RADIANT-PET: LLM reasoning cuts PET/CT false positives
Reinforcement learning trains a local LLM to classify lesions like a radiologist.
RADIANT-PET tackles the persistent challenge of false positives in PET/CT lesion segmentation, where physiologic tracer uptake and artifacts can mimic malignant signal. The framework uses a deliberately permissive initial segmentation stage to capture all candidate uptake regions, then converts those voxel-level outputs into structured textual descriptions summarizing intensity, morphology, and anatomical context (both regional and global). A large language model (LLM) classifies each candidate as true lesion or false positive, with an option to incorporate the radiology report for additional clinical context.
To strengthen lesion-level reasoning, the researchers optimize a local LLM using reinforcement learning via Group Relative Policy Optimization (GRPO), rewarding correct lesion classification and anatomically concordant site assignment. Evaluated on the AutoPET benchmark and an Ohio State University (OSU) test cohort, RADIANT-PET consistently outperforms strong image-only baselines, with the largest gains when radiology reports are provided. This work demonstrates that LLM-based reasoning can suppress physiologic false positives and align voxel-level predictions with clinical interpretation, effectively adding a novel reasoning layer beyond conventional segmentation.
- RADIANT-PET pairs a high-sensitivity segmentation model with an LLM to reduce false positives in PET/CT lesion detection.
- A local LLM is fine-tuned via Group Relative Policy Optimization (GRPO) to reward correct lesion classification and anatomical site assignment.
- Outperforms image-only baselines on AutoPET and OSU cohorts; largest improvements occur when radiology reports are provided.
Why It Matters
LLM reasoning can reduce misdiagnoses in oncology imaging by mimicking radiologist judgment.