Deep Metamorphic Registration for Tumor-Affected Medical Image Alignment

Abstract

Accurate alignment of medical images is essential for effective treatment evaluation and disease monitoring. However, many existing image registration methods are designed for healthy images and face challenges when applied to pathological images, such as those containing tumors. Furthermore, methods that attempt to address these challenges often struggle with large pathological regions and unsmooth deformations around tumor areas. In this paper, we propose a deep learning-based metamorphic registration method that utilizes time-varying flow and intensity variations within pathological regions to achieve diffeomorphic deformations and accurate image matching. Our method is evaluated on the BraTS-Reg 2022 dataset, showing promising results in aligning 3D images with brain tumors, particularly in difficult cases involving large deformations and complex tumor structures. The source code is available at https://github.com/wjp718a/MetaLapIRN.