Automated Methods For Pulmonary Nodule Growth Rate Measurement: Early Computer-Aided Diagnosis Of Lung Cancer From Computed Tomography Images

Pulmonary nodules are visible as dense, opaque areas in the lung on computed tomography (CT) images and may be early indications of lung cancer. Pulmonary nodule growth rate is highly correlated with malignancy and therefore its evaluation is useful in clinical decision making. Automated methods have been developed for nodule growth rate measurements, but these methods exhibit large measurement error; reducing this error will enable radiologists to make better decisions regarding follow up and treatment, in turn improving patient outcomes. Four major aspects of pulmonary nodule measurement are addressed in this thesis. A formal procedure for the comparative evaluation of different computer algorithms for pulmonary nodule change measurement has been developed that involves a standardized set of 50 CT image pairs and an analysis method. This procedure for the first time addresses the need to be able to quantitatively compare the performance of different methods. A study has been conducted in which developers of 18 computer methods participated and the results form a baseline with which to compare current and future algorithms. Two different computer algorithm approaches were developed to reduce the uncertainty in growth rate measurements. The first approach, moment-based compensation (ZCOMP) was performed on segmented nodule images to address additional observed increased error in the z-direction compared to the xyplane. By applying ZCOMP, volumetric measurement variability was reduced from a 95% limits of agreement of (-24.0%, 18.2%) to (-12.4%, 12.7%) on zerochange nodules imaged on thin-slice scans of the same resolution. The second approach was developed to address difficult-to-segment nodules with complex shapes and attachments. Instead of explicitly segmenting the nodule from the lung parenchyma, the growth index from density method (GID ) uses the density change in a region of interest as a surrogate growth measure. The GID method had much lower variation, (-11.0%, 12.3%) compared to a volumetric segmentation method, (-25.2%, 18.6%). Finally, an automated method was developed for measuring murine pulmonary nodule growth from micro-CT scans, adapting work from methods developed for human patients. This provides improved accuracy for lesion growth measurements used in small animal pre-clinical studies. The method addresses the additional noise, lack of contrast, and poor calibration of micro-CT scans. The measured growth rate was compared to the exponential growth model, and on a dataset of six nodules with repeat scans, the method measured growth that was consistent with the model