Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

ST: Fachverband Strahlen- und Medizinphysik

ST 1: Medical Imaging

ST 1.3: Vortrag

Dienstag, 17. März 2026, 11:30–11:45, KH 01.013

From Compton Kinematics to Physics-Informed Neural Network Reconstruction for High-Energy Gamma Imaging — •Yazeed Balasmeh, Atharva Bahekar, Ivor Fleck, Mara Fries, Lars Maczey, and Devanshi Mehta — Experimentelle Teilchenphysik, Center for Particle Physics Siegen, Universität Siegen

Compton cameras enable collimator-free imaging of high-energy gamma rays and are attractive for medical imaging, online range verification in cancer therapy, and targeted alpha therapy. This work presents a physics-informed neural network for Compton-camera image reconstruction that converts measured interaction positions and energies into a 2D source probability map. The objective is a general-purpose model that is computationally efficient, accurate, and stable across operating conditions, including incident energies from about 500 keV to 2 MeV and sources with varying sizes, shapes, and spatial distributions. Physical consistency is encouraged through physics-inspired inputs (e.g., reconstructed scatter angle, interaction-ordering cues, and detector-geometry features) and loss terms that penalize violations of Compton kinematics and constrain predictions to physically feasible Compton cones. A major focus is performance in low-statistics regimes relevant to Ac-225 targeted alpha therapy, where only a small number of detectable 1.57 MeV photons is available. We train and evaluate on Geant4-based simulations spanning count levels, energy spectra, and source configurations, and report energy- and statistics-dependent SSIM and peak-distance error.

Keywords: Medical imaging; Neural Network; Compton camera; Nuclear Medicine; GEANT4