eduzhai > Applied Sciences > Engineering >

Compensating for visibility artefacts in photoacoustic imaging with a deep learning approach providing prediction uncertainties

  • Save

... pages left unread,continue reading

Document pages: 16 pages

Abstract: Conventional photoacoustic imaging may suffer from the limited view andbandwidth of ultrasound transducers. A deep learning approach is proposed tohandle these problems and is demonstrated both in simulations and inexperiments on a multi-scale model of leaf skeleton. We employed anexperimental approach to build the training and the test sets using photographsof the samples as ground truth images. Reconstructions produced by the neuralnetwork show a greatly improved image quality as compared to conventionalapproaches. In addition, this work aimed at quantifying the reliability of theneural network predictions. To achieve this, the dropout Monte-Carlo procedureis applied to estimate a pixel-wise degree of confidence on each predictedpicture. Last, we address the possibility to use transfer learning withsimulated data in order to drastically limit the size of the experimentaldataset.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×