice: GenAI Skills for Radiologists (2025)-M7-CIN07-2025

Video Transcription

Video Summary

The session “From Theory to Practice: Gen‑AI Skills for Radiologists” outlines how generative AI could help radiology address a widening workload–capacity gap (rising exam volume, more data per exam, slower workforce growth). Radiology leads medical AI adoption, with hundreds of FDA-authorized imaging AI tools, though most are discriminative (detection/classification) and there is still no FDA-cleared healthcare LLM. Speakers emphasize the shift toward foundation models—large, self-supervised, often transformer-based systems that can be adapted to many downstream tasks—alongside the need to close the medical data scale gap versus general AI.<br /><br />Key GenAI applications discussed include image enhancement (resolution, denoising, artifact reduction), synthetic data generation for rare diseases and privacy-preserving validation, workflow optimization (protocol selection, triage, appropriateness checks), and reporting (structuring free text, error checking, information extraction, and patient-friendly summaries). Retrieval-augmented generation (RAG) is highlighted as essential to ground outputs in current, traceable medical sources.<br /><br />A practical segment on LLM use explains next-token prediction, risks like “sycophancy” from RLHF, and how better prompts and persistent personalization/system instructions improve usefulness. The diffusion-model talk compares VAEs, GANs, and diffusion models, stressing diffusion’s superior stability and fidelity but high compute and slow sampling, plus barriers to clinical translation and the need to measure diagnostic utility—not just image-similarity metrics.

Keywords

generative AI in radiology

foundation models

workload-capacity gap

image enhancement and denoising

synthetic medical imaging data

workflow optimization and triage

radiology report generation

retrieval-augmented generation (RAG)

diffusion models vs GANs and VAEs