Cascaded deep learning for synthesis of ¹⁸F-FDG PET from ¹⁸F-octreotide: model parameter optimization and performance evaluation.

Authors

Abstract

To investigate a cascaded conditional generative adversarial network (cGAN)-based method for synthesizing ¹⁸F-FDG PET images from ¹⁸F-octreotide, optimize model parameter configurations, and evaluate the quality of the synthesized images. Imaging data from 71 tumor patients who underwent dual-tracer PET/CT examinations with ¹⁸F-FDG and ¹⁸F-octreotide between May 2022 and November 2025 were retrospectively collected. A cascaded deep learning architecture (first stage: 3D U-Net; second stage: self-attention 3D V-Net) was employed to evaluate the effects of image preprocessing, resampling size, and patch size on synthesis quality. Three preprocessing strategies (log transformation, normalization, and a combined approach) were compared. Image quality was assessed using five objective metrics (SSIM, PSNR, MSE, MAE, and NMI) and Likert 5-point ratings from three nuclear medicine physicians; liver and lesion-level SUV agreement were further evaluated. In the held-out test cohort (<i>n</i> = 14), the combined preprocessing method (log transformation + normalization) significantly outperformed the individual methods (SSIM: 0.8454 ± 0.0328 vs. 0.7831 ± 0.0337 vs. 0.7438 ± 0.0362, <i>P</i> < 0.001). On the identical held-out test cohort (<i>n</i> = 14), the Group 2 parameter configuration (resampling size: 256 × 256 × 512; patch size: 128 × 128 × 128) outperformed Group 1, with an SSIM of 0.8854 ± 0.0455 and a PSNR of 28.15 ± 3.35 dB, representing relative improvements of 19.1% and 83.5%, respectively, over Group 1 (paired-samples t-test, both <i>P</i> < 0.001). The mean Likert score was 4.53 ± 0.19. Quantitative analysis of liver SUVmean showed no significant difference between synthesized and real images (<i>P</i> = 0.967, paired t-test). Lesion-level quantitative analysis demonstrated high agreement between synthetic and real ¹⁸F-FDG images, with Pearson correlation coefficients of 1.0000, 0.9995, and 0.9991 for SUVmax and SUVmean (<i>n</i> = 14) and SUVpeak (<i>n</i> = 13), respectively (all <i>P</i> < 0.001), no significant differences on paired t-tests (all <i>P</i> > 0.05), and all Bland-Altman differences within the 95% limits of agreement. The cascaded deep learning model enabled cross-tracer synthesis from ¹⁸F-octreotide to ¹⁸F-FDG PET images. In this exploratory, single-center pilot cohort, the optimized configuration demonstrated promising quantitative and visual performance, supporting image-synthesis feasibility and potential clinical utility of a single-injection, dual-parameter molecular imaging approach for neuroendocrine tumors. Prospective multicenter validation is warranted before any clinical application.

Tags

Topics