Revisiting LLM Adaptation for 3D CT Report Generation: A Study of Scaling and Diagnostic Priors

Short summary

RAD3D-Prefix framework enables efficient LLM adaptation for 3D CT report generation while reducing clinical hallucination. Key finding: smaller LLMs benefit from full fine-tuning, while larger models (1B+) achieve better generalization by freezing the LLM and training only lightweight projection layers. Method outperforms parameter-efficient baselines with substantial computational savings.

• •New parameter-efficient framework (RAD3D-Prefix) bridges semantic gap between visual features and clinical terminology
• •Smaller LLMs (96.1M) benefit from fine-tuning; larger LLMs (1B+) generalize better when frozen
• •Validated against clinical reader study; outperforms baselines with far fewer trainable parameters

Generated with AI, which can make mistakes.