RESEARCH / MEDICAL AI
Shows how structured schemas, staged extraction and batch processing turn radiology reports into consistent labels for training and evaluation. The paper reports an internal study of accuracy, reliability and processing cost.
Read on this pageRadiology reports contain valuable information for training and evaluating imaging models, but their free text is difficult to use consistently at scale. This paper describes a system for converting CT chest reports into a fixed, structured set of findings while retaining distinctions such as absence, presence and uncertainty.
The approach uses a constrained output schema and separates the initial finding check from the more detailed description of findings that are present. It evaluates label accuracy, structural reliability and the cost of processing a large archive. For medical AI development, the contribution is a clearer route from existing reports to usable research labels, together with an explicit distinction between labels extracted by a model and ground truth independently established by radiologists.
A fixed schema makes report-derived labels consistent and machine-readable.
Two-phase extraction focuses detailed work on findings identified as present.
Accuracy, output validity and processing cost are evaluated separately.
The accuracy study uses an internal radiologist-annotated held-out set of about 100 studies. Archive outputs remain weak labels extracted from text, not independently verified clinical truth. They are for research and model development, not clinical decisions without independent review and appropriate authorisation.
The complete paper, with its original figures, tables and references.
35 pages · Read without leaving this pageDiscuss the approach, evidence and relevance to your team with the people building it.
