Abstract
95
Objectives: PET imaging provides quantitative measurement of physiological and biochemical processes. As the inverse problem to solve for the emission distribution from the acquired data is ill-posed, maximum a posteriori (MAP) algorithms have been developed to regularize the iterative process of image formation. The performance of MAP algorithms strongly depends on the weighting applied to the regularization, which results in different tradeoffs between variance and spatial resolution. Instead of trying to tune the regularization weight for a MAP algorithm, the goal of this study is to take advantage of all the information in images reconstructed with different weights. Artificial neural network (ANN) solves regression or discrimination problems by learning from examples and it has successful applications in the medical image analysis and image interpretation fields. We aim to build a local fusion scheme using an ANN to extract and fuse the information from a collection of MAP reconstructed image versions.
Methods: We designed an ANN serving as a nonlinear multivariate regression function with four hidden layers: two fully-connected lagers, a Hyperbolic Tangent (TanH) activation layer, and a Euclidean loss layer. Using simulated PET data from one subject of the BrainWeb phantom at a high-count level, we obtained the training images from reconstruction with the Green smoothness (GS) MAP algorithm applying three different regularization weights. A set of small patches extracted from the training images at the same location were the inputs fed to the ANN. Using the back-propagation algorithm, the ANN was trained to compute a patch with values as close as possible to those in the patch of the reference image, i.e., the true phantom activity image. We validated the trained ANN by fusing the reconstructed images from simulations at lower count levels of the same subject and various count levels of other subjects.We tested the developed ANN fusion scheme using patient brain PET imaging datasets taken on a second-generation Siemens HRRT scanner with the tracers DPA-173 and Florbetapir. The DPA-173 data corresponded to a 30 min acquisition and the Florbetapir data had two count levels corresponding to 0-70 min and 50-70 min acquisitions. We reconstructed each set of patient data using the GS-MAP algorithms with three different weighting factors and then performed the ANN fusion on the reconstructed images. To ensure the adaptability of the trained ANN model on the patient PET images, we normalized both training and testing images to the range of [0, 1] separately using their respective maximum intensity values. To evaluate the performance of the proposed ANN fusion technique, we calculated the tradeoff between the mean value and the normalized standard deviation on regions of interest (ROIs) of each brain image.
Results: For the DPA-173 study, the ANN fused image reaches the noise measure as low as the GS-MAP reconstruction achieves at heavy regularization while approaches the regional mean value the maximum likelihood (ML) algorithm obtained in most of the ROIs. Similarly, at both count levels of the Florbetapir studies, the ANN fusion suppresses noise as well as the GS-MAP algorithm with large weighting factor while arriving at mean regional values close to what the ML algorithm achieves in most of the ROIs.
Conclusion: An ANN fusion scheme was developed and evaluated to extract and fuse information from images reconstructed using a MAP algorithm of different smoothness weights. The proposed technique reached the lowest variance and best spatial resolution achieved by the GS-MAP algorithm at different weightings in patient PET studies of different tracers at different count levels. With robust performance, the ANN-based fusion contributes to PET image reconstruction by removing the need for tuning the regularization weight of MAP algorithms. Research Support: National Science Foundation ECCS-1454552
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.