Optimized U-Net for Left Ventricle Segmentation

Abstract

The left ventricle segmentation is an important medical imaging task necessary to measure a patient's heart pumping efficiency. Recently, convolutional neural networks (CNN) have shown great potential in achieving state-of-the-art segmentation for such applications. However, most of the research is focusing on building complicated variations of these networks with modest changes to its performance. There is little to no insights on how these CNNs work and most of them are unfortunately treated the neural network as a black box. In this thesis, the famous U-Net architecture is used to segment the left ventricle from cardiac magnetic resonance (MR) images because of its simplicity and ability to analyze images at multiple scales. Posterior analysis of the network functionality demonstrates that by replacing the first set of layers of the U-Net with fixed filters, there is little change in performance compared to its fully connected version. This optimization was achieved by performing a Fourier analysis and visualization of the convolution layers after the completion of the network training phase. This analysis allows us to discover that some early layers approximate uniform filters which can then be replaced by fixed uniform kernel weights. 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 to achieve faster prediction time without compromising the performance. Experimental results and analysis are presented.