Compact deep learning pipeline for particle track reconstruction in the pCT detector system
B. Dudás, G. Papp, G. G. Barnaföldi, G. Bírófor the Bergen pCT Collaboration, [1ex] Institute for Physics, Eötvös Loránd UniversityPázmány Péter sétány. 1/A, 1117, Budapest, Hungary · 2026
Read Compact deep learning pipeline for particle track reconstruction in the pCT detector system by B. Dudás, G. Papp, G. G. Barnaföldi, G. Bírófor the Bergen pCT Collaboration, [1ex] Institute for Physics, Eötvös Loránd UniversityPázmány Péter sétány. 1/A, 1117, Budapest, Hungary (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-30158.txt.
Introduction
B. Dudás, G. Papp, G. Barnaföldi, and G. Bíró, writing for the Bergen pCT Collaboration, build a compact deep learning pipeline for proton computed tomography. This imaging method needs fast, accurate tracking of each proton. Traditional distance-based matching fails under scattering and track crossings. It is also too slow for clinical image formation. Lightweight neural networks offer a faster route.
The pipeline has two stages working in sequence. A tracking module predicts each proton's expected hit in the next detector layer. That prediction acts as a physically informed prior for matching. Ambiguous tracks get flagged and dropped from the final analysis. A second network estimates each incoming proton's kinetic energy. Training data come from the OpenGATE simulation toolkit. The energy estimates prove accurate enough for current reconstruction methods. Deep learning parallelism keeps the whole pipeline within clinical time limits.
Catalog record
| CREATOR | B. Dudás, G. Papp, G. G. Barnaföldi, G. Bírófor the Bergen pCT Collaboration, [1ex] Institute for Physics, Eötvös Loránd UniversityPázmány Péter sétány. 1/A, 1117, Budapest, Hungary |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.30158
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Compact deep learning pipeline for particle track reconstruction in the pCT detector system B. Dudás, G. Papp, G. G. Barnaföldi, G. Bírófor the Bergen pCT Collaboration, [1ex] Institute for Physics, Eötvös Loránd UniversityPázmány Péter sétány. 1/A, 1117, Budapest, Hungary Abstract Proton computed tomography (pCT) requires both fast and accurate reconstruction of particle trajectories and kinetic energies to achieve clinically viable image formation. Traditional distance-based matching algorithms often fail under the combined effects of multiple Coulomb scattering and track crossings and most importantly many of them take too much computation time, motivating the use of lightweight deep learning models that can be evaluated rapidly. In this work, we develop a two-stage reconstruction pipeline consisting of (i) a neural-network-assisted tracking module and (ii) a kinetic-energy estimation model. For the tracking task, compact multilayer perceptrons are trained to predict the expected hit position in the subsequent detector layer, providing a physically informed prior that substantially reduces ambiguities in bipartite matching. Furthermore, ambiguous tracks are flagged and excluded from the final analysis. Our training data is provided by OpenGATE simulation toolkit, both for tracking and energy estimation, where we designed a fully connected network that processes detector hit information. This model predicts the incoming proton kinetic energy with sufficient accuracy for current pCT image reconstruction methods. The entire pipeline benefits from deep-learning parallelism and evaluates particle tracks fast enough for clinical time constraints. Together, these results demonstrate that compact deep learning models can reliably reconstruct particle trajectories and energies in a realistic pCT detector system, offering a computationally efficient and highly accurate alternative to traditional matching and tracking methods. 1 Introduction The research presented in this paper aims to contribute to Proton Computed Tomography [ 10 ] and by that the advancement of hadron therapy [ 33 ] , a form of radiation-based cancer treatment that employs protons and heavier ions such as helium, carbon, and oxygen to achieve greater dose concentration compared to conventional X-ray therapy. A key advantage of hadron therapy arises from the characteristic energy loss behavior of charged particles in matter. In contrast to photons (e.g., X-rays), which deposit a substantial fraction of their energy upon entering the tissue, the energy deposition characteristics of charged particles are defined by the Bragg peak [ 9 ] . This phenomenon ensures that the majority of the energy is released at a well-defined depth, which is determined by the initial beam energy ( E_{beam} ). This physical property enables highly localized dose delivery, allowing the treatment to be tailored so that the Bragg peak aligns with the tumor site. The ionization and thus therapeutic damage is concentrated within the malignant tissue, while sparing surrounding healthy cells. This method critically depends on the accurate determination of the relative stopping power (RSP) of the charged particles used for irradiation. The RSP can be inferred from measurements of the particles’ scattering angles and their residual kinetic energies after traversing a phantom [ 28 ] , which is a simplified representation of target tissues or organs intended for irradiation. Unlike X-rays, the energy loss of charged particles cannot be inferred directly from intensity changes, as their paths are significantly affected by scattering. This means that traditional CT [ 16 ] images result in unwanted uncertainties of the RSP, limiting the opportunities of the hadron therapy [ 31 ] for hadron therapy dosage planning. Within the Bergen pCT collaboration, our aim is to develop a dedicated particle detector system, together with algorithms capable of reconstructing charged-particle trajectories in order to accurately estimate both their residual kinetic energies and scattering angles [ 32 , 1 , 27 ] . In proton computed tomography, the first step involves directing charged particles of relatively high, O(100) MeV energy onto the object. While these energies are modest compared to those used in high-energy physics (GeV–TeV scales), they are sufficiently high that the majority of the energy deposition occurs not within the phantom but in an external detector assembly. In this work, we present two deep learning based algorithms designed to enhance charged particle identification within the pCT detector. The first algorithm reconstructs particle trajectories, while the second predicts the residual kinetic energy. Trajectory reconstruction is formulated as a bipartite matching problem, where the goal is to connect consecutive particle hits based on spatial proximity. 2 Related works In our previous work [ 2 ] , we demonstrated that compact machine learning models can be used to reconstruct particle trajectories. Although individual matching performance was promising, the range of tests was considerably narrower than in the present study, and both matching quality and overall track reconstruction accuracy were substantially lower. The Kalman filter [ 19 ] , which is widely employed in high-energy physics [ 12 , 23 ] , is not suitable for pCT applications because of its high computational cost. Other approaches [ 13 , 11 ] employ neural network architectures such as LSTMs or graph neural networks; however, these methods are typically developed for substantially higher particle energies and often rely on magnetic fields to guide scattering, conditions that do not apply in our experimental setup. An iterative score based algorithm was introduced by the pCT collaboration, which relied on no deep learning usage [ 15 ] . Later a reinforcement learning based method was presented to be able to reconstruct particle trajectories [ 20 ] . 3 Setup Our system follows a granular tracking-calorimeter design composed of two tracking layers behind the phantom, used to determine the scattering angles, followed by forty-one calorimetric layers, used to infer the incoming kinetic energy. Between the calorimetric layers, a 3.5 mm thick aluminum absorber [ 8 ] is placed to decelerate the particles, ensuring that they come to rest within the detector system. However, this material also introduces the possibility of inelastic scattering events. Each detector layer is constructed using ALICE ALPIDE sensors [ 25 ] , which provide binary information on whether a particle has traversed the pixel. From these sensor responses, we reconstruct the detector hits that are defined as the cluster of pixels [ 26 , 29 ] , the size of which reflects the amount of energy deposited as the particle passes through. In Fig. 1 we present the pCT pipeline for detector analysis. Figure 1: Step-by-step pipeline of detector data processing in the pCT pipeline. The dataset used in this study was generated with the OpenGATE [ 17 , 6 , 5 , 3 , 18 ] (also called GATE) toolkit, a medical extension of Geant4 [ 4 ] , the widely used particle physics simulation framework. This simulation environment allows us to produce clean, fully controlled datasets that are well suited for supervised learning and systematic evaluation of our models. OpenGATE immediately produces the centroids and energies of the clusters. The pCT collaboration created the digital twin of the detector system, which we used for particle scattering simulations. The models were trained using water only phantoms with varying thicknesses. An experiment is defined as we beam in 10000 primaries into the detector system and measure their respective detector information. Some of these particles stop in the phantom or suffer such scattering with such an angle that it prevents them entering the detector system. From the particles