eduzhai > Applied Sciences > Engineering >

Extracting and Leveraging Nodule Features with Lung Inpainting for Local Feature Augmentation

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 9 pages

Abstract: Chest X-ray (CXR) is the most common examination for fast detection ofpulmonary abnormalities. Recently, automated algorithms have been developed toclassify multiple diseases and abnormalities in CXR scans. However, because ofthe limited availability of scans containing nodules and the subtle propertiesof nodules in CXRs, state-of-the-art methods do not perform well on noduleclassification. To create additional data for the training process, standardaugmentation techniques are applied. However, the variance introduced by thesemethods are limited as the images are typically modified globally. In thispaper, we propose a method for local feature augmentation by extracting localnodule features using a generative inpainting network. The network is appliedto generate realistic, healthy tissue and structures in patches containingnodules. The nodules are entirely removed in the inpainted representation. Theextraction of the nodule features is processed by subtraction of the inpaintedpatch from the nodule patch. With arbitrary displacement of the extractednodules in the lung area across different CXR scans and further localmodifications during training, we significantly increase the noduleclassification performance and outperform state-of-the-art augmentationmethods.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×