Separation and Validation of Overlapping Particle Tracks Using HPT Net and Steal Extension of Topas
Abstract
Purpose
Overlapping tracks complicate particle classification and are unavoidable in pixelated semiconductor detectors due to their random nature. This study introduces a Hybrid Particle Track (HPT) net and a Simultaneous Track Energy And Length (STEAL) extension of TOPAS Monte Carlo simulation to enable robust particle identification.
Methods
Standard U‑net architectures do not summate pixel values when structures overlap. When particle tracks overlap, pixel intensities summate, producing hybrid tracks that cannot be assigned a particle type. To address this issue, a dedicated HPT net was developed to separate overlapping particle tracks. During HPT net preprocessing, summated intensities of overlapping tracks—along with individual track intensities—are provided. Particle track images contain 80–90% dark background, which can lead to artificially high and insensitive DICE scores. To address this, a dosimetry‑based loss function is introduced to improve sensitivity and agreement for particle track areas during HPT net training. During postprocessing, a linearity cut differentiates protons from electrons. Experimental data were acquired using Timepix3 and Timepix2 pixelated detectors in the Varian ProBeam system at three detector inclination angles. Measurement data were processed using Pixet software, then analyzed by HPT net and compared with STEAL simulation. When tracks terminate inside the detector, the STEAL extension accounts for missing track length, enabling accurate comparisons of track energy and length for secondary particles (scattered protons, light ions, and electrons).
Results
Measured primary proton lineal energies agree with STEAL simulation and NIST reference values within one standard deviation. We compare measured energy‑deposit spectra with STEAL results and show strong agreement, with Bhattacharyya coefficients of 0.88 (electron), 0.93 (light ion), and 0.98 (secondary proton). Residual disagreements arise from temporal broadening in the measured data not included in simulations.
Conclusion
HPT net combined with STEAL extension enables separation of overlapped particle tracks, addressing a longstanding challenge in microdosimetry QA for ion therapy.