The Segmentation of Nonsolid Pulmonary Nodules in Ct Images

Nonsolid nodules are a common radiographical finding in high resolution CT images of the lung. A main factor in determining a nodules malignancy status is the change in the nodule size over time. A method for automatically segmenting a nonsolid nodule from CT images is presented in this thesis. Precise image segmentation is a prerequisite for determining the volumetric growth rate from multiple image scans and the corresponding nodule malignancy status. There has been limited previous work on a segmentation technique for nonsolid nodules. The methods that have been proposed have lacked clinical validation with a radiologist ground truth and often include smaller datasets. The method in this thesis directly compares radiologist ground truth with our automated method and examines the consistency of growth measurement for further validation. The segmentation method consists of three stages; bilateral noise reduction, a probability based voxel classifier and geometric vessel removal. Parameter optimization and validation of the segmentation algorithm is facilitated with a dataset of 20 nonsolid nodule images in which a radiologist has established ground truth by outlining the boundary of the nodule in each image that it is visible. The optimal parameters were determined using the overlap metric and a training/testing methodology. The automated method achieved an average overlap of 0.43 with the radiologist ground truth. An experiment was conducted to determine whether the radiologist manual boundaries or the automated segmentations were more consistent at measuring the volumetric growth between three time scans of the same nodule. Results were determined for two different growth models (exponential and linear) on a dataset of 25 nonsolid nodules. The growth variation of the automated method was found to be 1.87 compared to the radiologist growth variation of 3.00. This suggests that, if the assumption of consistent nodule growth holds for nonsolid nodules, then the automated method provides a more precise growth rate estimate than the radiologist markings.