Qi, Jinyi, (University of California, Davis, Department of Biomedical Engineering, Davis, USA)
Shortcut:
MIC-18
Date:
Saturday, 14 November, 2026, 9:00 -
11:00
Location:
FEDERICO GARCÍA LORCA
Session type:
MIC Session
Contents
Click on an contribution to preview the abstract content.
The underlined authors are the presenters. The authors in bold are the main authors.
9:00
MIC-18-01
Steerable Kernel Based Emission Tomography Reconstruction (#2137)
H. Wei1, L. Livieratos1, 2, A. J. Reader1
1 Kings College London, School of Biomedical Engineering and Imaging Sciences, London, United Kingdom 2 Kings College London, Guy's and St Thomas' NHS Foundation Trust, London, United Kingdom
Abstract
Kernelized expectation maximization (KEM) has emerged as an effective and practical approach for improving image quality in emission tomographic reconstruction by incorporating high-quality prior information into the iterative reconstruction process. In conventional KEM, the kernel matrix is derived from a prior image and remains fixed throughout reconstruction. This may result in artifacts when the prior image and target emission data are not perfectly matched, especially in cross-modality or cross-scanner settings, such as MR-guided PET reconstruction and PET-guided SPECT reconstruction in theranostic imaging. In those settings, differences in scanner characteristics, acquisition time points, or tracer distributions may cause mismatch between the prior image and the target emission data. In this work, we propose a steerable kernel reconstruction framework that adapts the prior-derived kernel representation according to the measured data. The proposed method dynamically updates the kernel matrix during reconstruction to improve consistency with the measured data. We evaluate the proposed method on 2D brain phantom data for MR-guided PET reconstruction and on 2D thorax phantom data for PET-guided SPECT reconstruction in a theranostic setting. Experimental results demonstrate that the proposed steerable kernel method mitigates mismatch-induced artifacts and achieves improved reconstruction fidelity compared with conventional fixed-kernel approaches. These findings suggest that adaptive kernel steering offers a promising direction for robust prior-guided emission tomography reconstruction under realistic mismatch conditions.
References [1] G. Wang and J. Qi, “PET image reconstruction using kernel method,”IEEE Trans. Med. Imag., vol. 34, no. 1, pp. 61–71, Jan. 2015. [2] E. J. Hu, Y. Shen, P. Wallis et al., “LoRA: Low-rank adaptation of large language models,” 2021, arXiv:2106.09685. [3] K. Thielemans, C. Tsoumpas, S. Mustafovic et al., “STIR: Software for tomographic image reconstruction release 2,” Phys. Med. Biol., vol. 57, no. 4, pp. 867–883, Feb. 2012. [4] L. A. Polson, R. Fedrigo, C. Li et al., “PyTomography: A Python library for medical image reconstruction,” SoftwareX, vol. 29, p. 102020, Feb. 2025.
Cross-Tracer PET Image Reconstruction Using an 18F-FDG-Trained Deep Kernel Model (#2782)
S. Li1, X. Long2, Y. Lv2, G. Liu1, J. Qi1, S. R. Cherry1, R. D. Badawi1, L. Nardo1, B. A. Spencer1, G. Wang1
1 University of California, Davis, California, United States of America 2 United Imaging Healthcare, Shanghai, China
Abstract
PET image quality is affected by limited-count statistics and the ill-posed nature of tomographic image reconstruction. Deep learning methods have shown promise for improving PET image quality. However, most existing models have primarily been developed and validated using 18F-FDG datasets. The generalizability of 18F-FDG- trained models to less common radiotracers, particularly non-18F tracers with distinct biodistributions and physical properties, has not been well demonstrated. As the clinical and research use of diverse PET tracers continues to expand, there is a growing need for unified deep learning reconstruction methods capable of generalizing across tracers. We previously developed a pretrained deep kernel model using a large-scale 18F-FDG total-body PET dataset and demonstrated its feasibility for tomographic image reconstruction in static PET imaging by using a single-frame MLEM image as the image prior. Unlike post-reconstruction denoising methods that learn tracer-specific activity distributions and directly output images, the deep kernel method learns to construct a kernel matrix that reflects spatial correlations, making it potentially transferable across tracers. In this work, we extend the deep kernel pretraining by incorporating additional 18F-FDG data and then investigate whether the 18F-FDG-trained model can be directly applied to other non-18F tracers without any tracer-specific training or fine-tuning. The proposed method was evaluated on fourteen 68Ga-DOTATATE scans acquired with the uEXPLORER total-body scanner and compared against a state-of-the-art deep-learning denoising model that is also trained on 18F-FDG total-body datasets. Results demonstrate that the proposed method significantly reduces image noise and achieves improved lesion quantitative accuracy. This work demonstrates the feasibility of cross-tracer PET reconstruction using an 18F-FDG-trained deep kernel model, supporting future evaluation in larger multi-tracer datasets.
References [1] F. Hashimoto, Y. Onishi, Kibo Ote, H. Tashima, A. J. Reader, and Taiga Yamaya, “Deep learning-based PET image denoising and reconstruction: a review,” Radiological Physics and Technology, Feb. 2024, doi: https://doi.org/10.1007/s12194-024-00780-3. [2] S. Xue et al., “A cross-scanner and cross-tracer deep learning method for the recovery of standard-dose imaging quality from low-dose PET,” European Journal of Nuclear Medicine and Molecular Imaging, Dec. 2021, doi: https://doi.org/10.1007/s00259-021-05644-1. [3] H. Liu, J. Wu, W. Lu, J. A. Onofrey, Y.-H. Liu, and C. Liu, “Noise reduction with cross-tracer and cross-protocol deep transfer learning for low-dose PET,” Physics in Medicine and Biology, vol. 65, no. 18, pp. 185006–185006, Aug. 2020, doi: https://doi.org/10.1088/1361-6560/abae08. [4] B. Yu et al., “Robust whole-body PET image denoising using 3D diffusion models: evaluation across various scanners, tracers, and dose levels,” European Journal of Nuclear Medicine and Molecular Imaging, Feb. 2025, doi: https://doi.org/10.1007/s00259-025-07122-4. [5] G. Wang and J. Qi, “PET Image Reconstruction Using Kernel Method,” IEEE Transactions on Medical Imaging, vol. 34, no. 1, pp. 61–71, Jan. 2015, doi: https://doi.org/10.1109/tmi.2014.2343916. [6] S. Li and G. Wang, “Deep Kernel Representation for Image Reconstruction in PET,” IEEE Transactions on Medical Imaging, vol. 41, no. 11, pp. 3029–3038, Nov. 2022, doi: https://doi.org/10.1109/tmi.2022.3176002. [7] S. Li et al., “Static PET Image Reconstruction Using A Pretrained Deep Kernel Method,” 2025 IEEE Nuclear Science Symposium (NSS), Medical Imaging Conference (MIC) and Room Temperature Semiconductor Detector Conference (RTSD), pp. 1–1, Nov. 2025, doi: https://doi.org/10.1109/nss/mic/rtsd57106.2025.11287136.
Keywords: PET image reconstruction, Cross-tracer, FDG-trained deep kernel model, 68Ga-DOTATATE
9:30
MIC-18-03
List-mode PET Image Reconstruction with Neural Radiance Fields (#1346)
H. Lu1, 2, C. E. Davey1, 2, L. A. Johnston1, 2
1 University of Melbourne, Department of Biomedical Engineering, Parkville, Australia 2 University of Melbourne, Melbourne Brain Centre Imaging Unit, Parkville, Australia
Abstract
Numerous AI-based solutions to PET image reconstruction, an ill-posed inverse problem, have been implemented to enhance image quality or convergence speed, yet most suffer from reliance on high-quality paired dataset and lack of robustness. As an emerging methodology, implicit neural representations (INRs) demonstrate tremendous potential in continuous modelling of spatial distribution. However, the existing INR-based PET reconstruction methods generally suffer from loss of information due to compression in sinograms they utilise as input, and do not include corrections for noise caused by the physical factors like attenuation and normalisation. In this work, we present a self-supervised PET reconstruction method based on neural radiance fields (NeRFs), a special type of INR with list-mode coincidences as input, where sensitivity is incorporated as both the weight for sampling points along each LOR and combined corrections for attenuation and normalisation. LORs with no recorded coincidences function as data-driven regularisation and are augmented as part of the input. Results using simulated and experimental data show that our proposed method can achieve excellent image quality and improved contrast, and outperform conventional iterative algorithms and other AI approaches, particularly in recovering fine details in small hot regions.
AcknowledgmentThis research was supported by a Melbourne Research Scholarship from the University of Melbourne, The University of Melbourne’s Research Computing Services and the Petascale Campus Initiative. The authors acknowledge the facilities and scientific and technical assistance of the National Imaging Facility, a National Collaborative Research Infrastructure Strategy (NCRIS) capability, at the Melbourne Brain Centre Imaging Unit, University of Melbourne.
References [1] B. Yang, P. He, S. Liu, and L. Zhou, “Neural Radiance Fields Enabled Continuous PET Image Modeling and High-Resolution Reconstruction,” Journal of Nuclear Medicine, vol. 65, no. supplement 2, pp. 241396–241396, Jun. 2024. [2] Y. Long et al., “PD-INR: Prior-Driven Implicit Neural Representations for TOF-PET Reconstruction,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, J. C. Gee, D. C. Alexander, J. Hong, J. E. Iglesias, C. H. Sudre, A. Venkataraman, P. Golland, J. H. Kim, and J. Park, Eds., Cham: Springer Nature Switzerland, 2026, pp. 459–469. doi: 10.1007/978-3-032-04947-6_44. [3] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: representing scenes as neural radiance fields for view synthesis,” Commun. ACM, vol. 65, no. 1, pp. 99–106, Dec. 2021, doi: 10.1145/3503250. [4] H. M. Hudson and R. S. Larkin, “Accelerated image reconstruction using ordered subsets of projection data,” IEEE Transactions on Medical Imaging, vol. 13, no. 4, pp. 601–609, Dec. 1994, doi: 10.1109/42.363108. [5] A. R. De Pierro and M. E. B. Yamagishi, “Fast EM-like methods for maximum ‘a posteriori’ estimates in emission tomography,” IEEE Transactions on Medical Imaging, vol. 20, no. 4, pp. 280–288, Apr. 2001, doi: 10.1109/42.921477. [6] K. Gong, C. Catana, J. Qi, and Q. Li, “PET Image Reconstruction Using Deep Image Prior,” IEEE Trans Med Imaging, vol. 38, no. 7, pp. 1655–1665, Jul. 2019, doi: 10.1109/TMI.2018.2888491.
Dynamic image reconstruction for high-resolution SAVANT scanner using non-negative low-rank constraint (#3104)
Y. Najmaoui1, 2, Y. Djebra1, Y. Chemli1, M. Gaudreault2, K. Fontaine1, M. Dhaynaut1, J.-D. Gallezot1, M.-A. Tétrault2, R. Lecomte3, M. Normandin1, G. El Fakhri1, C. Ma1, T. Marin1
1 Yale University, Yale Biomedical Imaging Institute, New Haven, Connecticut, United States of America 2 Université de Sherbrooke, Institut interdisciplinaire d'innovation technologique, Sherbrooke, Québec, Canada 3 Université de Sherbrooke, Centre hospitalier universitaire de Sherbrooke, Sherbrooke, Québec, Canada
Abstract
Dynamic Positron Emission Tomography (PET) imaging is an essential tool for research in neurodegenerative diseases through the modeling of tracer kinetics. However, achieving ultra-high spatial resolution in PET systems tends to reduce their overall geometric sensitivity. This leads to a possible compromise between dynamic frame duration, spatial resolution, and noise. This study presents the application of a non-negative low-rank-constrained reconstruction method tailored for the ultra-high-resolution SAVANT scanner. Traditional algorithms reconstruct PET images frame-by-frame without taking into account the natural spatiotemporal correlations within the data. This redundancy is leveraged by this method to reconstruct all the dynamic frames simultaneously. The method alternates between updating a matrix representing the spatial distribution for each temporal basis, and updating a matrix representing the translation from temporal bases to frame. In order to accommodate for the large memory and computational requirements associated with the SAVANT, we extended the nonnegative low-rank model to support list-mode data and optimized its implementation through Graphical Processing Unit (GPU) acceleration. In vivo evaluations using a dynamic [18F]DG scan demonstrate a visible improvement in image quality when using the proposed method compared to the conventional frame-by-frame MLEM algorithm. The low-rank constraint allows to preserve greater detail and shows sharper anatomical structures, which better showcases the high resolution of the SAVANT system.
AcknowledgmentThis work was supported in part by the National Institutes of Health under the grants R01EB035093, U01EB027003.
References [1] Lecomte, Roger et al. 'Scanner Approaching in Vivo Autoradiographic Neuro Tomography (SAVANT): Progress Towards μL Resolution for Imaging the Human Brain', SNMMI 2022.
[2] Djebra, Yanis et al. ''LR-PET: A Subspace-Based Dynamic PET Imaging via Explicit Non-Negative Low-Rank Factorization", IEEE NSS/MIC 2025. [3] Parra L., Barrett H. "List-mode likelihood: EM algorithm and image quality estimation demonstrated on 2-D PET" IEEE Transactions on Medical Imaging 17, 228 (1998). [4] Najmaoui, Yassir et al. "YRT-PET: An Open-Source GPU-Accelerated Image Reconstruction Engine for Positron Emission Tomography" IEEE Transactions on Radiation and Plasma Medical Sciences 10, 535 (2026). [5] Paatero, P. and Tapper, U. "Positive matrix factorization: A non-negative factor model with optimal utilization of error estimates of data values" Environmetrics 5, 111 (1994). [6] Siddon, R. L. "Fast calculation of the exact radiological path for a three-dimensional CT array" Medical Physics 12, (1985)
A Generalized Tuy-Type Completeness Analysis for X-Ray Scattering Tensor Tomography (#2612)
R. Liu1, 2, R. Deng1, 2, P. Song1, L. Zhang1, 2, Z. Wang1, 2
1 Tsinghua University, Department of Engineering Physics, Beijing, China 2 Radiation Imaging of Ministry of Education, Key Laboratory of Particle, Beijing, China
Abstract
X-ray scattering tensor tomography aims to recover orientation-dependent microstructural information beyond scalar attenuation or scalar dark-field CT. For quantitative reconstruction, however, an aquisition trajectory satisfying Tuy's condition should also provide enough independent angular sensitivity for the tensor model being reconstructed. We formulate this requirement as a generalized Tuy-type completeness condition for scattering tensor tomography, linking conventional geometric visibility with tensor sensitivity. This analysis shows that the former 15-component model is intrinsically incomplete: under each plane, only 12 independent tensor modes can be probed, so arbitrary 15-component fields require additional prior assumptions. In contrast, the reduced 6-component model satisfies the planar completeness requirement under sufficient angular coverage. Planar singular-value tests confirm both the 12-mode limit and the full rank of the reduced model, providing practical criteria for tensor-model selection and aquisition trajectory design and optimization.
AcknowledgmentThis work was supported in part by National Natural Science Foundation of China (No. 62227804), Beijing Municipal Natural Science Foundation (No. L232115) and National Key Research and Development Program of China (No. 2023YFC2605802).
References [1] Kim, J, Kagias, M, Marone, F, Shi, Z & Stampanoni, M 2021, ‘Fast acquisition protocol for X-ray scattering tensor tomography’, Scientific Reports, vol. 11, article no. 23046. [2] Sharma, Y, Schaff, F, Wieczorek, M, Pfeiffer, F & Lasser, T 2017, ‘Design of acquisition schemes and setup geometry for anisotropic X-ray dark-field tomography (AXDT)’, Scientific Reports, vol. 7, article no. 3195. [3] Tuy, HK 1983, ‘An inversion formula for cone-beam reconstruction’, SIAM Journal on Applied Mathematics, vol. 43, no. 3, pp. 546–552. [4] Quinto, ET 1993, ‘Singularities of the X-ray transform and limited data tomography in R² and R³’, SIAM Journal on Mathematical Analysis, vol. 24, no. 5, pp. 1215–1225. [5] Denisjuk, A 2006, ‘Inversion of the X-ray transform for 3D symmetric tensor fields with sources on a curve’, Inverse Problems, vol. 22, no. 2, pp. 399–411. [6] Wieczorek, M, Schaff, F, Pfeiffer, F & Lasser, T 2016, ‘Anisotropic X-ray dark-field tomography: A continuous model and its discretization’, Physical Review Letters, vol. 117, no. 15, article no. 158101. [7] Malecki, A et al. 2014, ‘X-ray tensor tomography’, EPL (Europhysics Letters), vol. 105, no. 3, article no. 38002. [8] Liu, R, Zhang, L, Zhu, P & Wang, Z 2025, ‘Parameter constraints and ill-conditioning in tensor tomography reconstruction: a theoretical and numerical approach’, Optics Express, vol. 33, no. 9, pp. 18810–18823.
Hamiltonian Monte Carlo Ensemble Features for DATSCAN SPECT Analysis in Parkinson's Disease (#1511)
T. Leontiou1, 2, S. Jaafari2, 7, C. Likartsis3, A. Frixou4, Y. Parpottas1, 2, E. Stiliaris5, 4, C. N. Papanicolas4, 6
1 Frederick University, Department of Mechanical Engineering, School of Engineering, Nicosia, Cyprus 2 Frederick Research Center, Nicosia, Cyprus 3 Limassol General Hospital, Department of Nuclear Medicine, Limassol, Cyprus 4 The Cyprus Institute, CaSToRC, Nicosia, Cyprus 5 National and Kapodistrian University of Athens, Department of Physics, Athens, Greece 6 The Cyprus Academy of Sciences, Letters and Arts, Nicosia, Cyprus 7 European University Cyprus, Department of Histology and Embryology, School of Dentistry, Nicosia, Cyprus
Abstract
Single-photon emission computed tomography (SPECT) reconstruction is an ill-posed inverse problem, and standard clinical methods typically provide a single reconstructed image without an explicit characterization of uncertainty. In this work, we adopt a stochastic reconstruction approach based on Hamiltonian Monte Carlo (HMC), which generates an ensemble of radiotracer distributions consistent with the measured projection data and the assumed forward model. Rather than relying solely on the reconstructed mean image, this formulation enables both uncertainty-aware evaluation of conventional measures and the construction of new descriptors derived from ensemble variability. We investigate whether such ensemble-derived features can provide additional diagnostic information for DATSCAN SPECT analysis, complementing standard metrics such as striatal binding ratios (SBR). We show that normalized fluctuation measures extracted from the HMC ensemble exhibit distinct behavior between control and Parkinson’s disease cases. These findings suggest that the stochastic reconstruction framework may provide additional diagnostic information and enhance the diagnostic capabilities of SPECT.
AcknowledgmentData used in the preparation of this article were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (https://www.ppmi-info.org/data), RRID:SCR\_006431. For
up-to-date information on the study, visit https://www.ppmi-info.org.
This work was supported by the Cyprus Research and Innovation Foundation under the project PERSPECT (Project No. EXCELLENCE/0524/0410). PERSPECT project (https://perspect.frederick.ac.cy) is implemented in the frames of Cohesion Policy Programme “THALIA 2021-2027” and is co-funded by the European Union.
Keywords: Hamiltonian Monte Carlo, SPECT, DATSCAN, Parkinson's disease, stochastic reconstruction
10:30
MIC-18-07
Monte Carlo–based reconstruction in Ac-225 SPECT: impact on image quality, quantitative accuracy, and computational burden. (#1164)
C. Li1, 2, P. Esquinas1, 3, J. Fowler1, 2, X. Hou3, A. Rahmim1, 2, C. Uribe2, 3
1 The University of British Columbia, Department of Physics & Astronomy, Vancouver, British Columbia, Canada 2 BC Cancer Research Institute, Department of Basic & Translational Research, Vancouver, British Columbia, Canada 3 BC Cancer, Department of Molecular Imaging & Therapy, Vancouver, British Columbia, Canada
Abstract
Purpose: Quantitative 225Ac SPECT is challenging because of low counts, complex emissions, scatter, and septal penetration[1]. This study developed and evaluated a Monte Carlo (MC)-based reconstruction framework for 225Ac SPECT, focusing on image quality, quantitative accuracy, and computational cost. Methods: SIMIND was used as an MC forward-projector for a vendor-specific SPECT system with a high-energy collimator, while an analytical back-projector was retained to improve efficiency[2]. Unlike conventional MLEM with TEW scatter correction, the MC model incorporated patient scatter, collimator penetration/scatter, detector backscatter, and system response. Performance was evaluated using both simulated and experimental phantom data. The simulated phantom contained 20, 40, and 72 mm spheres, and the experimental phantom contained a 60 mm sphere, both with a 10:1 sphere-to-background activity ratio. Reconstructions used a 218 keV photopeak with a 20% window and were assessed using the recovery coefficient(RC) and contrast recovery coefficient(CRC). Results: In simulations, MC reconstruction improved contrast recovery and reduced scatter and septal-penetration artifacts. At 25 iterations, mean CRCs for the 72, 40, and 20 mm spheres were 90.1%, 61.3%, and 21.9%, compared with 45.4%, 29.1%, and 9.4% for conventional reconstruction. MC reconstruction directly estimated activity in MBq and achieved RCs of 90.2% and 94.5% for the 72 and 40 mm spheres. Experimental results showed consistent improvement, with MC reconstruction more than doubling CRC at 25 iterations (75.4% vs. 29.2%) and achieving an RC of 88.5% for the 60 mm sphere. Reducing photon histories from 600 to 30 million per projection preserved accuracy for the 72 mm sphere while reducing runtime from 53 to 3.1 min per iteration. Conclusion: MC-based reconstruction substantially improves image quality and quantitative accuracy for 225Ac SPECT and may support more reliable post-therapy imaging and dosimetry.
References [1] Zacherl, MJ, Gildehaus, FJ, Mittlmeier, L, Boening, G, Gosewisch, A, Wenter, V, Unterrainer, M, Schmidt-Hegemann, N, Belka, C, Kretschmer, A, et al. 2021, ‘First clinical results for PSMA-targeted α-therapy using 225Ac-PSMA-I&T in advanced-mCRPC patients’, Journal of Nuclear Medicine, 62(5), 669–674, Society of Nuclear Medicine. [2] Gustafsson, J, Brolin, G, Ljungberg, M 2018, ‘Monte Carlo-based SPECT reconstruction within the SIMIND framework’, Physics in Medicine & Biology, 63(24), 245012, IOP Publishing.
Keywords: Reconstruction, Quantitative SPECT, Monte Carlo
10:45
MIC-18-08
Sparse-View 177Lu SPECT Reconstruction for Pre-Treatment Dosimetry using Gaussian Mixtures (#2004)
H.J. Lee1, 2, S. Bae1, 2, J.S. Lee1, 2
1 Seoul National University Graduate School, Interdisciplinary Program in Bioengineering, Seoul, Republic of Korea 2 Seoul National University Graduate School, Integrated Major in Innovative Medical Science, Seoul, Republic of Korea
Abstract
Pre-treatment dosimetry for 177Lu-DOTATATE therapy requires accurate reconstruction from low-dose sparse-view SPECT imaging, which remains challenging due to severe noise and limited projection information. We propose a Gaussian Mixture Model (GMM)-based reconstruction framework with projection-domain optimization for sparse-view 177Lu SPECT imaging. Experimental results on multi-timepoint 177Lu-DOTATATE SPECT datasets demonstrated competitive reconstruction and dosimetry performance with substantially faster optimization under sparse-view conditions.
References [1] Kim, Keon Min, et al. "Pre-treatment Dosimetry using Synthetic Full-Dose SPECT generated from Low-Dose SPECT with Self-Supervised NERF." (2024): 241297-241297. [2] Kerbl, Bernhard, et al. "3d gaussian splatting for real-time radiance field rendering." ACM Trans. Graph. 42.4 (2023): 139-1.
