eduzhai > Applied Sciences > Engineering >

3D Self-Supervised Methods for Medical Imaging

  • Save

... pages left unread,continue reading

Document pages: 19 pages

Abstract: Self-supervised learning methods have witnessed a recent surge of interestafter proving successful in multiple application fields. In this work, weleverage these techniques, and we propose 3D versions for five differentself-supervised methods, in the form of proxy tasks. Our methods facilitateneural network feature learning from unlabeled 3D images, aiming to reduce therequired cost for expert annotation. The developed algorithms are 3DContrastive Predictive Coding, 3D Rotation prediction, 3D Jigsaw puzzles,Relative 3D patch location, and 3D Exemplar networks. Our experiments show thatpretraining models with our 3D tasks yields more powerful semanticrepresentations, and enables solving downstream tasks more accurately andefficiently, compared to training the models from scratch and to pretrainingthem on 2D slices. We demonstrate the effectiveness of our methods on threedownstream tasks from the medical imaging domain: i) Brain Tumor Segmentationfrom 3D MRI, ii) Pancreas Tumor Segmentation from 3D CT, and iii) DiabeticRetinopathy Detection from 2D Fundus images. In each task, we assess the gainsin data-efficiency, performance, and speed of convergence. Interestingly, wealso find gains when transferring the learned representations, by our methods,from a large unlabeled 3D corpus to a small downstream-specific dataset. Weachieve results competitive to state-of-the-art solutions at a fraction of thecomputational expense. We publish our implementations for the developedalgorithms (both 3D and 2D versions) as an open-source library, in an effort toallow other researchers to apply and extend our methods on their datasets.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×