eduzhai > Applied Sciences > Engineering >

Cardiac Segmentation with Strong Anatomical Guarantees

  • Save

... pages left unread,continue reading

Document pages: 11 pages

Abstract: Convolutional neural networks (CNN) have had unprecedented success in medicalimaging and, in particular, in medical image segmentation. However, despite thefact that segmentation results are closer than ever to the inter-expertvariability, CNNs are not immune to producing anatomically inaccuratesegmentations, even when built upon a shape prior. In this paper, we present aframework for producing cardiac image segmentation maps that are guaranteed torespect pre-defined anatomical criteria, while remaining within theinter-expert variability. The idea behind our method is to use a well-trainedCNN, have it process cardiac images, identify the anatomically implausibleresults and warp these results toward the closest anatomically valid cardiacshape. This warping procedure is carried out with a constrained variationalautoencoder (cVAE) trained to learn a representation of valid cardiac shapesthrough a smooth, yet constrained, latent space. With this cVAE, we can projectany implausible shape into the cardiac latent space and steer it toward theclosest correct shape. We tested our framework on short-axis MRI as well asapical two and four-chamber view ultrasound images, two modalities for whichcardiac shapes are drastically different. With our method, CNNs can now produceresults that are both within the inter-expert variability and alwaysanatomically plausible without having to rely on a shape prior.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×