Development, validation, and application of a generic image‐based noise addition method for simulating reduced dose computed tomography images
Njood Alsaihati,Justin Solomon,Erin McCrum,Ehsan Samei
DOI: https://doi.org/10.1002/mp.17444
IF: 4.506
2024-10-11
Medical Physics
Abstract:Background Major efforts in computed tomography (CT) have been focused on reducing radiation dose to patients while maintaining adequate diagnostic quality. To that end, research tools have been developed to simulate reduced‐dose images via either image‐based or projection‐based methods. The former is limited to fully capturing realistic texture, streak, and non‐stationary characteristics of reduced dose, while the latter is impractical clinically. Purpose To develop and validate an image‐based noise addition method that accounts for such attributes while being practical in clinical settings. Methods A noise addition method was developed to add realistic noise in the image domain. The method first estimates the noise power spectrum (NPS) of CT images, which are also forward‐projected to form synthetic projections. The projection data are supplemented with random white noise proportional to their attenuation values. The noise sinogram is then back‐projected onto the image, filtered by the NPS, and scaled according to the desired dose reduction level. The tool was evaluated using both phantom images and patient data. The phantom images were acquired using a multi‐sized image quality phantom (Mercury Phantom 3.0, Duke University), and a thorax anthropomorphic phantom (Lungman Phantom, Kyoto Kagaku) at different dose levels and reconstruction settings. The patient images consisted of two dose levels of various CT examinations and reconstruction settings. The simulated and real reduced‐dose images were compared in terms of the noise magnitude and texture (i.e., NPS average frequency, NPS‐fav). The utility of this methodology was also assessed for routine clinical use for CT protocol review. Results For the phantom images, the percent errors in the noise magnitude between the simulated images and the actual images of the Mercury Phantom and anthropomorphic phantom images were 3.34% and 3.50%, respectively. The difference in fav was 0.07 mm−1 for the Mercury Phantom and 0.06 mm−1 for the anthropomorphic phantom between the simulated and actual images. The average noise magnitude percent error between the simulated and actual patient images was 4.61% with noise texture judged to be visually comparable with some kernel dependencies. When implemented clinically, the tool proved practical to simplify the process of estimating radiation dose reduction for CT protocols, resulting in a 50% dose reduction of our multiple myeloma protocol. Conclusions The method generated simulated CT images with realistic noise properties similar to images acquired at the same radiation exposure without needing access to raw projection data.
radiology, nuclear medicine & medical imaging