Neural-network-based method for intrathoracic airway detection from three-dimensional CT images

摘要

This paper presents a neural network-based method for intrathoracic airway detection and segmentation from 3D HRCT images. Two feed-forward neural networks are independently trained to identify various airway appearances in 3D CT images. While the first network identifies potential airways located adjacent to vessels, the second network identifies potential airways by assessing the existence of walls surrounding airways, The two networks are combined to construct a dual-network classifier taking its inputs from a 21 $MUL 21 moving subimage window: (1) raw gray-level subimage and (2) 4 directional profiles. By design, each network provides a superset of airways that are present in the CT images and only the airways identified by both networks are considered reliable. After the networks are trained by the generalized delta rule with momentum using limited number of airway/nonairway samples apart from the validation data sets, the generalization performance of the networks is assessed with two independent standards consisting of 282 and 167 observer-traced airways. The performance of the current method is compared with that of the conventional seeded region growing method. Our validation results indicate that the presented method indeed provide enhanced detection of peripheral airways compared to the conventional region growing method.