Task-Specific NLP Outperforms General LLMs for Lung Nodule Detection in Chest CT Reports

Key Details

• 1Task-specific NLP model ('FiNd') was compared with seven general-purpose LLMs (Gemma, Haiku, Sonnet 2, GPT-4o, DeepSeek, Phi-4, MedGemma).
• 2FiNd was developed using 21,542 radiology reports and tested on 1,016 chest CT reports.
• 3Performance was assessed for identifying incidental lung nodules (ILNs) and categorizing them by size (<6 mm, 6–7.9 mm, ≥8 mm).
• 4FiNd achieved 96.8% accuracy for nodules ≥6 mm and 97.4% for nodules ≥8 mm, outperforming all general LLMs.
• 5General-purpose LLM accuracy ranged from 77.7% to 88.6% for detection tasks; specificity varied widely among models.
• 6Researchers urge adaptation of LLMs using radiology-specific training for better clinical integration.