Novel imaging techniques for interventional applications
Non-circular orbit design for interventional cone-beam CT
We develop novel acquisition and reconstruction techniques for interventional cone-beam CT, including non-circular orbits for large field-of-view imaging and metal artifact reduction.
Selected presentations
- Non-circular CT orbit design for elimination of metal artifacts - SPIE Medical Imaging 2020
Spectral digital subtraction angiography
Digital subtraction angiography is sensitive to motion between mask and contrast frames. We develop spectral-temporal processing methods that use energy-resolved measurements to improve vessel visibility, material separation, and robustness to motion in interventional angiography.
Selected presentations
- Three-material decomposition using a dual-layer flat-panel detector in the presence of soft tissue motion - SPIE Medical Imaging 2023
Spectral imaging
Detector modeling and system design
We model spectral imaging systems to improve the performance of quantitative spectral imaging tasks. Predictive models for bias, variance, and task-based performance indices are developed to improve spectral detector and system designs.
- Detector modeling for photon-counting and multi-layer detectors
- Material decomposition algorithms and image-quality prediction
Selected presentations
- Theoretical prediction of bias in model-based material decomposition - SPIE Medical Imaging 2026
Deep learning image formation
Assessment of deep learning reconstruction
Deep learning reconstruction and post-processing methods are nonlinear and data dependent. We develop analysis methods to evaluate how learned algorithms change image content and diagnostic performance, including how uncertainty and variability can be communicated to radiologists.
Selected presentations
- Communicating variability in deep learning CT reconstruction to radiologists - Deep Reconstruction Workshop 2025
Image synthesis with radiomics supervision
We develop domain-guided generative models for image synthesis. We developed a generative adversarial network with radiomics supervision to encourage synthesized and training samples to have similar radiomic feature distributions. We later extended this concept to radiomics-conditioned diffusion posterior sampling, treating radiomic features as consistency constraints to control the synthesized lesions.
- Radiomics-conditioned image synthesis
- Domain knowledge incorporated into generative imaging models
Selected presentations
- A generative model of lung CT conditioned on radiomics features - SPIE Medical Imaging 2026
- Generative adversarial networks and radiomics supervision for lung lesion synthesis - SPIE Medical Imaging 2021
Physics-guided radiomics standardization
Radiomics feature standardization
Scanner-induced variability is a major impediment to the clinical translation of radiomics studies. We use prior knowledge of spatial resolution and noise to predict and mitigate scanner-dependent feature variability for robust, reproducible, and generalizable downstream quantitative analysis.
- Standardization of histogram- and texture-based radiomics features
- Prediction and correction of scanner-dependent feature shifts from blur and noise
- Evaluation of feature recovery in classification tasks
Selected presentations
- End-to-end modeling for predicting and estimating radiomics: application to GLCM in CT - SPIE Medical Imaging 2021
Image quality and task-driven imaging
Image chain modeling and system design
We model the imaging chain from data acquisition through reconstruction and image interpretation. The goal is to use task-based detectability indices to optimize system design and improve imaging performance.
Selected presentations
- Joint optimization of fluence field modulation and regularization for multi-task objectives - SPIE Medical Imaging 2018
- Task-driven acquisition and reconstruction in cone-beam CT - SPIE Medical Imaging 2015
Image analysis
Automatic working-angle prediction for neuro-interventional procedures
We develop image-analysis and quantitative feature-extraction methods for procedural planning and outcome evaluation. In neuro-interventional applications, this includes automatic working-angle prediction for intracranial aneurysm treatment using anatomical models derived from 3D angiographic imaging.
- Segmentation and quantitative feature extraction from 3D angiography
- Automatic neck-view and barrel-view prediction for aneurysm treatment
Selected presentations
- Automatic viewing angles prediction for endovascular treatment in intracranial aneurysm - SPIE Medical Imaging 2026