Deep learning-based simultaneous separation of overlapping projection data and attenuation and scatter correction for...

Multi-pinhole SPECT offers improved sensitivity and spatial resolution; however, overlapping projection data and multiple sources of physical image degradation remain major challenges. We propose a novel convolutional neural network (CNN)-based framework that simultaneously separates overlapping projections and corrects for gamma-ray attenuation, scatter, quantum noise, and aperture effects without requiring an attenuation coefficient map (μ-map) derived from X-ray CT.

Brain imaging was investigated using a three-detector stationary SPECT system equipped with a 44-pinhole collimator (pinhole diameter, 2 mm). A U-Net++-based CNN was trained to separate overlapping projections and correct multiple physical degradation factors. The input comprised overlapping projection data generated by Monte Carlo photon transport simulations, whereas the target data were ideal projections calculated using a ray-tracing method without projection overlap, attenuation, scatter, or other physical effects. A publicly available brain image dataset comprising 60 cases was used for training (4,050 images), validating (144 images) and testing (126 images). Corrected projections were reconstructed using the maximum-likelihood expectation-maximization (ML-EM) method. Image quality was evaluated visually and quantitatively using the normalized mean squared error (NMSE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and regional uptake bias. One-sided t-tests were performed against images reconstructed with 11 non-overlapping pinholes, and a fan-beam collimator method.

The proposed framework accurately separated overlapping projections and simultaneously suppressed attenuation, scatter, quantum noise, and aperture-related degradation. It also reduced artifacts caused by the limited number of projections. The proposed method with a quasi-stationary data acquisition achieved an NMSE of 0.18 ± 0.02, a PSNR of 18.42 ± 0.69 dB, and an SSIM of 0.81 ± 0.04, outperforming 11-pinhole reconstruction (NMSE, 0.23 ± 0.05; PSNR, 15.72 ± 1.27 dB; SSIM, 0.78 ± 0.06) and fan-beam reconstruction with a 20-fold longer acquisition time (NMSE, 0.20 ± 0.03; PSNR, 15.93 ± 1.12 dB; SSIM, 0.78 ± 0.05).

The proposed CNN-based framework enables simultaneous projection separation and correction of multiple physical degradation factors in stationary multi-pinhole SPECT without CT-derived attenuation correction. This approach may reduce radiation exposure and system cost while enabling high-sensitivity, high-resolution, and quantitatively accurate SPECT imaging.