3D brain registration with intensity shift robustness

Technological advances in medical imaging are enabling us to understand healthcare datasets in great detail. Machine Learning enabled methods, specifically, deep neural networks are continuously achieving benchmark performances in terms of accuracy and computational efficiency. However, the lack of agreed-upon standard procedures, variations in the devices by different vendors, and artifacts induced by the physical phenomenon in the sensors make the data inconsistent and noisy. These variations in the data are detrimental to the performance of learning-based methods. In this study, we analyze the behavior of traditional and deep learning-based image registration methods and explore strategies to handle the problem of intensity distributional shifts without compromising the performance. To achieve this, we propose an intensity- based loss function and demonstrate that the models trained with our proposed loss function are better at handling unseen data from different sites using machines from different vendors. In addition, our trained model is superior in preserving the boundaries of anatomical regions after registration.