Learning to Segment the Lung Volume from CT scans based on Semi-Automatic Ground-Truth

Patrick Sousa, Adrian Galdran, Pedro Costa, Aurélio Campilho

Abstract: Lung volume segmentation is a relevant task within the design of Computer-Aided Diagnosis systems related to automated lung pathology analysis. Isolating the lung from CT volumes can be a challenging process due to the considerable deformations and pathologies that can appear in different scans. Deep neural networks can be an effective mechanism in order to model the spatial relationship between different lung voxels. Unfortunately, this kind of models typically require large quantities of annotated data, and manually delineating the lung from volumetric CT scans can be a cumbersome process. In this paper, we propose to train a 3D Convolutional Neural Network to solve this task based on semi-automatically generated annotations. To achieve this goal, we introduce an extension of the well-known V-Net architecture that can handle higher-dimensional input data. Even if the training set labels are noisy and may contain some errors, we experimentally show that it is possible to learn to accurately segment the lung relying on them. Numerical comparisons performed on an external test set containing lung segmentations provided by a medical expert demonstrate that the proposed model generalizes well to new data, reaching an average 98.7% Dice coefficient. In addition, the proposed approach results in a superior performance when compared to the standard V-Net model, particularly on the lung boundary, achieving a 0.576 mm Average Symmetric Surface Distance with respect to expert validated ground-truth.

Keywords: Lung segmentation, Deep Learning, CT images

Author Affiliation: INESC TEC, Faculty of Engineering of University of Porto