Deformable Registration of Tumor-Diseased Brain Images

This paper presents an approach for deformable registration of a normal brain atlas to visible anatomic structures in a tumor-diseased brain image. We restrict our attention to cortical surfaces. First, a model surface in the atlas is warped to the tumor-diseased brain image via a HAMMER-based volumetric registration algorithm. However, the volumetric warping is generally inaccurate around the tumor region, due to the lack of reliable features to which the atlas can be matched. Therefore, the model structures for which no reliable matches are found are labeled by a Markov Random Field-Maximum A Posteriori approach. A statistically-based interpolation method is then used to correct/refine the volumetric warping for those structures. Finally, with the good initialization obtained by the above steps and the identification of the part of the model anatomy that can be recognized in the patient’s image, the model surface is adaptively warped to its counterpart that is visible in the tumor-diseased brain image through a surface registration procedure. Preliminary results show good performance on both simulated and real tumor-diseased brain images.