![]() | Only 14 pages are availabe for public view |
Abstract Studying pathological brain images has significant importance for several technical and clinical aspects, including computer-aided diagnosis systems, longitudinal brain studies, and preoperative planning. Nevertheless, several previous studies proved that damaged brain tissues have adverse effects on the outputs of brain image processing tasks, such as brain tissue segmentation and nonlinear registration. Besides, most brain image analysis software also assumes spatial consistency of input images or treats pathological brains as exceptional cases. Thereby, designing dedicated image process ing solutions for each type of brain damage is highly complex as brain tissue degener ation has numerous categories and medical causes, such as tumors, multiple sclerosis, encephalitis, and brain hemorrhage. Accordingly, image inpainting was introduced as a preprocessing method to replace abnormal brain tissues with healthy-looking ones to reduce the distortions caused by tissue inhomogeneity and restore spatial consistency. This study presents an extensive literature review discussing the methods developed for deep image inpainting and image inpainting in the brain imaging domain. Then, a new deep adversarial model is proposed for user-guided free-form inpainting high resolution T1-weighted brain magnetic resonance images using a large dataset ac quired from the Human Connectome Project. In addition, a novel linear-time algo rithm is proposed to generate random masks for training images on the fly. Quantita tive and qualitative comparative evaluations demonstrated that the proposed method outperforms the current state-of-the-art methods for brain image inpainting. Besides, the proposed network shows promising results of brain tumor inpainting and subse quent partial tissue segmentation. |