An extendable registration similarity metric for anatomical image sequence alignment

A brain anatomical image sequence obtained through histology posed a new challenge to medical image registration. Aligning hundreds to thousands of image slices using a pairwise registration technique may cause error propagation or introduce random error. Information across multiple adjacent image slices must be considered for the alignment. We developed a new similarity metric called minimum entropy of bad prediction (MEBP) that is suitable for pairwise image registration and image sequence alignment (ISA). MEBP is intensity-based, but it outperforms almost all other intensity-based metrics. When MEBP is used in ISA, it scales very well. MEBP has been applied to a rabbit brain digital atlas construction, and it is applicable to many similar problems.