PROJECT TITLE :

Progressively Trained Convolutional Neural Networks for Deformable Image Registration

ABSTRACT:

The quick registration periods of deep learning-based algorithms for deformable picture registration make them viable alternatives to conventional methods. For complicated deformation fields, a multi-resolution technique is necessary to estimate bigger displacements. The solution proposed in this article is to train neural networks incrementally. A convolutional neural network is first trained on low-resolution pictures and deformation fields before being trained on a larger version of the network. During training, additional layers are added to the network that are trained on more detailed data. For example, we show that a network trained this manner can learn greater displacements without reducing registration precision. This results in a more robust system that is less sensitive to massive misregistrations. On the intra-patient lung CT registration problem, we produce huge numbers of ground truth examples using random synthetic alterations applied to a training set of pictures. To determine how the progressive learning technique affects training, we examine the learnt representations in the gradually increasing network. A gradual training process leads to better registration accuracy while learning complicated deformations, as demonstrated in this study