Abstract 5016: Enhancing prostate pelvic multimodality data generating with conditional generative models: A Pix2Pix-based approach for MRI-to-PET synthesis

Abstract Prostate cancer remains the second leading cause of cancer-related mortality among men globally, underscoring the critical need for early detection of treatment failure and effective assessment of metastatic risk. Positron Emission Tomography (PET), particularly with prostate-specific membrane antigen (PSMA) tracers, has demonstrated superior sensitivity for identifying prostate lesions, including metastases. However, the accessibility of PET imaging is often limited by its high costs and associated radiation exposure. To overcome these challenges, we developed a deep learning model to synthesize PET images from Magnetic Resonance Imaging (MRI) image, facilitating treatment response evaluation.High-resolution T2-weighted MRI and PSMA PET images acquired within a close timeframe were retrieved from 10 prostate cancer patients, who underwent definitive radiotherapy. The PSMA-PET scans were registered to the MRI images within the Eclipse (Varian Medical Systems), and the PET images were cropped to match the same size and resolution of MRI images, resulting in 321 pairs of MRI-PET 2D images. Preprocessing included grayscale transformation, z-score normalization, and pixel value inversion to enhance model learning. A Pix2Pix framework was implemented, employing a U-Net generator and a PatchGAN discriminator. The loss function used consisted of a combination of adversarial loss, to ensure the realism of the generated images, and L1 loss, to maintain pixel-wise consistency between the generated and target images. Model evaluation was performed using leave-one-out-cross-validation (LOOCV), where all slices from one patient was used for testing and the remaining 9 patients for training.The model achieved an average Peak Signal-to-Noise Ratio (PSNR) of 14.55 and a Structural Similarity Index Measure (SSIM) of 0.648. Although these quantitative metrics were moderate, qualitative evaluation demonstrated precise and clinically meaningful localization of lesions, offering utility for aiding physician visualization of high-risk regions.This study highlighted the feasibility of leveraging Pix2Pix-based conditional generative models for synthesizing PET-equivalent images from MRI data as a cost-effective alternative to enhance prostate cancer imaging. Future efforts will focus on expanding the dataset and investigating advanced architectures, including 3D-to-3D generative models and diffusion techniques, to further improve the accuracy of prostate lesion localization and clinical applicability. Citation Format: Ruobing Liu, Shuo Wang, Shibiao Wan, Jieqiong Wang. Enhancing prostate pelvic multimodality data generating with conditional generative models: A Pix2Pix-based approach for MRI-to-PET synthesis [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5016.