سامانه همایش‌ها TSTA | ثبت‌نام و ارسال مقاله کنفرانس‌ها و کنگره‌های ملی

صفحه اصلی / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Comparative Assessment of U-Net and Pix2Pix for Applying Direct Attenuation Correction in the Image Domain in 68Ga-PSMA PET/CT Imaging

نویسندگان :

Negin Hamidiyan¹ Hadi Taleshi Ahangari² Pardis Ghafarian³ Hossein Arabi⁴ Mohammad Reza Ay⁵

1- دانشگاه علوم پزشکی سمنان 2- دانشگاه علوم پزشکی سمنان 3- دانشگاه علوم پزشکی شهید بهشتی 4- Geneva University Hospital 5- دانشگاه علوم پزشکی تهران

کلمات کلیدی :

Attenuation Correction،PET/CT،PSMA،Deep Learning،U-Net،Pix2Pix

چکیده :

Attenuation correction (AC) is crucial for achieving accurate quantitative positron emission tomography (PET) imaging; however, it remains a challenge in dedicated PET systems that lack simultaneous computed tomography (CT) imaging. In recent years, deep learning (DL) approaches have been explored for this purpose, though direct comparisons between models are still limited. In this study, we directly compared the performance of two widely applied DL architectures, U-Net and Pix2Pix, for direct AC of whole-body 68Ga-PSMA PET images using the same set of 95 patient data sets. For each data set, CT-based attenuation-corrected PET (PET-CTAC) was used as the reference. Quantitative evaluation included mean error (ME) of mean of standardized uptake value (SUVmean), normalized root mean square error (NRMSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). Both U-Net and Pix2Pix generated PET images of comparable quality to those from PET-CTAC, but Pix2Pix generally showed better quantitative metrics. Specifically, U-Net achieved ME, NRMSE, SSIM, and PSNR values of 0.037 ± 0.02, 0.006 ± 0.005, 12.88 ± 2.73, and 0.98 ± 0.14, respectively, whereas Pix2Pix achieved 0.015 ± 0.015, 0.005 ± 0.004, 13.93 ± 2.48, and 0.99 ± 0.004. Statistical analysis, using paired t-tests or Wilcoxon signed-rank tests depending on data normality, demonstrated that Pix2Pix produced SUV estimates closer to those of PET-CTAC, with lower bias and variability than U-Net. In conclusion, both DL models enabled direct AC of whole-body 68Ga-PSMA PET, but Pix2Pix provided more accurate and reliable AC when the two models were directly compared, indicating Pix2Pix is the stronger candidate for clinical use in dedicated PET systems without CT imaging.