FLARE: Selectively Routing Lesion Evidence for Comprehensive Volumetric Radiology Reporting

Recent multimodal large language models (MLLMs) have enabled radiology report generation from full 3D CT volumes, but clinically important focal lesions can remain underrepresented within global volumetric context. A natural remedy is to augment the global input with lesion-focused visual tokens. However, our experiments show that this naive global-plus-local approach yields inconsistent benefits, improving finding-level F1 while degrading other report-generation metrics. We introduce FLARE, a parameter-efficient post-training framework that explicitly controls when local lesion evidence influences generation. FLARE combines a lesion-focused adapter for extracting localized evidence with a token-wise router that selectively incorporates this evidence into a frozen global report generator. This separates extracting lesion information from deciding when it should affect the report. Using ground-truth lesion ROIs on the ReXGroundingCT evaluation cohort, FLARE improves all evaluated clinically oriented metrics over both full-report supervised fine-tuning and naive local augmentation, while maintaining broadly comparable textual similarity to the global-only baseline. Matched-versus-random ROI comparisons and token-level analyses further support that the intended lesion evidence contributes substantially to the gains and preferentially affects finding-relevant content. These results support selective evidence routing as a parameter-efficient approach to strengthening local lesion descriptions within comprehensive volumetric reports.