Geant4-Based Simulation and Image Reconstruction of Tc-99m for Radiopharmaceutical Imaging

Abstract

This paper presents a framework for simulatingand reconstructing radiopharmaceutical images using a Compton camera (CC) for potential use in personalized dosimetry.We developed a Geant4-based Monte Carlo simulation modelof a two-stage CC and a Technetium-99m (Tc-99m) source,simulating gamma-ray emissions and their interactions across multiple camera orientations. This synthetic data was used to evaluate two 3D image reconstruction algorithms: simple back projection (SBP) and a kernel weighted back projection (KWBP) method. While KWBP demonstrated superiority overSBP by producing reconstructions with significantly less noise and better-preserved source geometry, both methods were limitedby artifacts and blurring due to restricted angular sampling. To address these limitations, we integrated a deep learning-based image enhancement step into the reconstruction pipeline. We trained and evaluated two convolutional neural networks, a REDCNN and a U-Net, to denoise the reconstructed images. Our results show that both networks effectively suppressed noise, but the U-Net, trained with a hybrid mean squared error (MSE) and structural similarity index measure (SSIM) loss function, delivered superior performance. Quantitative analysis on a heldout test set showed that the U-Net achieved a lower average MSE of 0.0041, compared to the RED-CNN’s 0.0103. Furthermore, UNet qualitatively outperformed the RED-CNN by more accurately preserving the structural integrity and shape of the sources. These findings establish a two-step strategy—combining physics-based reconstruction with data-driven denoising—as a promising pathway for refining Compton camera images for clinical applications. Index Terms—Compton camera, image reconstruction, radiopharmaceutical imaging, deep learning, gamma-ray