Deep image inpainting for automatic brain tumor extraction using weak labels.

Abstract

Brain tumors are one of the leading cancer-related causes of death in all ages. The diversity of tumor shapes and the varying degrees of the visibility of their edges makes the analysis of tumors complex. The development of automatic tools can enhance tumor visualization and improve understanding and support of tumor-focused tasks. We propose an automatic brain tumor extraction method based on image inpainting. Using weak labels containing the approximate shape of the tumor, we are able to successfully remove the tumor from a Magnetic Resonance Image (MRI) by replacing it with non-tumor tissue through a partial convolution neural network trained over non-tumor tissue regions. Brain tumor extraction is then performed by calculating the residual between the original MRI and the reconstructed image without the tumor. The isolated tumor in the extracted tumor image is amenable to further analysis. To demonstrate the extracted tumor image potential, we performed tumor delineation using an active contour method. By clearly showing the tumor, the proposed method is valuable in helping experts come to an agreement when segmenting biomedical images.