About

Craig Levin is a Professor of Radiology at Stanford University, affiliated with the Molecular Imaging Program at Stanford and Nuclear Medicine. He also holds courtesy appointments in Physics, Electrical Engineering, and Bioengineering. His research focuses on artificial intelligence in medicine and imaging, contributing to the development and application of AI technologies in healthcare. Levin is actively involved in the Center for Artificial Intelligence in Medicine & Imaging (AIMI), where he engages in advancing research, education, and industry collaborations to improve medical imaging and diagnostics.

Research topics

• Computer Science
• Physics
• Optics
• Artificial Intelligence
• Medical physics
• Medicine
• Telecommunications
• Nuclear medicine
• Nuclear physics
• Materials science
• Nanotechnology
• Biomedical engineering
• Engineering
• Algorithm
• Optoelectronics
• Biology
• Electronic engineering

Selected publications

• Alternating Update Method for Joint EM Estimation of Multiple Tracer Distributions in Multiplexed PET ( <i>m</i> PET)

  IEEE Transactions on Radiation and Plasma Medical Sciences · 2026-01-01

  articleSenior author
  PublisherDOI

• Generative deep learning synthesizes high signal-to-noise ratio sensitivity maps for PET from low count direct normalization data

  Physics in Medicine and Biology · 2026-01-28

  articleSenior author

  Abstract Objective. An accurate and precise normalization procedure is essential to correct for variations in detector efficiency in reconstructed positron emission tomography (PET) images. Direct normalization is a conventional approach that requires a large number of counts per line of response from a known normalization source, which is time-consuming due to the need to acquire very high statistics with a reasonable source strength that does not saturate the system. Approach. To address the challenge of acquiring high signal-to-noise ratio (SNR) PET sensitivity maps efficiently, particularly with the often relatively low-count direct normalization data, this work develops a novel PET data processing and image reconstruction pipeline. This framework integrates sensitivity map features with generative modeling to synthesize high-quality maps, significantly reducing acquisition time while ensuring accurate and efficient normalization. Key contributions comprise a conditional attention-guided generative adversarial network that preserves the geometric and detector-specific characteristics of sensitivity maps, a robust assessment framework to verify synthesized map plausibility, and a comprehensive evaluation of the model’s performance across a range of acquisition and scanner conditions. Main Results. Quantitative evaluations were performed by testing the model on totally unseen normalization data, acquired to reconstruct images of a Hoffman brain phantom, a contrast phantom, and a uniform cylinder phantom. This evaluation used high-count, low-count (1%–15% of high count scan), and synthetic high-count sensitivity maps. The Hoffman brain image volume normalized using a synthetic sensitivity map with 15% count statistics as input produced results that closely matched that using the high count normalization data, with peak SNR (PSNR), structural similarity index measure (SSIM), and normalized root mean square error (NRMSE) values (mean ± standard error) of 30.68 ± 0.31, 0.95 ± 0.00, and 0.35 ± 0.00, respectively. In comparison, the unprocessed sensitivity map with 15% count statistics yielded substantially worse PSNR, SSIM, and NRMSE values of 15.93 ± 0.43, 0.54 ± 0.01, and 1.84 ± 0.03, respectively. Significance. This novel, fast, and effective approach enables high SNR direct normalization of PET image volumes through deep learning using synthetic correction factors obtained from a short normalization scan.

  PublisherDOI

• Triplexed PET: Experimental Results of Imaging Three Isotopes Simultaneously in a Single PET Imaging Session

  2025-11-01

  articleSenior author

  Positron emission tomography (PET) employs radiotracers-molecules labeled with positronemitting isotopes-to visualize biological processes via detection of 511 keV annihilation photon pairs. Conventional PET is typically limited to imaging a single radiotracer at a time due to its exclusive reliance on these annihilation photons. However, some positron-emitting isotopes also emit a prompt gamma photon with energy <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$&gt;511 \text{keV}$</tex>. By distinguishing between two- and threephoton coincidences, these isotopes enable the differentiation of multiple radiotracers within a single scan-a technique referred to as multiplexed PET (mPET). In this study, we demonstrate triplexed PET imaging of three distinct radioisotopes–<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">64</sup>Cu, <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">124</sup>I, and <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">55</sup>Co–using a custom PET system capable of detecting three-photon coincidences within an extended energy window. This approach will enable simultaneous visualization of three complementary biomarkers, offering a more comprehensive molecular view of diseases such as cancer, cardiovascular disorders, and neurological conditions. We present experimental results demonstrating the feasibility of this approach. Qualitative images confirm the expected spatial distribution of isotopes, with signal contributions aligning with the known energy windows of <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">64</sup>Cu, <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">124</sup>I, and <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">55</sup>Co. However, quantitative analysis using a linear unmixing method reveals limitations, such as cross-talk between single-isotope images and an average absolute error of 13 % in activity estimation. These findings underscore both the promise of simultaneous imaging of three isotopes and the need for improved correction strategies and nonlinear unmixing techniques to enhance quantitative accuracy in mPET.

  PublisherDOI

• A magnetic field compatible readout circuit for enhanced coincidence time resolution in BGO Cherenkov radiation–based TOF‐PET detectors

  Medical Physics · 2025-01-29 · 1 citations

  articleSenior authorCorresponding

  Abstract Background Developing time‐of‐flight positron emission tomography/magnetic resonance imaging (TOF‐PET/MRI) detectors that exploit prompt Cherenkov photons from bismuth germanate (BGO) crystals for estimating 511 keV photon arrival time. Purpose To present a low‐noise, high‐speed electronic readout circuit design for BGO‐based TOF‐PET detectors that achieves enhanced coincidence time resolution (CTR) in presence of a strong magnetic field. Methods The CTR of a BGO‐based TOF‐PET test detector employing a high‐speed, low‐noise electronic readout chain was evaluated in a strong magnetic field produced by a permanent magnet placed directly on top of the circuit. For these experiments, which exploit Cherenkov radiation for precise measurement of annihilation photon time arrival time difference, a point source of 22 Na was positioned between a pair of 3 × 3 × 15 mm 3 polished BGO crystals wrapped in Teflon tape and optically coupled to 3 × 3 mm 2 ultra‐violet (UV)‐sensitive silicon photomultipliers (SiPMs). Results By incorporating both Cherenkov (prompt) and standard (slow) luminescence components, 283 ± 8 ps and 275 ± 10 ps full‐width‐half‐maximum (FWHM) CTR were achieved without and with the permanent magnet present, respectfully. These values improved to 236 ± 4 ps and 216 ± 17 ps FWHM when only the Cherenkov components of the timing signal (events with the fastest rise time) were considered. Conclusions Results indicate we have designed a high‐performance readout circuit that achieves significantly the same CTR in BGO with or without a strong magnetic field present. This further demonstrates that UV SiPMs can effectively operate in a strong magnetic field while remaining highly advantageous for detecting Cherenkov radiation, thus highlighting their potential to be used in BGO‐based TOF‐PET/MRI scanners.

  PublisherDOI

• Regularization in Multiplexed Pet Reconstruction

  2025-11-01

  articleSenior author

  In this paper we present a novel application of regularized EM reconstruction to enhance the unmixed individual tracer images in multiplexed PET (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$m$</tex> PET). Multiplexed PET using one tracer labeled with a positron plus prompt gamma ray emitter is limited by low sensitivity to triplephoton coincidences, leading to high noise in both reconstructed images. The use of regularization has been previously shown to increase the image quality of single tracer PET images with low signal-to-noise. First, we present how block sequence regularization expectation maximization (BSREM) can be used to regularize simultaneously acquired mPET data. Next, we use Monte Carlo to simulate an IQ phantom with a pure positron (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${ }^{18} ~\mathrm{F}$</tex>) and a positron plus prompt gamma (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${ }^{44} \text{Sc}$</tex>) emitter. For a reconstruction method that simultaneously estimates the two tracer distributions while performing regularization, we show that a relative difference penalty (RDP) provides an optimal tradeoff in image quality and quantitative accuracy over quadratic and log-cosh penalty functions, with a CNR in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${ }^{18} ~\mathrm{F} /{ }^{44} \text{Sc}$</tex> of 13.3/5.9 and 6.8/3.4 for the RDP and quadratic penalties in the smallest spheres, respectively. However, this simultaneous reconstruction with regularization has a worse crosstalk ratio in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${ }^{18} ~\mathrm{F} /{ }^{44} \text{Sc}$</tex> of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.83 / 0.83$</tex> as opposed to the unregularized reconstruction which has <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.64 / 0.47$</tex>. Finally, we further optimize the reconstruction by alternating the update of each tracer with regularization to achieve a CNR in the smallest <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${ }^{18} ~\mathrm{F} /{ }^{44} \text{Sc}$</tex> spheres of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$15.7 / 6.2$</tex>, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 5-15 \%$</tex> improvement over simultaneously regularized images and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 25-40 \%$</tex> improvement over no regularization. This method also reduces the crosstalk ratio for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${ }^{18} ~\mathrm{F} /{ }^{44} \text{Sc}$</tex> in the smallest spheres to 0.56/0.37, better than the unregularized reconstruction.

  PublisherDOI

• Modular Readout Electronics System for a 3D Position Sensitive 100 ps Coincidence Time Resolution TOF-PET Detector Configuration

  2025-11-01

  articleSenior author

  This study presents a modular PET readout architecture that aims to achieve near-100 ps coincidence time resolution (CTR) within a system-level time-of-flight PET configuration, while also enabling 3D positioning of one or more interactions per event. Building on our previous sidereadout detector design and CPLD-based front-end electronics, we developed a new stackable and buttable front-end module and a customized back-end FPGA board that together implement a cross-chip single-event processing framework. Each front-end module performs local digitization of energy and position signals, while the back-end FPGA handles time-to-digital conversion (TDC) with <10 ps rms jitter, global synchronization, and data aggregation. To evaluate coincidence timing performance under realistic multi-module conditions, we conducted a coincidence test using a 4 × 2 array of 3 × 3 × 10 mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> LYSO:Ce crystals side-coupled to a 4 × 6 array of 3.16 × 3.16 mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> SiPMs. In the new configuration, the two detectors were read out by separate back-end FPGA boards via a 200 MHz global clock delivered over 2-meter-long SMA cables through a treestructured distribution and synchronization network to test cross-module clock distribution and synchronization. The intrinsic CTR between two identical 24-SiPM detector units was determined to be 118.07 ± 2.19 ps FWHM, closely matching the 121.28 ± 3.35 ps result from a previous singleFPGA setup. These results validate the proposed architecture, clock distribution, and synchronization strategy for scalable, multi-module TOF-PET 3D position sensitive scintillation detector systems.

  PublisherDOI

• Front-End Electronics Design for 3-D Position Sensitive TOF-PET Detector That Achieves ~120-ps CTR and ~1.2-mm DOI Resolution

  IEEE Transactions on Radiation and Plasma Medical Sciences · 2025-02-14

  articleSenior author

  This study introduces and evaluates a new front-end electronics design for time-of-flight (TOF) 3-D position sensitive (TOF-3-DPS) detectors with a side-readout configuration. This design employs an RF amplifier and summing circuit-based timing multiplexing scheme to achieve 24:1 timing multiplexing. Additionally, complex programmable logic devices are utilized for precise energy measurement and 3-D positioning, accommodating both single and multiinteraction intercrystal scatter (ICS) events within a detector unit. Experimental results on a single <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$3\times 3\times $ </tex-math></inline-formula> 10 mm3 LYSO:Ce crystal side-coupled to three <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$3\times $ </tex-math></inline-formula> 3 mm2 SiPMs in a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4\times 6$ </tex-math></inline-formula> SiPM array demonstrated a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$9.17\pm 0.20$ </tex-math></inline-formula>% energy resolution, a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.20\pm 0$ </tex-math></inline-formula>.26 mm FWHM depth-of-interaction (DOI) resolution, and a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$112.46\pm 1.91$ </tex-math></inline-formula> ps FWHM coincidence time resolution (CTR) after DOI-related time skew correction. Further tests on a detector unit comprising a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4\times 2$ </tex-math></inline-formula> array of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$3\times 3\times $ </tex-math></inline-formula> 10 mm3 LYSO:Ce crystals, side-coupled with the same <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4\times 6$ </tex-math></inline-formula> SiPM array, yielded a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$10.56\pm 1.05$ </tex-math></inline-formula>% energy resolution and a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$121.28\pm 3.35$ </tex-math></inline-formula> ps FWHM DOI-calibrated CTR. The ICS event ratio for each crystal element within the detector unit was also preliminarily assessed. The front-end readout circuit consumes approximately 0.75 W per 24-SiPMs detector unit and features a compact <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$27\times $ </tex-math></inline-formula> 95 mm2 footprint capable of reading out two units, enabling easy stacking of multiple units to form a complete TOF-3-DPS detector module.

  PublisherDOI

• Image SNR Enhancement for a Short Axial FOV Brain PET System Using Generative Deep Learning

  IEEE Transactions on Radiation and Plasma Medical Sciences · 2025-04-21 · 1 citations

  articleSenior author

  The signal-to-noise ratio (SNR) of positron emission tomography (PET) images is determined by several factors including the geometry of the scanner. Low system sensitivity caused by a short axial field of view (FOV) results in a low reconstructed image SNR that can complicate clinical decision-making. Therefore, a longer FOV is highly desirable (e.g., a total body geometry). However, this raises the scanner’s cost by increasing the volume of crystals, number of detectors, and readout electronics. We have developed a deep-learning framework to enhance the image quality of data acquired from a prototype brain-dedicated PET insert system for PET/MRI with an axial FOV of just 2.8 cm. We employed a retrospective analysis on 18F-fluorodeoxyglucose PET scans of 28 patients with either Glioblastoma (n=9) or Alzheimer’s disease (n=19) acquired on a commercial PET/MRI scanner with 60 cm diameter and 25 cm axial FOV. From this data we reconstructed low statistics PET images mimicking that acquired from the 2.8 cm axial FOV brain PET prototype using the 25 cm axial FOV commercial system dataset using a "fault-tolerant reconstruction" algorithm, which allowed us to constrain the count statistics from a set of detectors in a single ring of the latter system to match the geometry of the former system. A conditional generative adversarial network (cGAN) was trained and tested using the simulated short axial FOV images as input, with the paired 25 cm axial FOV image data as the target. We performed 5-fold cross-validation and compared the deep learning (DL)-enhanced images to the target images using 4 metrics: peak-signal-to-noise-ratio (PSNR), root mean squared error (RMSE), mean absolute error (MAE), and structural similarity index (SSIM). The DL-enhanced PET images from the 2.8 cm axial FOV system had a median PSNR of 39.09 (interquartile range (IQR): 32.80–45.32), a median SSIM of 0.98 (IQR: 0.97–0.99), a median RMSE of 0.07 (IQR: 0.04–0.09), and a median MAE of 0.004 (IQR: 0.000–0.009). We also assessed the pretrained cGAN model’s performance in a zero-shot denoising task using patient data collected with our first generation PETcoil system. The ability of the cGAN model to enhance the quality of PET images acquired with a short axial FOV suggests a potential method to provide high-quality, high-accuracy images comparable to those of large axial FOV systems.

  PublisherDOI

• Characterization of a Novel Detector Design for an Ultra-High-Resolution, MR-Compatible, Breast-Dedicated Multiplexed Pet (mPET) System

  2025-11-01

  articleSenior author

  We are developing a high-resolution (~ 1 mm) MR-compatible, breast-dedicated PET insert that enables imaging of two tracers in one PET imaging session(a.k.a multiplexed PET or mPET). To achieve this goal, we propose a single-ended, DOI detector design with high resolution and detection efficiency, featuring a trapezoidal shape to maximize system sensitivity. We propose using regular crystal rod elements of varying lengths to create a trapezoidal crystal array, simplifying fabrication and enhancing reliability. The detector consists of LYSO crystal elements ranging from 5 mm to 35 mm in length, with a 1.28 mm pitch, read out by an MPPC array and PETsys ASIC. A light-sharing approach, combined with unpolished crystal surfaces and top-side light redirection, enables DOI encoding. A prototype detector array module was evaluated for crystal separation, energy resolution, and DOI resolution. Measurements were performed with both side-irradiation and front-irradiation setups to assess the detector's performance across various interaction depths in the different crystal elements. The prototype detector successfully resolved crystal elements even with a 35 mm length. DOI resolutions ranged from 4 to 5 mm for crystal lengths between 10 mm and 30 mm, and 7 to 8 mm for 35 mm length. Energy resolutions varied from 13 % to 26 %, with degradation observed at positions farther from the MPPC plane. The DTR is 331 ps for 5 mm long crystals and between 400-660 ps for 10 mm to 35 mm long crystals. The calibrated DOI results from the front irradiation setup were consistent with those from the side-irradiation setup. In summary, this work introduces a novel, single-ended, DOI-capable detector design that optimizes sensitivity while achieving ultra-high spatial resolution for a breast-dedicated mPET system through a unique trapezoidal crystal array and light-sharing DOI encoding.

  PublisherDOI

• Evaluation of Event Positioning Strategies to Optimize Contrast Recovery in a TOFPET Brain Insert for Simultaneous PET/MRI

  2025-11-01

  articleSenior author

  Simultaneous PET/MRI offers important advantages for neurological imaging, demanding high spatial resolution and contrast recovery. This study evaluates multiple event positioning strategies within the scintillation crystals of a Time-of-Flight (TOF) PET brain insert (PETcoil) designed for PET/MRI imaging. Using GATE Monte Carlo simulations and a custom contrast recovery phantom with embedded hot spheres, we test four positioning approaches-front face, back face, center, and average depth of interaction (DOI)-for their impact on contrast recovery. Simulation data was reconstructed using OSEM and analyzed across two axial positions. Results demonstrate that event positioning significantly influences percent contrast, with center and back face strategies yielding the highest performance <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(84.6 \pm 3.2 \%$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$84.3 \pm 3.6 \%$</tex>, respectively), while the front face strategy resulted in the lowest contrast (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$68.8 \pm 2.5 \%$</tex>). These findings inform future optimization of the PETcoil system and other high-performance brain PET systems.

  PublisherDOI

Recent grants

• NIH Grant R21EB003283

  NIH · $287k · 2005

• A new direction to achieve ultra-fast timing for positron emission tomography

  NIH · $636k · 2017–2021

• NIH Grant R01EB02512501

  NIH · $416k · 2019

• Stanford Molecular Imaging Scholars (SMIS) Program

  NIH · $436k · 2016–2022

• NIH Grant R01EB025125

  NIH · $568k · 2021

Frequent coauthors

• Peter D. Olcott

  RefleXion Medical (United States)

  95 shared

• Garry Chinn

  Stanford University

  92 shared

• Joshua W. Cates

  Lawrence Berkeley National Laboratory

  76 shared

• Derek Innes

  Stanford University

  62 shared

• Chen‐Ming Chang

  National Chung Cheng University

  48 shared

• Matthew F Bieniosek

  Enable Biosciences (United States)

  48 shared

• A.M.K. Foudray

  Stanford Medicine

  46 shared

• Li Tao

  Hainan Medical University

  35 shared