Jean-Benoit Delbrouck

Research blog All posts →

Long-form reading notes on radiology AI papers, with a focus on report generation.

X-ray and CT report generationMICCAI 2025Main13 Oct 2026~10 min read

RRG-DPO, a cheap fix for missed findings, less for false alarms

RRG-DPO trains report generators to prefer the true report over a look-alike with the wrong findings: recall rises in six X-ray and CT models, precision much less, and the model with the highest F1 barely moves. What you can use today.

Endoscopy report generation and hallucination checkingMICCAI 2025Main13 Oct 2026~10 min read

Gut-VLM, 7 in 10 GPT-4o endoscopy descriptions had an error

Gut-VLM releases 1,816 GPT-4o descriptions of endoscopy images with expert tags and corrections: what it holds, what you can use today, and what its hallucination-aware fine-tuning result shows.

CT lesion description and question answeringMELBA 2026MICCAI Open Data12 Oct 2026~10 min read

CT-Bench, a four-choice lesion exam where GPT-4V scores 28%

CT-Bench adds a description and a size to 20,335 DeepLesion CT lesions and builds 2,850 four-choice questions on them: what the release holds, what you can use today, and what its baselines and reader study say.

Experience

1. Jul 2026 – nowAI Researcher · Eden
2. Mar 2025 – Jul 2026AI Researcher · HOPPR
3. Jun 2024 – Mar 2025AI Researcher · Hugging Face
4. Sep 2020 – Mar 2025Postdoc, then Academic Staff (Research) · Stanford University, AIMI Center Coordinated a research lab: publications, mentoring PhD students and postdocs, funding.