2019/10/01 by Tobi Olatunji, Yao Li, Olatunji, Tobi +1
Computer Science · Medicine · #AI in cancer detection #COVID-19 diagnosis using AI #Computation and Language (cs.CL) #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Radiomics and Machine Learning in Medical Imaging #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1910.00673
openalex publication_date 2019/10/01 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28
Bootstrapping labels from radiology reports has become the scalable\nalternative to provide inexpensive ground truth for medical imaging. Because of\nthe domain specific nature, state-of-the-art report labeling tools are\npredominantly rule-based. These tools, however, typically yield a binary 0 or 1\nprediction that indicates the presence or absence of abnormalities. These hard\ntargets are then used as ground truth to train image models in the downstream,\nforcing models to express high degree of certainty even on cases where\nspecificity is low. This could negatively impact the statistical efficiency of\nimage models. We address such an issue by training a Bidirectional Long-Short\nTerm Memory Network to augment heuristic-based discrete labels of X-ray reports\nfrom all body regions and achieve performance comparable or better than\ndomain-specific NLP, but with additional uncertainty estimates which enable\nfiner downstream image model training.\n