Abstract
242507
Introduction: Myocardial perfusion SPECT (MPS) is a widely used imaging procedure. Current clinical protocols for acquiring MPS images are generally similar across most patients. However, individuals with atypical anatomical features, such as large breasts, often have suboptimal image quality when conventional protocols are used [1]. Finding solutions to enhance image quality for such patients without resorting to increased dose remains challenging. One potential approach to address this issue involves personalizing image acquisition protocols based on the specific characteristics of each patient. Current conventional acquisition spends the same amount of time at each projection angle; however, it is known different projection views contain different amounts of information for perfusion defect detection on MPS [2], providing an opportunity for patient-specific adapted protocols. This can be done using information from a CT scan, often acquired for purposes of attenuation compensation, which provides the anatomical characteristics of the patient. Towards this, we propose a strategy for patient-specific protocol optimization for MPS, namely PREcision SPECT (PRESPECT), to improve image quality of MPS images as measured on the task of detecting cardiac defects.
Methods: The proposed PRESPECT approach is designed to optimize acquisition times at each projection view such that the performance of an observer on detecting defects is maximized. A loss function based on acquisition time and information for defect detection of each projection is optimized, with minimum and maximum time constraints for all projection views to enable reconstruction. We evaluated the performance of PRESPECT versus conventional uniform acquisition on the task of defect detection in an anthropomorphic observer study on low dose SPECT. We generated N = 15 female virtual patients with different anatomical parameters, including height and myocardium dimensions, with breast volumes above 99th percentile [3]. Then, for each patient, a test set was generated, consisting of N = 800 extended cardiac and torso (XCAT) phantoms [4]. These phantoms had fixed anatomy (the anatomy of each patient), but varying activity uptake sampled from a representative distribution [5]. Half of the phantoms had a perfusion defect, where the extents and locations of the defect varied. Image acquisition of the phantoms was simulated by SIMIND [6] for both PRESPECT and conventional uniform acquisition, yielding SPECT projections at 10% of normal dose. The projections were reconstructed using ordered subsets expectation maximization with attenuation correction, followed by Butterworth low-pass filtering and reorientation to short axis view. An anthropomorphic channelized Hotelling observer (CHO) was implemented for defect detection. Performance was measured by area under the ROC curve (AUC), and the AUCs of PRESPECT and conventional uniform acquisition were compared using DeLong’s test [7].
Results: We observed that for all 15 patients, PRESPECT had significantly higher AUC (p<0.05) than conventional uniform acquisition. These improvements were seen over a wide range of AUC values. Further analysis revealed a significant improvement in mean signal difference (and thus signal-to-noise ratio) for PRESPECT. Sample reconstructions also indicated improved visualization of defects on PRESPECT. Further, PRESPECT-suggested protocols spent more time on views closest to the heart, since those projection views maximize the information for defect detection.
Conclusions: In this study, PRESPECT yielded improved defect detection performance over conventional uniform acquisition for patients with large breast volumes. These observations motivate further evaluation of PRESPECT on clinical data, as well as similar optimizations for other tomographic nuclear imaging applications. More broadly, these results indicate that personalized approaches to imaging, especially for outlier patients such as those with large breasts, may yield improved image quality.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.