GAN Converts CT to PET for Early Metastases

In a groundbreaking advancement poised to transform prostate cancer diagnostics, researchers have unveiled a novel deep learning approach to synthesize [^18F]PSMA-1007 PET bone images directly from CT scans. This innovative strategy leverages generative adversarial networks (GANs) to produce high-fidelity synthetic PET images, potentially eliminating the need for additional costly and radiation-intensive PET/CT scans. The pioneering study, recently published in the reputable journal BMC Cancer, demonstrates the feasibility and accuracy of this technique in the early detection of bone metastases in prostate cancer patients.

Prostate cancer remains one of the most prevalent malignancies among men worldwide, and its progression often leads to bone metastases—a critical factor influencing patient prognosis and treatment strategy. Conventional detection methods heavily rely on combined imaging modalities such as [^18F]FDG and [^18F]PSMA-1007 PET/CT scans. While effective in visualizing metastatic lesions, these methods are associated with significant drawbacks, including high operational costs and increased radiation exposure to patients. Addressing these limitations, the research team explored deep learning methods to synthesize functional PET images using only structural CT data, thereby promising a non-invasive, cost-effective alternative.

The study amassed a robust dataset comprising paired whole-body [^18F]PSMA-1007 PET/CT images from 152 subjects, carefully curated through retrospective analysis. These included 123 patients clinically and pathologically diagnosed with prostate cancer and 29 with benign lesions serving as comparative controls. The mean patient age was 67.48 years, with an average lesion size of approximately 8.76 millimeters. Such comprehensive data enabled the research to construct detailed bone structure images by preprocessing and segmenting both low-dose CT and PET scans, a crucial step for effective model training.

Central to the methodology was the deployment of two distinct GAN architectures: Pix2pix and CycleGAN. Both models are renowned for their capabilities in image-to-image translation tasks, but they approach the synthesis differently. Pix2pix operates on paired datasets with supervised learning, while CycleGAN leverages unpaired data through cycle consistency to achieve transformation. By training these networks to convert CT bone images into synthetic [^18F]PSMA-1007 PET images, the study rigorously assessed performance across multiple quantitative metrics including mean absolute error (MAE), mean squared error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and importantly, the target-to-background ratio (TBR) relevant for identifying metastatic lesions.

Results from this extensive validation imparted compelling evidence of model efficacy. The Pix2pix model outperformed CycleGAN, attaining an exceptional SSIM of 0.97, indicative of near-perfect structural similarity between synthetic and real PET images. Additionally, a PSNR of 44.96 and low error rates (MSE at 0.80 and MAE at 0.10) underscored the precision of synthetic image generation. Particularly significant was the strong correlation (Pearson’s r > 0.90) observed between TBR values calculated from synthesized versus actual PET bone images—this parameter being critical for differentiating malignant bone lesions from healthy tissue with statistical insignificance in difference (p

Tags: BMC Cancer publicationcost-effective cancer detection solutionsCT to PET conversiondeep learning in medical imagingearly detection of bone metastasesGAN for imaging synthesisgenerative adversarial networks in healthcarenon-invasive cancer imaging methodsprostate cancer diagnosticsprostate cancer imaging advancementsradiation exposure reduction in imagingsynthetic PET imaging technology