PerFact: Perception-Derived Fact Prompting for 3D Brain MRI Report Generation

TL;DR - PerFact generates reports from 3D multi-sequence brain MRI by prompting a LoRA-adapted vision-language model with structured facts produced by upstream segmentation and classification models. The study finds that grounding quality, rather than backbone choice or scale, is the main controllable factor in report quality.

• Five identically fine-tuned backbones spanning three model families and an order of magnitude in scale performed only marginally differently.
• Perception-derived facts outperformed retrieved prior reports, and retrieval became redundant when those facts were included.
• Predicted facts remained effective without ground-truth annotations at inference; the gap to oracle facts was attributed primarily to fact granularity.
• Adding closed-ended visual question answering caused no measurable reduction in report quality.