Comparative Evaluation of Gaussian Filtering, U-Net, and MedGAN for Restoration of Simulated Low-Count Whole-Body Bone Scintigraphy: Image Quality Assessment with Auxiliary BONENAVI Analysis

Objectives: To compare Gaussian filtering, U-Net, and MedGAN for restoration of simulated low-count whole-body bone scintigraphy and to explore the effects of these image-processing methods on BONENAVI-derived quantitative outputs. Material and Methods: This single-centre retrospective study included 210 male patients with prostate cancer who underwent whole-body bone scintigraphy. Simulated low-count images corresponding to 70%, 50%, 30%, and 10% of the original counts were generated by Poisson resampling. Gaussian filtering was optimised for each count level, and U-Net and MedGAN were trained for image restoration. Image quality was evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) against full-count reference images. An expanded auxiliary BONENAVI analysis was performed on the 30 patients from the independent held-out test set at the 30% count level to examine bone scan index and hotspot counts in full-count images, unprocessed 30%-count images, Gaussian-filtered images, U-Net-restored images, and MedGAN-restored images. Results: MedGAN showed the most consistent restoration performance across count levels. At the 30% count level, MedGAN achieved the highest PSNR (44.05 ± 5.02 dB) and SSIM (0.96510 ± 0.04230). At the 10% count level, U-Net and MedGAN showed comparable restoration performance, whereas MedGAN showed clearer advantages at the 30% and 50% count levels. In the expanded auxiliary BONENAVI analysis, 30%-count images overestimated bone scan index and hotspot counts, whereas Gaussian-filtered, U-Net-restored, and MedGAN-restored images reduced these deviations toward the full-count values to varying degrees. Conclusion: In this preliminary technical validation using simulated low-count data, MedGAN showed more consistent restoration performance than Gaussian filtering and U-Net, particularly at the 30% to 50% count levels. The expanded auxiliary BONENAVI analysis indicated that image processing affected BONENAVI-derived indices, with denoising and restoration reducing low-count-related deviations to varying degrees. These findings should not be interpreted as evidence supporting immediate routine clinical implementation, and further validation with prospectively acquired low-count or shortened-time data and reader-based diagnostic evaluation is warranted.