INFORMATION EXTRACTION FROM RADIOLOGY REPORTS FOR A POPULATION BASED CANCER REGISTRY

摘要

A complete system of Cancer Information Extraction for apopulation based Cancer Registry is introduced. The analysisinvolves the classification and annotation of radiologyimaging reports to identify the components needed to completecancer staging and recurrence extraction. Besides traditionalsupervised learning methods such as ConditionalRandom Fields and Support Vector Machines, active learningapproaches are investigated to bring further improvementto the information extraction system performance. Areportability classifier, separating cancer from non-cancerreports, has achieved a performance of 97.74% sensitivityand 96.00% specificity on the held-out test set. The accuraciesof Report Purpose classifier and Tumour Stream classifierare approximately 80% on 10-fold cross-validation(CV) experiments. The overall F-score of the tagging systemis over 93% on 5-fold CV with approximately 487000instances from more than 3000 reports manually annotated.