ADAPT: Anatomy-Guided Token Pooling with Dynamic Anatomical Guidance for Radiology Report Generation
Published in SSRN (Preprint), 2026
This paper proposes ADAPT, an anatomy-guided framework that constructs compact visual prompts for LLM-based radiology report generation. Anatomy-Guided Token Pooling (AGTP) introduces anatomical segmentation masks as soft additive biases in cross-attention, steering learnable region queries toward relevant patches while retaining access to the complete patch sequence. A Dynamic Scale Predictor (DSP) estimates the bias strength for each examination, region, and attention head. Experiments on IU-Xray and MIMIC-CXR show gains over an R2GenGPT baseline; on MIMIC-CXR, ADAPT improves BLEU-4 from 0.117 to 0.134 and CheXbert precision from 0.506 to 0.545.
Recommended citation: M. Sao, D.-P. Dao, M. Lee, and H.-J. Yang, "ADAPT: Anatomy-Guided Token Pooling with Dynamic Anatomical Guidance for Radiology Report Generation," SSRN Preprint, 2026. [Online]. Available: https://ssrn.com/abstract=7553243
Download Paper