Multi-pinhole SPECT drops the CT attenuation map: one network separates overlapping projections and corrects attenuation and scatter
Synopsis
Multi-pinhole SPECT gains sensitivity but stacks overlapping projections with attenuation, scatter, quantum noise and aperture effects, and attenuation correction usually needs an attenuation coefficient map from X-ray CT. A U-Net++ convolutional network takes Monte Carlo-simulated overlapping projections as input and ray-traced ideal projections as target, separating overlaps and correcting the degradations together before maximum-likelihood expectation-maximization reconstruction. On 60 public brain cases under quasi-stationary acquisition, it reached NMSE 0.18±0.02, PSNR 18.42±0.69 dB and SSIM 0.81±0.04, better than 11-pinhole reconstruction and fan-beam reconstruction with 20-fold longer acquisition.
Interpretation
Multi-pinhole SPECT can separate overlapping projections and correct attenuation, scatter, quantum noise and aperture effects in one step without a CT-derived attenuation map. Multi-pinhole imaging previously faced projection overlap and multiple physical degradations, and attenuation correction typically required an attenuation coefficient map derived from X-ray CT. On 126 test images from a public brain dataset, corrected projections were reconstructed with maximum-likelihood expectation-maximization and compared by one-sided t-tests against 11-pinhole and fan-beam reconstruction.
The network learns a mapping from overlapping projections contaminated by physical effects to ideal projections computed by ray tracing without overlap, attenuation or scatter. Training inputs came from Monte Carlo photon transport simulations of overlapping projections, while targets were ray-traced clean projections, so the network does not model each degradation factor explicitly. Training, validation and testing used 4,050, 144 and 126 images respectively, on a three-detector stationary SPECT system with a 44-pinhole collimator of 2 mm pinhole diameter.
Under quasi-stationary acquisition, the method's quantitative metrics exceed both comparison reconstructions. 11-pinhole reconstruction gave NMSE 0.23±0.05, PSNR 15.72±1.27 dB and SSIM 0.78±0.06; fan-beam reconstruction with 20-fold longer acquisition gave NMSE 0.20±0.03, PSNR 15.93±1.12 dB and SSIM 0.78±0.05. The method gave NMSE 0.18±0.02, PSNR 18.42±0.69 dB and SSIM 0.81±0.04, with evaluation also including visual inspection and regional uptake bias.
Perspective
The conclusion rests on Monte Carlo-simulated overlapping projections and a public brain image dataset, with a three-detector stationary SPECT system using a 44-pinhole collimator of 2 mm pinhole diameter, and the evaluated object is brain images reconstructed by maximum-likelihood expectation-maximization.
All results come from simulated projection data and have not been validated on real acquired projections; quasi-stationary acquisition is the specific condition the reported metrics correspond to, and performance under other acquisition schemes is not given in the material.