It doesn't take a whole U-Net to find brain tumors: towards fast brain metastasis segmentation with sparse predictions

Egor Krivov, Mikhail Belyaev

Abstract: In this work, we focus on brain metastasis delineation from 3D MRI scans. Currently, state-of-the-art fully-convolutional neural networks segment the whole image to solve this task. However, we suppose that it might be useful to split the problem into detection and delineation. The goal of the paper is to evaluate the first step of such approach. We propose a simple network architecture for metastasis detection by predicting downsampled segmentation. We observed an increased detection sensitivity and lower computational cost in comparison with 3D U-Net.

Keywords: detection, brain metastasis, volumetric convolutional networks