High-Resolution MRI Brain Inpainting

Analyzing diseased brain images is often challenging as the presence of abnormal brain tissues hinders several image processing tasks, like tissue segmentation and non-rigid registration, leading to distorted or biased results. This paper presents a solution to this problem by replacing abnormal brain tissues with healthy ones using free-form image inpainting to restore spatial consistency. Inspired by the recent success of Gated Convolution, we propose a user-guided deep adversarial inpainting model to fill irregularly shaped holes in high-resolution T1-weighted MR brain images. Besides, we propose a novel linear-time algorithm to generate irregular masks on the fly for training images. Comparative qualitative and quantitative analyses demonstrated that the proposed model outperforms recently proposed deep learning methods for brain lesion filling and significantly reduces blurriness and boundary artifacts. Moreover, the proposed network shows promising performance in inpainting brain tumors and improving tissue segmentation outputs.