LungSeg: A Comparative Study of Neural Network Architecture for Lung Segmentation

Abstract

Medical professionals have relied on medical imaging as one of the key sources of data for diagnosis over the years. With the growth of the field of Artificial Intelligence and Machine Learning, medical imaging can be further innovated to improve healthcare. Using image segmentation, which can be implemented using convolutional neural networks, a single region can be isolated from any medical image and a very focused investigation can be performed on such regions. The focus of this paper is to compare various neural network models for their ability to segment a given region from a series of CT scans and recommend a suitable neural network model for segmentation. We have focused on isolating the lungs from CT scans and observed the effects of changing hyperparameters as well as augmentation on the results. For this research, we have taken three models for comparison, namely Universal Network, Pyramid Scene Parsing Network and Feature Pyramid Network. Quantitative metrics like Hausdorff Distance, Dice score, Intersection over Union, Structural Similarity, Mean Squared Error and qualitative metrics like inferred segmentation masks are used to compare the results. Based on the results we identify the most suitable model for lung segmentation application.