Stephan Naunheim

Dr. Stephan Naunheim

orcid

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 October 2023 in IEEE Transactions on Neural Networks and Learning Systems
Verified
Journal Article
Open access
SNStephan Naunheim
Stephan Naunheim
YKYannick Kuhl
Yannick Kuhl
DSDavid Schug
David Schug
VSVolkmar Schulz
Volkmar Schulz
FMFlorian Mueller
Florian Mueller
Artificial intelligence (AI) is entering medical imaging, mainly enhancing image reconstruction. Nevertheless, improvements throughout the entire processing, from signal detection to computation, potentially offer significant benefits. This work presents a novel and versatile approach to detector optimization using machine learning (ML) and residual physics. We apply the concept to positron emission tomography (PET), intending to improve the coincidence time resolution (CTR). PET visualizes metabolic processes in the body by detecting photons with scintillation detectors. Improved CTR performance offers the advantage of reducing radioactive dose exposure for patients. Modern PET detectors with sophisticated concepts and read-out topologies represent complex physical and electronic systems requiring dedicated calibration techniques. Traditional methods primarily depend on analytical formulations successfully describing the main detector characteristics. However, when accounting for higher-order effects, additional complexities arise matching theoretical models to experimental reality. Our work addresses this challenge by combining traditional calibration with AI and residual physics, presenting a highly promising approach. We present a residual physics-based strategy using gradient tree boosting and physics-guided data generation. The explainable AI framework SHapley Additive exPlanations (SHAPs) was used to identify known physical effects with learned patterns. In addition, the models were tested against basic physical laws. We were able to improve the CTR significantly (more than 20%) for clinically relevant detectors of 19height, reaching CTRs of 185(450–550 keV).

9 January 2023 in Physics in Medicine & Biology
Verified
Journal Article
Open access
SNStephan Naunheim
Stephan Naunheim
YKYannick Kuhl
Yannick Kuhl
TSTorsten Solf
Torsten Solf
DSDavid Schug
David Schug
VSVolkmar Schulz
Volkmar Schulz
FMFlorian Mueller
Florian Mueller
Objective. Positron emission tomography (PET) detectors providing attractive coincidence time resolutions (CTRs) offer time-of-flight information, resulting in an improved signal-to-noise ratio of the PET image. In applications with photosensor arrays that employ timestampers for individual channels, timestamps typically are not time synchronized, introducing time skews due to different signal pathways. The scintillator topology and transportation of the scintillation light might provoke further skews. If not accounted for these effects, the achievable CTR deteriorates. We studied a convex timing calibration based on a matrix equation. In this work, we extended the calibration concept to arbitrary structures targeting different aspects of the time skews and focusing on optimizing the CTR performance for detector characterization. The radiation source distribution, the stability of the estimations, and the energy dependence of calibration data are subject to the analysis. Approach. A coincidence setup, equipped with a semi-monolithic detector comprising 8 LYSO slabs, each 3.9 mm × 31.9 mm × 19.0 mm, and a one-to-one coupled detector with 8 × 8 LYSO segments of 3.9 mm × 3.9 mm × 19.0 mm volume is used. Both scintillators utilize a dSiPM (DPC3200-22-44, Philips Digital Photon Counting) operated in first photon trigger. The calibration was also conducted with solely one-to-one coupled detectors and extrapolated for a slab-only setup. Main results. All analyzed hyperparameters show a strong influence on the calibration. Using multiple radiation positions improved the skew estimation. The statistical significance of the calibration dataset and the utilized energy window was of great importance. Compared to a one-to-one coupled detector pair achieving CTRs of 224 ps the slab detector configuration reached CTRs down to 222 ps, demonstrating that slabs can compete with a clinically used segmented detector design. Significance. This is the first work that systematically studies the influence of hyperparameters on skew estimation and proposes an extension to arbitrary calibration structures (e.g. scintillator volumes) of a known calibration technique.

12 January 2024 in Medical Physics
Verified
Journal Article
Open access
YKYannick Kuhl
Yannick Kuhl
FMFlorian Mueller
Florian Mueller
SNStephan Naunheim
Stephan Naunheim
MBMatthias Bovelett
Matthias Bovelett
JLJanko Lambertus
Janko Lambertus
VSVolkmar Schulz
Volkmar Schulz
Background: Preclinical research and organ‐dedicated applications use and require high (spatial‐)resolution positron emission tomography (PET) detectors to visualize small structures (early) and understand biological processes at a finer level of detail. Researchers seeking to improve detector and image spatial resolution have explored various detector designs. Current commercial high‐resolution systems often employ finely pixelated or monolithic scintillators, each with its limitations.Purpose: We present a semi‐monolithic detector, tailored for high‐resolution PET applications with a spatial resolution in the range of 1 mm or better, merging concepts of monolithic and pixelated crystals. The detector features LYSO slabs measuring (24 × 10 × 1) mm3, coupled to a 12 × 12 readout channel photosensor with 4 mm pitch. The slabs are grouped in two arrays of 44 slabs each to achieve a higher optical photon density despite the fine segmentation.Methods: We employ a fan beam collimator for fast calibration to train machine‐learning‐based positioning models for all three dimensions, including slab identification and depth‐of‐interaction (DOI), utilizing gradient tree boosting (GTB). The data for all dimensions was acquired in less than 2 h. Energy calculation was based on a position‐dependent energy calibration. Using an analytical timing calibration, time skews were corrected for coincidence timing resolution (CTR) estimation.Results: Leveraging machine‐learning‐based calibration in all three dimensions, we achieved high detector spatial resolution: down to 1.18 mm full width at half maximum (FWHM) detector spatial resolution and 0.75 mm mean absolute error (MAE) in the planar‐monolithic direction, and 2.14 mm FWHM and 1.03 mm MAE for DOI at an energy window of (435–585) keV. Correct slab interaction identification in planar‐segmented direction exceeded 80%, alongside an energy resolution of 12.7% and a CTR of 450 ps FWHM.Conclusions: The introduced finely segmented, high‐resolution slab detector demonstrates appealing performance characteristics suitable for high‐resolution PET applications. The current benchtop‐based detector calibration routine allows these detectors to be used in PET systems.

15 May 2023 in IEEE Transactions on Radiation and Plasma Medical Sciences
Verified
Journal Article
Open access
YKYannick Kuhl
Yannick Kuhl
SNStephan Naunheim
Stephan Naunheim
DSDavid Schug
David Schug
VSVolkmar Schulz
Volkmar Schulz
FMFlorian Mueller
Florian Mueller
Typical positron emission tomography (PET) detectors consist of one-layer segmented scintillators coupled to silicon photomultipliers (SiPMs). Light-sharing detectors, e.g., semi-monoliths, additionally provide depth-of-interaction (DOI) estimation, performing best when calibrated individually. To establish those designs in large PET systems, scalable (re-)calibration methods are needed, possibly transferable to assembled systems. Here, two DOI calibration methods, potentially allowing in-system calibration, are evaluated and compared with an established calibration scheme. Both methods are based on angular detector irradiation using a fan-beam slit collimator and gradient tree boosting (GTB) for 3-Dimensional (3D) position estimation. The positioning performance was assessed for irradiation angles between 0∘ (lateral) and 90∘ (detector normal). With lateral irradiation, a unique DOI position is given, whereas with angular irradiation a gamma-individual reference position must be retrieved. The first method employs one angular beam and calculates DOI from the beam path and planar position estimation. The second method uses two intersecting beams. The intersection defines DOI for gamma interactions that are spatially localized there. Those gamma photons are identified by light distribution comparison using a nearest-neighbor routine. The methods were evaluated on a semimonolithic LYSO slab detector (32×3.9×19 mm slabs). Both methods performed similarly to the benchmark lateral irradiation within 1 %, and 6 %, respectively, for shallow irradiation angles up to 45∘.

26 July 2024 in Physics in Medicine & Biology
Verified
Journal Article
Open access
SNStephan Naunheim
Stephan Naunheim
FMFlorian Mueller
Florian Mueller
VNVanessa Nadig
Vanessa Nadig
YKYannick Kuhl
Yannick Kuhl
JBJohannes Breuer
Johannes Breuer
VSVolkmar Schulz
Volkmar Schulz
Objective. Modern PET scanners offer precise TOF information, improving the SNR of the reconstructed images. Timing calibrations are performed to reduce the worsening effects of the system components and provide valuable TOF information. Traditional calibration procedures often provide static or linear corrections, with the drawback that higher-order skews or event-to-event corrections are not addressed. Novel research demonstrated significant improvements in the reachable timing resolutions when combining conventional calibration approaches with machine learning, with the disadvantage of extensive calibration times infeasible for a clinical application. In this work, we made the first steps towards an in-system application and analyzed the effects of varying data sparsity on a machine learning timing calibration, aiming to accelerate the calibration time. Furthermore, we demonstrated the versatility of our calibration concept by applying the procedure for the first time to analog readout technology. Approach. We modified experimentally acquired calibration data used for training regarding their statistical and spatial sparsity, mimicking reduced measurement time and variability of the training data. Trained models were tested on unseen test data, characterized by fine spatial sampling and rich statistics. In total, 80 decision tree models with the same hyperparameter settings, were trained and holistically evaluated regarding data scientific, physics-based, and PET-based quality criteria. Main results. The calibration procedure can be heavily reduced from several days to some minutes without sacrificing quality and still significantly improving the timing resolution from ( 304 ± 5 ) ps to ( 216 ± 1 ) ps compared to conventionally used analytical calibration methods.Significance. This work serves as the first step in making the developed machine learning-based calibration suitable for an in-system application to profit from the method’s capabilities on the system level. Furthermore, this work demonstrates the functionality of the methodology on detectors using analog readout technology. The proposed holistic evaluation criteria here serve as a guideline for future evaluations of machine learning-based calibration approaches.

26 January 2026 in arXiv
Verified
Preprint
Open access
PDF file
SNStephan Naunheim
Stephan Naunheim
BPBrandon Pardi
Brandon Pardi
GMGuneet Mummaneni
Guneet Mummaneni
CTCarlotta Trigila
Carlotta Trigila
EREmilie Roncali
Emilie Roncali
Monte Carlo simulations of optical photon transport are computationally prohibitive for large-scale optical systems including detector arrays and PET systems, restricting practical use to single-crystal studies. This work presents an enhanced conditional generative adversarial network (optiGAN) replacing optical simulations at the crystal array level, extending our single-crystal approach to a 3x3 BGO array. We enhance the Wasserstein-GAN framework with Fourier feature encoding, a learnable latent mapping network, and a physics-informed loss enforcing momentum conservation. Training data is reduced eight-fold by exploiting symmetry. Evaluation employs three studies: a full array evaluation testing generalization from the fundamental domain to the complete geometry, a high-resolution study probing out-of-distribution generalization to untrained positions, and a pencil beam $γ$-photon study assessing practical applicability for experimental detector characterization. Performance is benchmarked against GATE10/Geant4 ground truth, using intrinsic fluctuations between independent Monte Carlo runs as baseline. OptiGAN achieves sliced Wasserstein similarity within 3$σ$-agreement of the baseline across all conditions, demonstrating successful generalization to the full array. The model transitions from electron-emission training data to realistic $γ$-photon interactions, producing flood maps that reproduce characteristic patterns including photopeak clusters and inter-crystal scatter lines. This proof-of-concept demonstrates that physics-informed generative models can accurately simulate optical photon transport in segmented scintillator arrays. The reproduction of experimentally relevant flood map features validates optiGAN for PET detector development and establishes a foundation for models generalizing across diverse array configurations.