CNN-BASED UNSUPERVISED REGISTRATION OF TIME-LAPSE MICROSCOPY IMAGE SEQUENCES
Keywords: Image Registration, Deep Learning, Convolutional Neural Networks, Live Cell Imaging
Abstract. Image registration is widely used in live cell microscopy image analysis to compensate for the cell motion. It is a challenging task as the cell is not only moving (which causes rotation and translation), but also changes its form in time making the motion non-rigid. To address this, we propose a CNN-based unsupervised method for non-rigid registration of live cell image sequences. Our network predicts both the deformation field between a pair of images of the sequence and an affine transformation matrix for the cell motion compensation. The method can be used alone or in combination with other approaches. The proposed approach was successfully applied to real live cell microscopy image sequences.We conducted an experimental comparison with existing methods including contour-based, intensity-based and a deep learning based joint denoising and registration method. In addition, we analyzed different deformation regularizers and their impact on the alignment results. In combination with contour-based method we outperformed the existing approaches in average registration accuracy for two metrics on the standard evaluation dataset.