Training Hacks and a Frugal Man’s Net with Application to Glioblastoma Segmentation

Abstract

In this paper, we investigate the effectiveness of training a sparse Neural Network on a limited number of samples in the context of brain tumor segmentation. Nowadays, Deep Learning architectures are getting deeper, more sophisticated and environmentally unfriendly in an effort to improve their segmentation performance. We use a brain tumor segmentation dataset and apply simple practices to reduce the needed computational resources to allow cheap and fast training. We also present a lighter, cheaper version of the U-Net dubbed Frugal U-Net stemming from our investigation on how far we can push the original U-Net by decreasing its parameter count using Depth-Wise Separable Convolutions instead of regular ones, all the while preserving the minimum levels of accuracy required in Medical Imaging. Our methodology is useful in clinical facilities where high-computation resources are limited.