Direct measurements of airway lumen and wall areas are potentially useful as a diagnostic tool and as an aid to understanding the pathophysiology underlying lung disease. Direct measurements can be made from images created by high resolution computer tomography (HRCT) by using computer-based algorithms to segment airways, but current validation techniques cannot adequately establish the accuracy and precision of these algorithms.
A detailed review of HRCT airway segmentation algorithms was undertaken, from which three candidate algorithm designs were developed. A custom Windows-based software program was implemented to facilitate multi-modality development and validation of the segmentation algorithms.
The performance of the algorithms was examined in clinical HRCT images. A centre-likelihood (CL) ray-casting algorithm was found to be the most suitable algorithm due to its speed and reliability in semi-automatic segmentation and tracking of the airway wall. Several novel refinements were demonstrated to improve the CL algorithm’s robustness in HRCT lung data.
The performance of the CL algorithm was then quantified in two-dimensional simulated data to optimise customisable parameters such as edge-detection method, interpolation and number of rays. Novel correction equations to counter the effects of volume averaging and airway orientation angle were derived and demonstrated in three-dimensional simulated data.
The optimal CL algorithm was validated with HRCT data using a plastic phantom and a pig lung phantom matched to micro-CT. Accuracy was found to be improved compared to previous studies using similar methods. The volume averaging correction was found to improve precision and accuracy in the plastic phantom but not in the pig lung phantom. When tested in a clinical setting the results of the optimised CL algorithm was in agreement with the results of other measures of lung function.
The thesis concludes that the relative contributions of confounders of airway measurement have been quantified in simulated data and the CL algorithm’s performance has been validated in a plastic phantom as well as animal model. This validation protocol has improved the accuracy and precision of measurements made using the CL algorithm