Multi-Modal Volume Registration MIT Artificial Intelligence Lab and Brigham and Women's Hospital Overview We developed an algorithm to register multi-modal medical images using a prior intensity model. Various imaging modalities provide different information about the underlying tissue. It is common for many scans to be taken, which then need to be registered or aligned into a common coordinate frame. Framework of Algorithm The flowchart of the algorithm is shown on the right. A registered pair of training images is used to build the joint intensity model (below, left). This model measures the probability of two intensity values co-occurring in a registered image pair. Novel images are registered by sampling the intensities at corresponding points in the two images under an hypothesized pose and maximizing the pose given the prior model. Registration The image on the left shows the joint intensity distribution of a registered training pair. The image on the right shows two scans being registered overlayed in a checker-pattern. Notice that the intensity histogram (upper-right corner) appears more like the prior model as the registration converges to the correct alignment. Results The results of registering 36 PD/T2 scans are shown on the left, and the registrations of 36 SPGR/PD scans are tabulated on the right. The error is an RMS bounding box error, measured in millimeters. In most cases, the algorithm achieved sub-voxel alignment.