A fusion method for pulmonary nodule segmentation in chest CT image sets

Abstract

This paper presents a fusion algorithm combining two methods for pulmonary nodule segmentation in Chest CT scans. Segmentation is an important task for both diagnosis and therapy monitoring. Since manual segmentation of volumes is time-consuming and there is an inter-observer variability between each radiologist, there have been numerous studies on developing schemes for automated segmentation of lung nodules. In order to segment nodules automatically, we have merged two proposed methods from the literature. First, a semi-automated segmentation method is used to find an initial estimate of the nodule using a non-fixed thresholding scheme. Then, the vessels and tissues connected to the nodule segment are removed by using a refinement method based on geodesic distances. The initial results indicate a 0.618 Jaccard index on 50% consensus ground truth data extracted from 18 non-solid and solid nodules from the LIDC-IDRI database, representing a significant improvement for this type of nodules. It is shown that fusion improves the segmentation process by combining the well performing parts of individual algorithms. We expect to further improve the results by fusing other algorithms.