Feature-based supervised lung nodule segmentation

Abstract

Lung nodule segmentation allows for automatic measurement of the nodule's size or volume which is of utmost importance in lung cancer diagnosis. It is a challenging task since there are many different types of nodules (solid or non-solid, solitary or multiple, etc). A supervised lung nodule segmentation method uses a shape-based, contrast-based and intensity-based feature set to produce three preliminary segmentations and an artificial neural network to obtain a more accurate segmentation. This method was applied to 20 computer tomography studies, all containing nodules. The data has 10 images of solid nodules and 10 images of ground glass opacity nodules, all with ground-truth. The segmentation uses a region growing approach and the volumetric shape index is used for nodule detection and providing a seed point. In the first and second segmentation the probability of each neighbor belonging to the nodule is estimated using the volumetric shape index and the convergence index filter, respectively. The third segmentation is obtained using a feature set region regression method where for each neighbor the probability of belonging to the nodule or not is obtained using k nearest neighbor regression. Then, using a leave-one out method, an artificial neural network uses the three preliminary segmentations as input and is trained to obtain a more accurate segmentation. Results obtained a 12% relative volume error, 88% and 93% Jaccard and Dice coefficient respectively.