Computer aided diagnosis of acute pulmonary embolism

This work focused on using artificial intelligence techniques and texture analysis for computer aided diagnostic tasks in nuclear medicine. Specifically, artificial neural networks (ANNs) were developed to integrate texture features of ventilation-perfusion lung scans into a diagnosis of pulmonary embolism.;First, fractal texture analysis was discovered to be a useful method for assisting in the diagnostic interpretation of perfusion lung scans. The average fractal dimension (FD) of normal lung regions of interest (ROIs) was significantly higher than that of abnormal ROIs. Furthermore, the average FD of abnormal ROIs with pulmonary embolism was significantly lower than the FDs of ROIs with chronic obstructive pulmonary disease present.;Next, a computer aided diagnosis (CAD) tool was developed, which comprised two modules. The first module performed multifractal texture analysis on the ROIs within the posterior view of the perfusion scan. The second module was a decision algorithm that merged the discovered multifractal parameters into a diagnosis regarding the presence or absence of pulmonary embolism. Linear and non-linear decision models were evaluated for the diagnostic task. A consensus neural network significantly outperformed all decision models including the physicians.;A CAD tool was then developed for the diagnosis of pulmonary embolism in whole lungs. This CAD tool used multifractal texture analysis to improve the specificity of perfusion lung scans. This tool was shown to have a significant performance advantage over physicians in the diagnosis of pulmonary embolism from perfusion lung scans, which suggests that the performance of the diagnostic tool is not compromised if used by different operators.;Enhancement of the performance of the previously developed CAD tools was then attempted with the addition of features extracted from the ventilation scan. While the performance of this network was not significantly different than that of the perfusion-only network, its classification performance on a case-by-case basis was better for lungs that were classified as normal or obstructive disease.;Finally, two preliminary studies were performed to further explore the application of texture analysis to the diagnosis of pulmonary embolism. First, spatial texture analysis was examined to determine its diagnostic potential. The spatial texture features were found to capture important texture information that contributed to the ANN's ability to diagnose pulmonary embolism in lungs. Next, the techniques developed throughout this dissertation were validated on digitally acquired data. While this effort was hindered by a lack of data, the results indicate that the use of multifractal texture analysis has potential application for digitally acquired ventilation-perfusion lung scans.;In conclusion, a CAD approach is presented using texture analysis and artificial neural networks to assist physicians in the diagnosis of PE.