Adversarial Affine Registration for Real-Time Intraoperative Registration of 3-D US-US for Brain Shift Correction

One of the most frequent tumors in the central nervous system is glioma. The high-grade gliomas grow relatively fast and eventually lead to death. The tumor resection improves the survival rate. However, an accurate image-guidance is necessary during the surgery. The problem may be addressed by image registration. There are three main challenges: (i) the registration must be performed in real-time, (ii) the tumor resection results in missing data that strongly influence the similarity measure, and (iii) the quality of ultrasonography images. In this work, we propose a solution based on generative adversarial networks. The generator network calculates the affine transformation while the discriminator network learns the similarity measure. The ground-truth for the discriminator is defined by calculating the best possible affine transformation between the anatomical landmarks. This approach allows real-time registration during the inference and does not require defining the similarity measure that takes into account the missing data. The work is evaluated using the RESECT database. The dataset consists of 17 US-US pairs acquired before, during, and after the surgery. The target registration error is the main evaluation criteria. We show that the proposed method achieves results comparable to the state-of-the-art while registering the images in real-time. The proposed method may be useful for the real-time intraoperative registration addressing the brain shift correction.