ArXiv Preprint
Registration of brain scans with pathologies is difficult, yet important
research area. The importance of this task motivated researchers to organize
the BraTS-Reg challenge, jointly with IEEE ISBI 2022 and MICCAI 2022
conferences. The organizers introduced the task of aligning pre-operative to
follow-up magnetic resonance images of glioma. The main difficulties are
connected with the missing data leading to large, nonrigid, and noninvertible
deformations. In this work, we describe our contributions to both the editions
of the BraTS-Reg challenge. The proposed method is based on combined deep
learning and instance optimization approaches. First, the instance optimization
enriches the state-of-the-art LapIRN method to improve the generalizability and
fine-details preservation. Second, an additional objective function weighting
is introduced, based on the inverse consistency. The proposed method is fully
unsupervised and exhibits high registration quality and robustness. The
quantitative results on the external validation set are: (i) IEEE ISBI 2022
edition: 1.85, and 0.86, (ii) MICCAI 2022 edition: 1.71, and 0.86, in terms of
the mean of median absolute error and robustness respectively. The method
scored the 1st place during the IEEE ISBI 2022 version of the challenge and the
3rd place during the MICCAI 2022. Future work could transfer the inverse
consistency-based weighting directly into the deep network training.
Marek Wodzinski, Artur Jurgas, Niccolo Marini, Manfredo Atzori, Henning Muller
2022-11-14
