Deep Learning-Based Attenuation Map Generation for Low-Dose and Few-Angle Dedicated Cardiac SPECT

Introduction: Myocardial perfusion imaging (MPI) by single-photon emission computed tomography (SPECT) is widely applied for the non-invasive clinical diagnosis of cardiovascular diseases. Attenuation maps (μ-maps) derived from computed tomography (CT) scans are commonly utilized for attenuation correction (AC) of cardiac SPECT and has been shown to improve diagnosis of CAD. However, CT scans can lead to many issues including additional hardware expenses, additional radiation exposure, SPECT-CT misalignments, etc. In addition, the majority of SPECT systems on the market are SPECT-only without CT. Previous studies have shown promising results in deep learning-based μ-map generation from dedicated cardiac SPECT images. However, these results were generated based on SPECT images reconstructed using full-dose (FD) and full-angle projections. In clinical practice, reducing the dose of the radioactive tracers in SPECT scans is important to minimize radiation exposure. Additionally, collecting projections from fewer projection angles using fewer detectors can largely reduce the hardware expense of the dedicated scanners. However, low-dose (LD) or few-angle projections result in noisier reconstructed images. In this paper, we investigated the feasibility of generating μ-maps from SPECT images reconstructed using LD and few-angle projections on dedicated cardiac SPECT scanners, and then compared the generated μ-maps with those using FD and full-angle projections.

Methods: The study dataset consists of 500 anonymized clinical hybrid one-day SPECT/CT stress/rest MPI studies with list-mode data. Each study was acquired following the injection of ^99mTc-tetrofosmin on a GE Discovery NM/CT 570c dedicated scanner at Yale New Haven Hospital. The GE 570c stationary scanner consists of 19 cadmium zinc telluride (CZT) detectors with pinhole collimators arranged on a cylindrical surface in three columns. Few-angle projections were produced by selecting only the 9 angles (9A) in the central column of the full 19-angle (19A) projections, simulating the cost-effective GE MyoSPECT ES geometry. LD projections of 10% dose were simulated by randomly decimating the normal FD list-mode data at a 10% downsampling rate. The LD-9A SPECT images of one photopeak (133–148 keV) and three scatter windows (55–80, 80–105, 105–130 keV) were channel-wise concatenated as the input of neural networks to generate the predicted μ-maps. The μ-maps generated from FD-19A SPECT images were used as references for comparison. 200, 100, and 200 cases were used for training, validation, and testing, respectively. Quantitative metrics including the voxel-wise normalized mean squared error (NMSE) and correlation coefficient (Corr. Coef.) were applied to evaluate the voxel-wise accuracy of the predicted μ-maps and AC SPECT images.

Results: The generated μ-maps using LD-9A SPECT images are highly consistent with that using FD-19A SPECT images. The correlation maps of the generated μ-maps from LD-9A images demonstrate quite similar distributions to those from FD-19A images. Additionally, the AC SPECT images of the generated μ-maps from LD-9A images also show quite consistent accuracy to those using the generated μ-maps from FD-19A images. The generated μ-maps or AC SPECT images using both photopeak- and scatter-window SPECT images are more accurate than that using photopeak-window SPECT images alone, for both LD-9A (μ-map, 11.51 ± 4.26% vs. 10.61 ± 3.85%, p < 0.001) and FD-19A SPECT images (μ-map, 10.42 ± 3.80% vs. 9.56 ± 3.59%, p < 0.001).

Conclusions: It is feasible to generate μ-maps from dedicated cardiac SPECT images reconstructed using LD and few-angle projections, potentially for cost-effective CZT-based scanners. Concatenating both photopeak- and scatter-window images leads to better results than that using photopeak-window images alone.

• Download figure
• Open in new tab
• Download powerpoint