Data-Driven Strategies for Reliable Imaging: Optimized Training Sets and Unsupervised Motion Correction

Reinhard Heckel

Session: Machine learning meets computational imaging (Sara Fridovich-Keil, Mahdi Soltanolkotabi)

Keywords: Reliable Imaging, Unsupervised Motion Correction

TL;DR: Data-Driven Strategies for Reliable Imaging: 
Optimized Training Sets and Unsupervised Motion Correction

Abstract: Deep neural networks trained on example images perform excellent for medical and scientific imaging, but can be sensitive to distribution shifts and motion perturbations. In this talk, I’ll discuss how imaging reliability and performance, in particular performance under distribution shifts can be improved through dataset design. Second, I’ll discuss effective un-supervised deep learning based method for 3D rigid motion estimation.

Submission Number: 129