On-demand instance-wise adaptation of large language models for radiology report generation

Radiology report generation (RRG) requires models to transform radiographs into structured and clinically precise findings, yet available radiograph-report datasets are limited and long-tailed. General-domain large language models (LLMs) can generate fluent text but may miss pathology-specific evidence, while conventional supervised fine-tuning requires substantial labeled data and prompt-based few-shot learning does not update parameters for the current case. We study ODAM, an instance-wise test-time adaptation method for RRG. ODAM selects a small training set for each test radiograph using training data selection (TDS), which combines visual feature comparing (VFC) with pathology matching (PM) so that visually similar but clinically different cases are less likely to dominate the selected set. It then updates the LLM on the selected radiograph-report pairs and applies contrastive instance discrimination (CID), where selected and unselected training instances provide positive and negative references. Experiments on two widely used RRG datasets show that ODAM improves natural language generation and clinical efficacy metrics over zero-shot, few-shot, and fine-tuned baselines. We further clarify the boundary of the method, including visually similar but clinically different cases, rare or out-of-distribution findings, and deployment latency constraints.