Blind MRI Brain Lesion Inpainting Using Deep Learning

José V Manjón , José E Romero , Roberto Vivo-Hernando , Gregorio Rubio , Fernando Aparici , Maria de La Iglesia-Vaya , Thomas Tourdias (1) , Pierrick Coupé (2, 3)

1 Institut François Magendie
2 LaBRI - Laboratoire Bordelais de Recherche en Informatique
3 PICTURA - Patch-based processing for medical and natural images

In brain image analysis many of the current pipelines are not robust to the presence of lesions which degrades their accuracy and robustness. For example, performance of classic medical image processing operations such as non-linear registration or segmentation rapidly decreases when dealing with lesions. To minimize their impact, some authors have proposed to inpaint these lesions so classic pipelines can be used. However, this requires to manually delineate the regions of interest which is time consuming. In this paper, we propose a deep network that is able to blindly inpaint lesions in brain images automatically allowing current pipelines to robustly operate under pathological conditions. We demonstrate the improved robustness/accuracy in the brain segmentation problem using the SPM12 pipeline with our automatically inpainted images.

Mots clés: Lesion inpainting; MRI; deep learning; robust segmentation

Domaines: Imagerie médicale

José V Manjón, José E Romero, Roberto Vivo-Hernando, Gregorio Rubio, Fernando Aparici, et al.. Blind MRI Brain Lesion Inpainting Using Deep Learning. International Workshop on Simulation and Synthesis in Medical Imaging, Oct 2020, LIMA, Peru. pp.41-49, ⟨10.1007/978-3-030-59520-3_5⟩. ⟨hal-02966536⟩