Analysis of U-Net Neural Network Training Parameters for Tomographic Images Segmentation

Abstract

Image segmentation is one of the main resources in computer vision. Nowadays, this procedure can be made with high precision using Deep Learning, and this fact is important to applications of several research areas including medical image analysis. Image segmentation is currently applied to find tumors, bone defects and other elements that are crucial to achieve accurate diagnoses. The objective of the present work is to verify the influence of parameters variation on U-Net, a Deep Convolutional Neural Network with Deep Learning for biomedical image segmentation. The dataset was obtained from Kaggle website (www.kaggle.com) and contains 267 volumes of lung computed tomography scans, which are composed of the 2D images and their respective masks (ground truth). The dataset was subdivided in 80% of the volumes for training and 20% for testing. The results were evaluated using the Dice Similarity Coefficient as metric and the value 84% was the mean obtained for the testing set, applying the best parameters considered.

Keywords

• deep learning
• biomedical image segmentation
• fully convolutional networks
• u-net
• computed tomography