Room: Exhibit Hall | Forum 6
Purpose: Nonlinear image processing algorithms in CT pose significant challenges in assessment and optimization. Traditional image quality metrics such as impulse and edge response may no longer be meaningful for characterizing nonlinear system response. This work investigates metrics applicable for nonlinear system analysis, and develops a data-driven approach to efficiently predict image properties.
Methods: To account for data-dependence in nonlinear methods, we define perturbation response and its variation as the difference between the mean and the variance of system outputs with and without a stimulus. Metrics were assessed for two nonlinear algorithms: convolutional neural network (CNN) denoising, and model-based reconstruction (MBIR) with an edge-preserving penalty. To develop predictive abilities, we leveraged the universal approximation property of neural networks to model nonlinear responses using a data-driven strategy. The neural network was developed for perturbation response prediction for the MBIR algorithm as a function of lesion sizes, locations, and regularization parameters.
Results: Perturbation response of the CNN shows highly nonlinear behavior. System response to a larger lesion is no longer the superposition of impulse responses, as in linear systems. Lesions of lower contrast may be misrepresented in size. Perturbation variation indicates that certain algorithms may present â€œjitterâ€? in edge positions, indicating that mean response alone may not represent individual image performance and that traditional assessment (e.g. task-based metrics) based on mean response may need to be reconsidered. The neural network predictor yielded excellent agreement with the gold-standard output as a function of all variables.
Conclusion: Perturbation response overcomes limitations of traditional metrics to quantify the nonlinear behavior of tomographic imaging algorithms. A neural network predictor has shown promise in efficient and accurate modeling of nonlinear algorithms. With the rapid proliferation of nonlinear data processing approaches, these tools have great potential to provide meaningful quantitative assessments to facilitate understanding and robust clinical implementation.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH grant R21CA219608.