3D Deconvolution with Deep Learning

3D deconvolution is a challenging problem faced by astronomers and micrsoscopists. The goal of deconvolution algorithms is to reverse the effects of convolution on the observed data. Convolution can be understood as a blurring operation on the ground truth that results in a relatively blurred/unclear observation. It is challenging to model the exact convolution operation because of various factors like noise, optical aberrations and the complexity of optical setup. The Point Spread Function (PSF) is the impulse response of the imaging system which helps us to model a major part of the image forming process. In this report, we look at various methods used for deconvolution of 3D microscopy images. In particular, we compare the performance of typical non-blind iterative algorithms like Richardson-Lucy (RL) and Alternating Direction Method of Multipliers (ADMM) that require knowledge of the PSF. The second set of techniques are based on Deep Learning methods that use Convolutional Neural Networks (CNNs) to learn a non-linear mapping from the observed data to the ground truth. In this context, we consider two network architectures and analyze the results of all the methods.