Residual 3D U-Net with Localization for Brain Tumor Segmentation.

Gliomas are brain tumors originating from the neuronal support tissue called glia, which can be benign or malignant. They are considered rare tumors, whose prognosis, which is highly fluctuating, is primarily related to several factors, including localization, size, degree of extension and certain immune factors. We propose an approach using a Residual 3D U-Net to segment these tumors with localization, a technique for centering and reducing the size of input images to make more accurate and faster predictions. We incorporated different training and post-processing techniques such as cross-validation and minimum pixel threshold.