Optimizing U-Net to Segment Left Ventricle from Magnetic Resonance Imaging

Abstract

Left ventricle segmentation is an important medical imaging task to measure several diagnostic parameters related to the heart such as ejection fraction and stroke volume. Recently, convolutional neural networks (CNN) have shown great potential in achieving state-of-the-art segmentation results for such applications. However, most of the existing research is focusing on building complicated variations of the neural networks with modest changes to their performance. In this study, the popular U-Net architecture is optimized by analyzing its behaviour once fully trained from which one can simplify its architecture by fixing layers weights or eliminating some of them completely. For instance, by performing a Fourier analysis of the convolution at each layer, we were able to discover that some early layers can be approximated by simple uniform filters. Furthermore, in a separate experiment by removing the middle layers of the U-Net one can reduce the number of U-Net parameters from 31 million to 0.5 million weights without compromising its performance. The experimental evaluations show that the new optimized U-Net achieves 0.93 for the Dice score in comparison to manual ground truth.