Project 8: Structural position of radiology findings¶
Mentors: To be announced
Problem: Radiology reports are assumed to follow a predictable structure, with findings landing in roughly the same place every time, but nothing measures whether that's actually true.
Context: Built on Open-I de-identified chest X-ray reports (Indiana University), anchored to ML4LLM Ch.7 · proj45: Minkowski distance, mutual information, and token positions (helper).
Goals: Do "impression"/finding statements reliably cluster at a predictable structural position within a radiology report?
Deliverables: A notebook that tags finding-related sentences in Open-I reports, computes mutual information between token position (normalized by report length) and finding-mention, and compares against Minkowski distance between finding and non-finding token-position distributions to quantify how structured radiology reporting is.
Showcase: TBD
References:
- Learning to Summarize Radiology Findings: treats the Findings-to-Impression structure of radiology reports as a modeling target, the same structural regularity this project measures directly with mutual information and Minkowski distance.