AI-Driven Enhancement of Low-Count Lung Scintigraphy: Optimizing Radiation Safety and Imaging Efficiency with cGANs

Background

Lung ventilation/perfusion (V/Q) scintigraphy is a crucial imaging tool for diagnosing pulmonary embolism (PE). Transitioning exclusively to single photon emission computed tomography (SPECT) acquisitions presents challenges for clinicians accustomed to interpreting traditional planar lung scintigraphy. Additionally, standard acquisition protocols for lung SPECT and planar V/Q scintigraphy, designed to produce high-quality images, are time-consuming and prone to patient motion artifacts. This can lead to patient discomfort and necessitate repeat imaging to obtain clinically acceptable scans. This study explores the potential of conditional generative adversarial networks (cGANs) to generate high-quality pseudo-planar images from either low-dose SPECT projection data or low-count resampled planar images. By leveraging AI-driven enhancement, this approach aims to significantly reduce radiation dose, minimize patient motion artifacts, and decrease the need for redundant imaging while preserving diagnostic accuracy.

Methods

We retrospectively analyzed 704 patients from The Ottawa Hospital who underwent V/Q scans for suspected PE between June 2017 and January 2023. Only perfusion images acquired using 99m-Tc MAA were included. Perfusion images were obtained in six standard projections—anterior (ANT), posterior (POST), right and left posterior oblique (RPO, LPO), and right and left anterior oblique (RAO, LAO)—using eight SPECT machines from two vendors. Each projection was recorded until reaching a total of 600K counts, using a 256 × 256 matrix. Planar acquisition averaged 120.0 ± 54.7 seconds. SPECT acquisition followed immediately, using 128 projections with an acquisition time of 8 seconds per stop and a 128 × 128 matrix. The best-matching SPECT projection for each planar projection was determined automatically by selecting the one with the highest Pearson correlation coefficient. The SPECT/planar count ratio was calculated and used to Poisson-resample the planar images, generating synthetic low-count images with Poisson noise levels matching the SPECT projections.

To enhance these low-count images, a conditional generative adversarial network (cGAN) was trained using an L1+Perceptual+GAN loss function over 300 epochs. The training dataset consisted of synthetic resampled planar images paired with their corresponding full-count planar images. Image intensity values were normalized between 0 and 1, and SPECT projections were upsampled to a 256 × 256 matrix before being fed into the cGAN. Training, validation, and test sets were created with an 80:10:10 split.

To evaluate the model’s performance, mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) were calculated, comparing low-count synthetic images, low-count SPECT projections, and their AI-enhanced results against high-count planar images. Statistical comparisons were performed using the Wilcoxon rank-sum test.

Results

The planar to SPECT projection count ratios were 0.078 ± 0.047. By visual inspection, we show that subsegmental and segmental perfusion defects can still be discerned after enhancement and that no new defects are introduced, demonstrating that diagnostic information is preserved despite ~10-fold count loss. Synthetic and SPECT projections had similar performance metrics both before and after AI enhancement. All performance metrics showed significant improvements with AI enhancement. Specifically, for the SPECT projection, the median ± interquartile range (IQR) of mean squared error (MSE) decreased from 0.59 ± 0.08 to 0.72 ± 0.07, peak signal-to-noise ratio (PSNR) increased from 21.1 ± 1.9 to 27.7 ± 1.6, and structural similarity index (SSIM) improved from 7.75 × 10⁻³ ± 3.39 × 10⁻³ to 1.70 × 10⁻³ ± 8.70 × 10⁻⁴, with all changes being statistically significant (p < 10⁻⁵).

Conclusion

The proposed cGAN model effectively enhances low-count lung scintigraphic images, generating high-quality pseudo-planar images from both low-dose SPECT projection data and low-count resampled planar images. Given that both low-dose and fast-scan paradigms inherently are two sides of the same coin, Poisson noise, which degrades image quality and diagnostic reliability, AI-driven enhancement offers a crucial solution. By mitigating the effects of noise, this approach enables a reduction in administered radiopharmaceutical activity, thereby lessening patient absorbed dose while maintaining diagnostic accuracy. Additionally, it supports faster acquisition protocols, improving clinical efficiency, reducing patient discomfort, and minimizing motion-related artifacts. Future research will explore its applicability to ventilation studies, further reinforcing its role in optimizing nuclear medicine imaging.