Information Source Mapping in Prior-Image-Based Reconstruction
J Stayman*, J Prince, J Siewerdsen, Johns Hopkins University, Baltimore, MDTU-A-213CD-3 Tuesday 8:00:00 AM - 9:55:00 AM Room: 213CD
Purpose: While reconstruction methods that include prior images can provide spectacular results on highly undersampled and noisy CT data, the reliability of features in volume estimates remains a major question. Specifically, how can one determine whether a specific feature arises from the measurements or the prior image? We present a first step toward answering this question using a novel decomposition of prior-image-based reconstructions that traces the contribution of information from each source (i.e., from the current measurement data and from the prior image).
Methods: We have developed a novel analytic framework for decomposing prior-image-based reconstructions, such as those formed by prior-image-constrained compressive sensing (PICCS) or prior-image penalized likelihood estimation (PI-PLE) reconstructions. Two terms arise - one a function of the measurement data and the other a function of the prior image. Each term also represents an image volume that can be used to form an 'information source map' (ISM), illustrating the spatially varying contribution of information and the source of specific features in the reconstruction.
Results: We apply the analytic decomposition and ISM approach to a simulated lung nodule surveillance scenario where a prior image (without a lung nodule) is incorporated into both PICCS and PI-PLE reconstructions of sparse tomographic data (in which a lung nodule appears). Information source maps correctly identify the measurement data as the source of the nodule feature and quantify the how various adjustable PICCS and PI-PLE parameters increase the reliance on the prior image when larger prior-image parameter strengths are selected.
Conclusion: The proposed ISM analysis allows one to visualize the propagation of information from the prior image and the current measurement data into resulting reconstructions. These maps quantify and provide confidence that specific features arise from the current data (and not the prior image), identify regions of change, and guide selection of reconstruction parameters.