About

Grace Jianan Gang, Ph.D., is an Associate Professor of Radiology at the Perelman School of Medicine at the University of Pennsylvania. Her research expertise includes computed tomography, cone-beam computed tomography, image quality, observer models, optimization, image reconstruction, spectral CT, dual energy CT, tomosynthesis, x-ray imaging, radiomics, and deep learning. She has contributed to the development and evaluation of advanced imaging techniques and algorithms, including the use of generative adversarial networks with radiomics supervision for lung lesion generation, multi-material decomposition using spectral diffusion posterior sampling, and strategies for CT reconstruction with diffusion posterior sampling. Dr. Gang's work also involves creating patient-derived phantoms for evaluating clinical imaging performance and dose reduction capabilities of deep learning CT reconstruction algorithms, as well as improving CT detectability assessment through 3D printed phantoms. Her research aims to enhance imaging quality, optimize reconstruction methods, and advance the application of deep learning in medical imaging.

Research topics

• Computer Science
• Artificial Intelligence
• Mathematics
• Medicine
• Nuclear medicine
• Medical physics
• Radiology
• Statistics

Selected publications

• Automatic viewing angles prediction for endovascular treatment in intracranial aneurysm

  2026-04-01

  articleSenior author

  Timely and accurate selection of working angles for 2D digital subtraction angiography (DSA) on biplane C-arms system is critical for successful endovascular treatments of intracranial saccular aneurysms. Image guidance under fluoroscopy acquired at these working angles ensure clear visualization of coil positioning and allow the detection of coil protrusion into the parent vessel. Currently, these angles are manually chosen, which can be time-consuming and subject to variability, particularly among less experienced clinicians. In this work, we propose an automatic method to predict two working angles using deep learning and rule-based geometric criteria. Aneurysms are first segmented from parent vessels using a point transformer-based model. The two working angles constituting the “neck” and “barrel” views are then predicted based on geometric criteria. For the neck view, a weighted voting strategy considers maximizing neck diameter, minimizing aneurysm-vessel overlap, minimizing vessel overlap, and maximizing neck shoulder clearance. For the barrel view, a similar strategy considers minimizing aneurysm-vessel overlap and optimizing parent vessel alignment. Finally, joint optimization ensures that all vessel branches are clearly visible in at least one view. The algorithm was evaluated on 84 aneurysm models from an open source dataset. The predicted angles showed strong agreement with ground truth annotated by an experienced neurosurgeon, with ≤ 20° differences in 75% of the neck views and 90% of the barrel views. The method has the potential for clinical integration to improve the efficiency and consistency in aneurysm treatment planning.

  PublisherDOI

• Deep learning CT image restoration using system blur and noise models

  Journal of Medical Imaging · 2025-02-03 · 2 citations

  articleOpen access

  Purpose: The restoration of images affected by blur and noise has been widely studied and has broad potential for applications including in medical imaging modalities such as computed tomography. Recently, deep learning approaches have demonstrated the potential to enhance image quality beyond classic limits; however, most deep learning models attempt a blind restoration problem and base their restoration on image inputs alone without direct knowledge of the image noise and blur properties. We present a method that leverages both degraded image inputs and a characterization of the system's blur and noise to combine modeling and deep learning approaches. Approach: Different methods to integrate these auxiliary inputs are presented, namely, an input-variant and a weight-variant approach wherein the auxiliary inputs are incorporated as a parameter vector before and after the convolutional block, respectively, allowing easy integration into any convolutional neural network architecture. Results: The proposed model shows superior performance compared with baseline models lacking auxiliary inputs. Evaluations are based on the average peak signal-to-noise ratio and structural similarity index measure, selected examples of top and bottom 10% performance for varying approaches, and an input space analysis to assess the effect of different noise and blur on performance. In addition, the proposed model exhibits a degree of robustness when the blur and noise parameters deviate from their true values. Conclusion: Results demonstrate the efficacy of providing a deep learning model with auxiliary inputs, representing system blur and noise characteristics, to enhance the performance of the model in image restoration tasks.

  PublisherOA PDFDOI

• Multi-Material Decomposition Using Spectral Diffusion Posterior Sampling

  IEEE Transactions on Biomedical Engineering · 2025-02-19

  articleOpen access

  Many spectral CT applications require accurate material decomposition. Existing material decomposition algorithms are often susceptible to significant noise magnification or, in the case of one-step model-based approaches, hampered by slow convergence rates and large computational requirements. In this work, we proposed a novel framework - spectral diffusion posterior sampling (spectral DPS) - for one-step reconstruction and multi-material decomposition, which combines sophisticated prior information captured by one-time unsupervised learning and an arbitrary analytic physical system model. Spectral DPS is built upon a general DPS framework for nonlinear inverse problems. Several strategies developed in previous work, including jumpstart sampling, Jacobian approximation, and multi-step likelihood updates are applied facilitate stable and accurate decompositions. The effectiveness of spectral DPS was evaluated on a simulated dual-layer and a kV-switching spectral system as well as on a physical cone-beam CT (CBCT) test bench. In simulation studies, spectral DPS improved PSNR by 27.49% to 71.93% over baseline DPS and by 26.53% to 57.30% over MBMD, depending on the the region of interest. In physical phantom study, spectral DPS achieved a <1% error in estimating the mean density in a homogeneous region. Compared with baseline DPS, spectral DPS effectively avoided generating false structures in the homogeneous phantom and reduced the variability around edges. Both simulation and physical phantom studies demonstrated the superior performance of spectral DPS for stable and accurate material decomposition.

  PublisherOA PDFDOI

• Recovery of GLRLM features in degraded images using deep learning and image property models

  2025-04-11

  articleOpen accessSenior author

  Radiomics models have been extensively used to predict clinical outcomes across various applications. However, their generalizability is often limited by undesirable feature values variability due to diverse imaging conditions. To address this issue, we previously developed a dual-domain deep learning approach to recover ground truth feature values in the presence of known blur and noise. The model consists of a differentiable approximation for radiomics calculation and a dual-domain loss function. We demonstrated model performance for gray-level co-occurrence matrix (GLCM) and histogram-based features. In this work, we extend the method to gray-level run length matrix (GLRLM) feature recovery. We introduce a novel algorithm for the differentiable approximation of GLRLMs. We assessed the performance of the GLRLM feature restoration network using lung CT image patches, with a focus on the accuracy of recovered feature values and classification performance between normal and COVID-positive lungs. The proposed network outperformed the baselines, achieving the lowest MSE in GLRLM feature recovery. Furthermore, a classification model based on the recovered GLRLM features reached an accuracy of 86.65%, closely aligning with the 88.85% accuracy of models using ground truth features, compared to 82.00% accuracy from degraded features. These results demonstrate the potential of our method as a robust tool for radiomics standardization.

  PublisherOA PDFDOI

• Joint Temporal and Spectral Processing for Improved Digital Subtraction Angiography Using Photon-Counting Detectors

  IEEE Transactions on Biomedical Engineering · 2025-05-16

  articleOpen accessSenior author

  OBJECTIVE: Digital subtraction angiography (DSA) is the gold standard modality for diagnostics and guidance for interventional procedures. Spectral imaging has previously been explored for DSA, but severe noise amplification from material decomposition has impeded clinical adoption. We present a novel joint processing strategy that leverages both temporal and spectral information for material decomposition to address this issue. METHODS: We develop a model-based material decomposition approach that utilizes the pre- and post-contrast images simultaneously for material estimation. Performance was evaluated on a small-vessel phantom on a test bench with a photon-counting detector. Joint processing was compared with temporal subtraction and previously proposed spectral DSA techniques including hybrid subtraction and conventional three-material decomposition. Additional simulation was performed to investigate performance with perfectly calibrated spectral response and sensitivity to patient motion. RESULTS: The improved conditioning of the proposed method effectively reduces bias and noise in the spectral results and allows three-material decomposition with dual-energy spectral measurements. The method achieved more than an order of magnitude variance reduction compared to previously proposed spectral DSA techniques. Compared to temporal subtraction, a mean variance reduction of 23.9% was achieved in simulation and 10.8% in experimental data. The degree of reduction is object-dependent. Noise reduction achieved in physical experiments is slightly lower than that in simulation, likely due to bias from imperfect spectral calibration. The method is equally sensitive to motion compared to temporal subtraction. CONCLUSION: The proposed method addresses a major image quality challenge limiting previous approaches and outperforms temporal subtraction. SIGNIFICANCE: Such improvements facilitate the clinical translation of spectral angiography.

  PublisherOA PDFDOI

• Volumetric Material Decomposition Using Spectral Diffusion Posterior Sampling with a Compressed Polychromatic Forward Model.

  PubMed · 2025-03-28

  preprintOpen access

  We have previously introduced Spectral Diffusion Posterior Sampling (Spectral DPS) as a framework for accurate one-step material decomposition by integrating analytic spectral system models with priors learned from large datasets. This work extends the 2D Spectral DPS algorithm to 3D by addressing potentially limiting large-memory requirements with a pre-trained 2D diffusion model for slice-by-slice processing and a compressed polychromatic forward model to ensure accurate physical modeling. Simulation studies demonstrate that the proposed memory-efficient 3D Spectral DPS enables material decomposition of clinically significant volume sizes. Quantitative analysis reveals that Spectral DPS outperforms other deep-learning algorithms, such as InceptNet and conditional DDPM in contrast quantification, inter-slice continuity, and resolution preservation. This study establishes a foundation for advancing one-step material decomposition in volumetric spectral CT.

  PublisherOA PDF

• Simultaneous Material Decomposition and Detector Calibration in Photon-Counting CT

  2025-11-01

  article

  Photon-counting CT (PCCT) enables advanced spectral x-ray imaging by resolving the energy of individual photons. However, physical limitations such as charge sharing, pulse pileup, and detector instability introduce spatial, spectral, and temporal variations in detector response. These effects complicate detector calibration and degrade the accuracy of subsequent material quantification. In this work, we propose a unified framework–Joint Diffusion Posterior Sampling (Joint DPS)–for simultaneous one-step material decomposition and online detector calibration. Unlike conventional decomposition approaches that rely on a pre-calibrated spectral response and stability of the detector between calibration and scan time, Joint DPS seeks to estimate unknown pixel-wise spectral distortions using a low-dimensional parameterized spectral model. A Denoising Diffusion Probabilistic Model (DDPM) is trained to provide a sophisticated learned prior for material images. The overall approach formulates a joint optimization problem to estimate both material density maps and pixel-wise spectral parameters from projection data. Simulation studies demonstrate that Joint DPS effectively compensates for a mismatched detector response, substantially improving decomposition accuracy and reducing image artifacts without requiring separate calibration scans, thereby enhancing the robustness of PCCT imaging.

  PublisherDOI

• Joint CT reconstruction of anatomy and implants using a mixed prior model

  Journal of Medical Imaging · 2025-10-18

  article

  Purpose: Medical implants, often made of dense materials, pose significant challenges to accurate computed tomography (CT) reconstruction, especially near implants due to beam hardening and partial-volume artifacts. Moreover, diagnostics involving implants often require separate visualization for implants and anatomy. In this work, we propose a approach for joint estimation of anatomy and implants as separate volumes using a mixed prior model. Approach: We leverage a learning-based prior for anatomy and a sparsity prior for implants to decouple the two volumes. In addition, a hybrid mono-polyenergetic forward model is employed to accommodate the spectral effects of implants, and a multiresolution object model is used to achieve high-resolution implant reconstruction. The reconstruction process alternates between diffusion posterior sampling for anatomy updates and classic optimization for implants and spectral coefficients. Results: Evaluations were performed on emulated cardiac imaging with stent and spine imaging with pedicle screws. The structures of the cardiac stent with 0.25 mm wires were clearly visualized in the implant images, whereas the blooming artifacts around the stent were effectively suppressed in the anatomical reconstruction. For pedicle screws, the proposed algorithm mitigated streaking and beam-hardening artifacts in the anatomy volume, demonstrating significant improvements in SSIM and PSNR compared with frequency-splitting metal artifact reduction and model-based reconstruction on slices containing implants. Conclusion: The proposed mixed prior model coupled with a hybrid spectral and multiresolution model can help to separate spatially and spectrally distinct objects that differ from anatomical features in single-energy CT, improving both image quality and separate visualization of implants and anatomy.

  PublisherDOI

• A Photon-Counting CT Simulator with Charge Sharing, Pulse Pileup, and Nonuniform Response

  2025-11-01

  article

  Photon-counting computed tomography (PCCT) has entered clinical use and is a valuable tool for high-resolution and energy-resolved x-ray imaging. However, physical effects such as charge sharing (CS), pulse pileup (PP), and detector nonuniformity limit imaging accuracy. Accurate and efficient modeling of these effects is essential for both imaging algorithm validation and system design optimization. In this work, we present a fast and flexible PCCT simulation framework that comprehensively incorporates CS, PP, and detector nonuniformity. The framework integrates polychromatic forward projection, physics-based CS simulation with spatial-energetic correlation, analytical PP simulation based on a probabilistic model, and nonuniform detector response simulation. The simulator is implemented using the PyTorch framework with GPU acceleration, enabling efficient parallelized computation. Validation against a Monte Carlo simulation demonstrates good agreement in terms of count mean, variance, and inter-channel correlations. Simulated patient CT scans with 720 projections on a 1536x5 pixel detector further demonstrate the capability for realistic, patient-scale simulations with configurable conditions and reasonable runtimes (<2 seconds for CS, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 60$</tex> seconds for CS+PP with modest computational resources); in contrast to full Monte Carlo simulation which is typically computationally infeasible for patientscale simulation.

  PublisherDOI

• Joint reconstruction and scatter estimation in cone-beam CT using diffusion posterior sampling

  2025-04-08

  articleOpen access

  X-ray scatter degrades image quality in computed tomography, particularly in cone-beam CT due to wide cone angles. While mono-energetic CT scatter correction is well-studied, spectral CT imaging presents additional challenges due to its sensitivity to unmodeled biases in material decomposition and density estimation. This work presents a joint estimation approach that simultaneously estimates scatter and material densities by integrating the scatter component into a spectral CT forward model. Using Diffusion Posterior Sampling method, we leverage the combination of prior knowledge from large dataset training and the physical model for joint density and scatter estimation. Tested on simulated and phantom data, our method significantly reduce artifacts associated with unestimated scatter, improving spectral CT image quality.

  PublisherOA PDFDOI

Frequent coauthors

• J. Webster Stayman

  124 shared

• Peter B. Noël

  MIT World Peace University

  44 shared

• Jeffrey H. Siewerdsen

  Johns Hopkins University

  41 shared

• Wojciech Zbijewski

  Johns Hopkins University

  36 shared

• Matthew Tivnan

  Massachusetts General Hospital

  25 shared

• J. H. Siewerdsen

  Johns Hopkins University

  23 shared

• Nadav Shapira

  University of Pennsylvania

  21 shared

• Kai Mei

  University of Pennsylvania

  16 shared

Education

• PhD, Institute of Biomaterials and Biomedical Engineering

  University of Toronto

  2014