Spatial gating improves accuracy of AI radiology report generation
Diagnosis-Conditioned Spatial Gating and Decoder-Level Supervised Contrastive Learning for Radiology Report Generation
Computer Vision and Pattern Recognition
Summary
Radiology AI systems can write reports but often make mistakes about what they see in images. The paper introduces a new way to better connect predicted medical findings with specific image parts, helping the AI focus where it matters. It also uses a method to teach the AI to better understand shared findings across cases, improving report accuracy. These ideas together help AI generate more reliable radiology reports without needing extra detailed image labels.
What this means in practice
- •For hospital it teams: Deploy improved AI systems that produce more accurate radiology reports by integrating spatially aware diagnosis conditioning and contrastive training.
- •For medical device manufacturers: Incorporate new AI techniques to enhance diagnostic support features in medical imaging software products for clinical use.$Commercial implications: Enables creation of advanced radiology software products with better automated report accuracy, appealing to healthcare providers.
Authors
Md Mustafizur Rahman, Mylene C. Q. Farias
Abstract
Radiology report generation models can produce fluent text while still containing finding-level inaccuracies. Diagnosis-driven methods improve generation by conditioning on predicted findings, but these predictions do not directly modify the visual patch features provided to the decoder, and global gating applies the same modulation across spatial locations. We introduce a Position-Aware Gate (PAG) that uses predicted finding representations to modulate visual patches spatially without region supervision. We also propose a Decoder-level Supervised Contrastive Loss (DSCL) that structures decoder representations using shared positive findings rather than instance identity. On MIMIC-CXR, PAG+DSCL improves clinical efficacy (CE) F1 from 0.484 to 0.502 over a matched global-gate reference, while PAG and DSCL individually reach 0.491 and 0.495. Without additional fine-tuning, the combined model achieves 0.226 CE F1 on IU X-Ray, compared with 0.211 reported by PromptMRG.