MedGemma Outperforms Fine-tuned RadBERT in Zero-Shot CT Report Labeling
August 7, 2026
A study of 40,833 CT reports shows MedGemma 27B and MedGemma-1.5 4B perform superior zero-shot labeling of organ systems compared to rule-based algorithms and fine-tuned RadBERT. The research highlights how labeling conventions influence macro-averaged F1 scores across multi-disease classification.
HOW THIS AFFECTS YOU
●
researcherThe findings provide a benchmark for evaluating zero-shot medical labeling performance.
●
healthOpen-weight medical LLMs show high potential for automating clinical report annotation.