Images obtained from serial section electron microscopy can contain defects that create discontinuous tissue deformation. Fixing such defects during image registration is especially challenging, as classical block matching registration techniques assume smooth motion within each block, and ConvNet based registration techniques must rely on smoothness assumption during training. We propose Caesar, a divide-and-conquer technique that breaks registered images into segments, such that most of discontinuity is confined to segment boundaries. Then, we align the segments independently and stitch the results back together. We provide extensive experimental evaluation on brain tissue serial section microscopy data that shows that segment-wise alignment reduces the average misalignment area around defects by 6-10x.