Denoisereg: Unsupervised Joint Denoising and Registration of Time-Lapse Live Cell Microscopy Images Using Deep Learning

Abstract

Image registration is important for analysing time-lapse live cell microscopy images. However, this is challenging due to significant image noise and complex cell movement. We propose a novel end-to-end trainable deep neural network for joint denoising and affine registration of temporal live cell microscopy images. Our network is trained unsupervised, and only a single network is required for both tasks which reduces overfitting. Our experiments show that the proposed network performs better than deep affine registration without denoising, and better than sequential deep denoising and affine registration. In combination with deep non-rigid registration, we outperform state-of-the-art non-rigid registration methods.