ttanida/rgrg

Uses a two-stage architecture combining an object detector to localize 29 anatomical chest regions with region-specific binary classifiers for saliency and abnormality detection, feeding selected region features into a language model for sentence generation. Enables interactive clinical workflows through anatomy-based and freeform bounding-box-based sentence generation, where radiologists can selectively query descriptions for specific regions rather than receiving fully automated reports. Trained on MIMIC-CXR and Chest ImaGenome datasets with evaluation via NLG metrics (BLEU, CIDEr, ROUGE) and clinical efficacy metrics aligned with radiologist observations.

210 stars. No commits in the last 6 months.