Effects of reconstruction filter and input parameter variability on object detectability in CT imaging

The purpose of this work is to investigate and quantify the effects of technical parameter variability and reconstruction algorithm on image quality and object detectability. To accomplish this, metrics of both noise and signal to noise ratio (SNR) are explored and then applied in object detection tasks using a computer aided diagnosis (CAD) system.;The noise power spectrum (NPS) is investigated as a noise metric in that it describes both the magnitude of noise and the spatial characteristics of noise that are introduced by the reconstruction algorithm. The NPS was found to be much more robust than the conventional standard deviation metric. The noise equivalent quanta (NEQ) is also studied as a tool for comparing effects of acquisition parameters (esp. mAs) on noise and, as NEQ is not influenced by reconstruction filter or other post-processing, its utility for comparison across different techniques and manufacturers is demonstrated.;The Ideal Bayesian Observer (IBO) and Non-Prewhitening Matched Filter (NPWMF) are investigated as SNR metrics under a variety of acquisition and reconstruction conditions. The signal and noise processes of image formation were studied individually, which allowed for analysis of their separate effects on the overall SNR. The SNR metrics were found to characterize the influence of reconstruction filter and technical parameter variability with high sensitivity.;To correlate the above SNR metrics with detection, signal images were combined with noise images and passed to a CAD system. A simulated lung nodule detection task was performed on a series of objects of increasing contrast. The average minimum contrast detected and corresponding IBO and NPWMF SNR values were recorded over 100 trials for each reconstruction filter and technical parameter condition studied. Among the trends discovered, it was found that detectability scales with SNR as mAs is varied. Furthermore, the CAD system appears to under-perform when sharp algorithms are used.;Conclusion. Robust noise metrics and SNR metrics were explored and applied under a variety of detections tasks. The results offer insight into both potential improvements for CAD, as well as for improving protocol design.