2 Nonlinear algorithms such as model-based reconstruction (MBR) and deep learning (DL) reconstruction have 3 sparked tremendous research interest in recent years. Compared to traditional linear approaches, the nonline- 4 arity of these algorithm transcends traditional signal-to-noise requirement and offer flexibility to draw information 5 from a variety of sources (e.g., statistical model, prior image, dictionary, training data). MBR has enabled numer- 6 ous advancements including low-dose CT and advanced scanning protocols. Deep learning algorithms are rap- 7 idly emerging and have demonstrated superior dose vs. image quality tradeoffs in research settings. However, 8 widespread clinical adoption of nonlinear algorithms has been impeded by the lack of a lack of systematic, quan- 9 titative methods for performance analysis. Nonlinear methods come with numerous dependencies on the imag- 10 ing techniques, the imaging target, and the prior information, and the data itself. The relationship between these 11 dependencies and image quality is often opaque. Furthermore, improper selection of algorithmic parameters can 12 lead to erroneous features (e.g., smaller lesions, texture) in the reconstruction. Therefore, methods to quantify 13 and predict performance permit efficient and quantifiable performance evaluation to provide the robust control 14 and understanding of imaging output necessary for reliable clinical application and regulatory oversight. 15 We propose to establish a robust, predictive framework for performance assessment and optimization that can 16 be generalized to any reconstruction method. We quantify performance in turns of the perturbation response and 17 covariance as a function of imaging techniques, system configurations, patient anatomy, and, importantly, the 18 perturbation itself. The perturbation response quantifies the appearance (e.g., biases, blurs, distortions), and, 19 together with the covariance, allows the computation of more complex metrics such as task-based performance 20 and radiomic measures including size, shape, and texture information. We illustrate utility of the approach in lung 21 imaging with the following specific aims:
Aim 1 : Develop a lesion library and generate perturbations encom- 22 passing clinically relevant features. We will extract lesions from public databases and develop methods lesion 23 emulation in for realistic CT simulation and physical data via 3D printing technology.
Aim 2 : Develop a gener- 24 alized prediction framework for perturbation response and covariance. Using analytical and neural network 25 modeling, we will establish a framework that predicts perturbation response and covariance across imaging 26 scenarios for classes of algorithms with increasing data-dependence including MBR with a Huber penalty, MBR 27 with dictionary regularization, and a deep learning reconstructor.
Aim 3 : Develop assessment and optimiza- 28 tion strategies to drive robust, low dose lung screening CT methods. We will optimize and adapt nonlinear 29 algorithms and protocols for lung cancer screening to achieve faithful representations of clinical features. This 30 work has the potential to drive much-needed quantitative assessment standards that directly relate image quality 31 to diagnostic performance and optimal strategies for robust, reliable clinical deployment of nonlinear algorithms. 32

Public Health Relevance

Major research efforts have been devoted to the development of nonlinear reconstruction algorithms ? from model-based reconstruction to deep learning, these algorithms have demonstrated many advantages such as improved image quality, reduced radiation dose, and additional diagnostic information that are not achievable with traditional linear reconstructions. However, only a disproportionately small number has reach the clinic due to the lack of a predictive image quality analysis framework to quantify diagnostic performance, control algorithm behavior, and ensure consistent performance for robust clinical deployment. The propose effort use a combination of analytic and machine learning approaches to drive much-needed quantitative assessment standards that directly relate image quality to diagnostic performance and establish optimal strategies for robust, reliable clinical deployment of nonlinear algorithms.

National Institute of Health (NIH)
National Cancer Institute (NCI)
Research Project (R01)
Project #
Application #
Study Section
Imaging Technology Development Study Section (ITD)
Program Officer
Redmond, George O
Project Start
Project End
Budget Start
Budget End
Support Year
Fiscal Year
Total Cost
Indirect Cost
Johns Hopkins University
Biomedical Engineering
Schools of Medicine
United States
Zip Code