| Popis: |
Regularization methods are commonly used for noise reduction in SPECT image reconstruction. Total variation (TV) is a well-accepted method for suppressing noise while preserving edges. However, using TV as regularizer implies that the image is piecewise constant. Usually this is untrue in clinical settings and leads to staircasing artifacts. High-order TV (HOTV) and infimal-convolution TV (ICTV) are proposed here to avoid such artifacts. Previously, we proposed the preconditioned alternating projection algorithm (PAPA) to address the nondifferentiability of the TV regularizer. Here we generalize PAPA to solve for these regularizers, and apply them to reconstruction of Monte Carlo-simulated SPECT data. We designed two numerical phantoms with lumpy background, with and without targets, to compare the performance of these algorithms. We analyzed the reconstructed images using the channelized Hotelling observer (CHO). The signal-to-noise ratio of CHO was used as a figure of merit. The hyperparameters of the regularization methods were selected using the point-spread function of point sources inside the phantom. Images reconstructed using PAPA-TV, PAPA-HOTV, and PAPA-ICTV, along with traditional EM-TV were studied. All such images have similar local spatial resolution. Images reconstructed using EM-TV and PAPA-TV show significant staircasing artifacts. While both PAPA-HOTV and PAPA-ICTV reduce these artifacts very well, PAPA-HOTV cannot suppress noise as well as PAPA-ICTV does, without compromising spatial resolution. According to CHO analysis, PAPA-ICTV provides higher lesion detectability, compared to PAPA-HOTV. We conclude that PAPA-ICTV is a superior image-reconstruction method and has high potential for use in clinical settings due to more efficient noise suppression and significantly improved image quality. |
