Small Language Models Achieve CPU-Deployable Radiology AI via LoRA Fine-Tuning

A recent study shows that small language models (SLMs) with 3-4 billion parameters can efficiently tackle various radiology tasks when fine-tuned with LoRA, even on standard consumer CPUs. This breakthrough tackles the difficulties of using large language models (LLMs) in resource-limited clinical settings. The research trained the Qwen2.5-3B-Instruct and Qwen3-4B models on 162,000 samples from nine radiology tasks, including RADS classification for ten systems, impression generation, and more, using data from 12 public datasets. Tests on up to 500 withheld samples per task revealed that LoRA fine-tuning led to impressive performance gains: RADS accuracy jumped by 53%, NLI by 60%, and N-staging by 89%. This makes AI in radiology more accessible without needing specialized hardware.