eduzhai > Applied Sciences > Engineering >

Paying Per-label Attention for Multi-label Extraction from Radiology Reports

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 14 pages

Abstract: Training medical image analysis models requires large amounts of expertlyannotated data which is time-consuming and expensive to obtain. Images areoften accompanied by free-text radiology reports which are a rich source ofinformation. In this paper, we tackle the automated extraction of structuredlabels from head CT reports for imaging of suspected stroke patients, usingdeep learning. Firstly, we propose a set of 31 labels which correspond toradiographic findings (e.g. hyperdensity) and clinical impressions (e.g.haemorrhage) related to neurological abnormalities. Secondly, inspired byprevious work, we extend existing state-of-the-art neural network models with alabel-dependent attention mechanism. Using this mechanism and simple syntheticdata augmentation, we are able to robustly extract many labels with a singlemodel, classified according to the radiologist s reporting (positive,uncertain, negative). This approach can be used in further research toeffectively extract many labels from medical text.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...