Deformable image registration of worm brains

In this project, VoxelMorph [1] was used to align all the frames of a confocal microscopy video of C. Elegans in order to generalize the segmentation data of seven red fluorescently-labelled neurons to all the frames. The goal was to reach similar or higher accuracy than state-of-the art registration methods but in a shorter time. First, a new pipeline was investigated to roughly pre-align some raw data and test unsupervised VoxelMorph in 2D as a proof-of-concept. In a second time, VoxelMorph was applied with unsupervised and semi-supervised learning on pre-aligned frames to assess the efficiency and the rapidity of the algorithm in 3D. The best results were obtained with the semi-supervised training MSE, a L2-regularizer λ of 5e−2, 32 epochs and a batch size of 4.