D2BGAN: Dual Discriminator Bayesian Generative Adversarial Network for Deformable MR-Ultrasound Registration Applied to Brain Shift compensation

Abstract

PurposeIn neurosurgery, image guidance is provided based on the patient to pre-operative data registration with a neuronavigation system. However, the brain shift phenomena invalidate the accuracy of the navigation system during neurosurgery. One of the most common approaches for brain shift compensation is using intra-operative ultrasound (iUS) imaging followed by registration of iUS with pre-operative magnetic resonance (MR) images. While, due to the unpredictable nature of brain deformation and the low quality of ultrasound images, finding a satisfactory multimodal image registration approach remains a challenging task. MethodsWe proposed a new automatic unsupervised end-to-end MR-iUS registration approach based on the Dual Discriminator Bayesian Generative Adversarial Network (D2BGAN). The proposed network consists of two discriminators and is optimized by introducing a Bayesian loss function to improve the generator functionality and adding a mutual information loss function to the discriminator for similarity measurement. An evaluation was performed using the RESECT training dataset based on the organizers manual landmarks. ResultsThe mean Target Registration Error (mTRE) after MR-iUS registration using D2BGAN reached 0.75{+/-}0.3 mm. The D2BGAN illustrated a clear advantage by 85% improvement in the mTRE of MR-iUS registration over the initial error. Also, the results confirmed that the proposed Bayesian loss function rather than the typical loss function outperforms the accuracy of MR-iUS registration by 23%. ConclusionThe D2BGAN improved the registration accuracy while allowing us to maintain the intensity and anatomical information of the input images in the registration process. It promotes the advancement of deep learning-based multi-modality registration techniques.