Algorithms for automatic lung scan registration

This thesis describes the design, implementation, and performance evaluation of ventilation (V) and perfusion (Q) lung scan registration algorithms. These algorithms transform images, superimpose images, produce composite images, and create V/Q ratios. V/Q functions are the key variable controlling the functionality of gas exchange in the lungs. Analyzing these images helps physicians make clinical decisions in consistent and unbiased ways.;Registration is an essential prerequisite for radionuclide lung scan analysis, classification of diseases, and detection of changes over time. Most current registration methods require some degree of user interaction. This study presents four new fully automated methods to register ventilation and perfusion (V/Q) lung scan images and tests them on a database of 49 pairs, which includes normal and abnormal images. Each V/Q scan was used to create a V/Q composite image, which is a superimposition of the registered ventilation and perfusion images. Color-enhanced V/Q composite images facilitate the measurement of V/Q ratios, the identification of defects, and the detection of temporal change and may assist in patient diagnosis and therapy.;The four algorithms are: the one-dimensional V/Q automatic registration (VQARM), the center-of-object (COO), the two-dimensional linear Pearson correlation with background subtraction (2DLPC), and the conditional entropy (CE) registration algorithms.;All four algorithms assume the ventilation and perfusion images have similar gray-levels. Similar object shapes between ventilation and perfusion images will result in RST-invariance and better error improvement. All algorithms are user friendly fully automated registration methods and result in reasonable V/Q ratios. VQARM, 2DLPC, and CE are categorized as pixel-based automated registration, while COO is feature-based automated registration. Each algorithm offers tradeoffs in terms of speed, complexity, and accuracy under different operating conditions. Considering the error improvement and deviation standard, 2DLPC is best. However, the COO algorithm has the fastest execution time with acceptable results. CE is less sensitive to noise compared with the other algorithms. These approaches provide a completely automated registration mechanism that also accelerates registration and visualization. Color visualization provides more functional details allowing a more comprehensive examination. The results of this research have been used to build an automatic registration system for use by physicians.