Room: Room 205
Purpose: As CT dose is lowered to reduce risks from radiation, image noise may negatively impact diagnosis. To quantify this effect, pilot studies were performed to measure the stochastic noise in CT clinical images and record radiologistsâ€™ estimate of tolerable noise levels. Observer performance was evaluated at varying noise levels to provide preliminary data for future observer experiments.
Methods: Abdomen and pelvic CT scans ordered for pediatric appendicitis (139 normal cases and 101 non-perforated cases) were reviewed by three radiologists. The cases were presented on a custom 3D viewer with axial/coronal views. For each image slice, a pixel-by-pixel noise estimate was calculated by a variance-propagation method, and synthetic noise fields were generated for creating simulated low-dose images. Readers identified and marked the appendix. They recorded the diagnosis with a 3-point confidence rating (3-definite to 1-possibly normal/appendicitis). The reader then adjusted the noise level to the â€˜just tolerableâ€™ amount using a noise-addition tool. At a later time, the readers were presented the image volumes at adaptively selected noise levels, and entered a diagnosis. Observer agreement, sensitivity/specificity, and confidence were analyzed at clinical and high-noise levels.
Results: Full-dose image noise levels (standard-deviation) were found to be 20.2HU (+/- 9.8) in the appendix region. Tolerable-noise levels varied by individual (66.4HU, 50.1HU, 42.0HU), indicating a theoretical potential for reducing dose by factors of 4-12X. When viewing images with noise greater than the designated tolerable noise levels, confidence fell from 2.5 to 1.2, agreement from 0.93 to 0.61, and sensitivity from 0.97 to 0.64 (with specificity unchanged)
Conclusion: By estimating in-situ stochastic noise and creating synthetic low-dose images, reader performance as function of noise level can be studied quantitatively. Additional comprehensive studies are required to define ALARA operating protocols based on trade-off between observer performance and stochastic noise.