Research

Real-time Motion Tracking and Motion Management

Challenges with treating abdominal tumors on X-ray imaging-equipped MV Linacs

The workflow for online plan and motion model adaptation, and image-based real-time motion tracking and radiation gating.

Enhance Locally Advanced Pancreatic Cancer (LAPC) treatments using Online Plan Adaptation and Real-time Image-Based Target Motion Tracking on X-ray Linacs

Real-time motion tracking of abdominal tumors, including LAPC, on regular X-ray Linac machines is an unsolved problem in radiation therapy. We are developing novel imaging methods and AI models to continuously and accurately monitor and track the 3-dimensional respiratory motion of the tumor targets and organs-at-risk on the radiation treatment day. The novel solutions will allow for precisely directing the radiation beams to the moving tumor, reducing gastrointestinal radiation toxicity, and increasing the treatment effectiveness and overall survival.

Abdomen Real-time Motion Tracking movie.mp4

Real-time 3D motion tracking by feeding the AI prediction model with rotational X-ray projection fluoroscopy video.

Utilize AI to Correct 4D-CBCT and Eliminate Streaking Artifacts. (Up) Original 4D-CBCT. (Down) After Improvement

The RSTAR-4DNet Architecture

4D Cone Beam CT

A 4D convolutional neural network model was developed to reduce the streak artifacts induced by sparse and clustered phase binning. The 4D gated-FDK images served as the input, and the outputs were the artifact-removed images. A lightweight separable 4D convolution module was designed to process all phase-gated 4D CBCT images as a whole 4D volume and fully utilize the correlated image information among all 4D phases.

Demonstration with An Abdominal Case

Stereotactic Arrhythmia Radioablation and Cardiac Motion Management

Cardiac Motion Assessment and Modeling

It is critical to quantitatively assess the cardiac motion and respiratory motion of the heart to determine the motion margin for treatment planning and optimization. We developed novel image processing methods to compute heart motion in the breath-hold cardiac 4DCTs and the free-breathing respiratory 4DCTs.

The workflow of our gCGF algorithm (Group-wise Coherent Point Drifting with Gaussian Mixture Models and Finite Element Modeling) to compute the respiratory motion of the heart in the respiratory 4DCTs

Analysis of the cardiac motion of the heart

Analysis of the respiratory motion of the heart

ICD Lead Metal Artifact Examples

The AI Inpatient Model Workflow

McKeown, Trevor, et al. "ICD lead and primary metal artifact detection and inpainting in cardiac CT images." Medical physics 52.7 (2025): e17947.

Implanted Cardiac Device Metal Artifact Reduction

We develop a methodology to automatically detect the ICD lead wires and surrounding primary metal artifacts in cardiac CT scans and inpaint the affected volume with anatomically consistent structures and values.

The inpainting results on patients’ 4DCT images, showing realistic inpainting of artifacts, and comparing to other established methods.

5DCT Imaging

It is critical to precisely define the critical cardiac substructures regardless of the motion in the CT images, and to accurately assess both respiratory and cardiac motion of these critical substructures. The regularly used respiratory 4DCTs ignore cardiac motion from the respiratory phase rebinning step, and thus, have severe cardiac motion artifacts and cannot accurately inform either cardiac motion or respiratory motion of the heart. Breath-hold cardiac 4DCTs are useful for assessing cardiac-only motion and defining cardiac substructures in the breath-hold state of image acquisition. Because cardiac motion and respiratory motion are not entirely independent, it is ineffective and inadequate to assess the respiratory and cardiac motion of the heart separately on the r4DCTs and c4DCTs. It is necessary to develop a new 5DCT imaging protocol to image patients breathing freely and relaxedly, thus to capture both respiratory motion and cardiac motion simultaneously.

Simulated 5DCT - Axial View

Simulated 5DCT - Coronal View

Simulated 5DCT - Sagittal View

AI and Medical Imaging Science

Human Anatomical Modeling and Image Synthesis

We are developing a novel method to model the 3D human body, organs, structures, and skeletal kinematics. A patient-specific anatomical model will be prepared by fitting a generic human body model to a patient’s CTs or MRIs, as well as the segmentations of organs and structures in CTs and MRIs. The digital anatomical model of the patient can enable or enhance many clinical applications.

The patient-specific anatomical model can serve as a digital twin of the patient. It can be registered to all medical images of the patient, acquired in any body and limb positions and poses. In this way, all medical scans of a patient, information, and derived data defined on the medical scans, such as treatment plans, segmentations, radiation dose, etc., can be managed and subsequently analyzed. Downstream clinical applications include treatment response evaluation, radiation dose accumulation, and decision support for treatments overlapping with prior treatments. There was no existing solution for such clinical applications, and the new digital anatomical model will fill the gap.

Another important application of the patient-specific anatomical model is to animate body and limb postures, and to synthesize new corresponding CT and MRI images. The anatomical model and data processing pipeline will essentially allow editing of CT or MRI scans into new body postures while retaining anatomical fidelity. The capability of results of editing CTs and MRIs post-acquisition will be broadly useful for supporting surgery and radiation therapy treatment simulation, planning, optimization, and education. Such capabilities to edit a patient’s CTs into different body and limb poses will shorten the time from consultation to treatments, reduce the medical image cost, and avoid extra radiation to the patients.

The workflow to prepare the patient-specific anatomical model

The workflow to map all images and derived treatment data to the common reference of the patient.

The workflow to animate the patient-specific anatomical model to simulate different body and limb poses for supporting different clinical needs.

The image synthesis workflow

Examples of synthesized liver CT data generated by the proposed pipeline.

Physics and Model Based Conditional Image Synthesis

Accurate delineation of liver blood vascular structures is crucial for planning and performing therapeutic interventions in liver-related medical procedures. However, current deep learning models often underperform due to the limited availability of high-quality annotated datasets, particularly for fine vascular branches.

We developed a comprehensive, physics-based simulation pipeline to synthesize ground truth liver vessel CT datasets. The pipeline allows flexible control over anatomical variation, vessel contrast, image noise and artifacts, generating synthetic data that closely mimics clinical imaging conditions.

The workflow of conditional diffusion model for abdominal organ segmentation ground-truth CT image synthesis. Our implementation of the conditional DDPM model followed the basic two-step process but with a few tweaks. To guide the image generation, the ground truth organ segmentation masks were assembled into a one-hot segmentation label map and used as the conditional 2nd channel, together with the random noise map, to initiate the image generation.

Deformable Image Registration and Validation

Our research on deformable image registration (DIR) included developing DIR algorithms, AI models, benchmarking dataset libraries, and automated procedures to verify case-specific DIR results. We have published heavily on DIR-related topics. Our recent efforts on DIR have been more focused on quantitative validation and benchmarking library generation.

Up to 2025, we have systematically worked through multiple anatomical sites and prepared DIR benchmarking dataset libraries for the thorax (CT), abdomen (CT), liver (CT), and brain (MRI).

Please check out the list of papers for all our DIR algorithms, DIR landmark detection algorithms, and DIR benchmarking dataset libraries.

The generic workflow for preparing the DIR benchmarking dataset libraries

An example of automatically detected benchmarks in a lung CT case

Diffuse Gliomas MRIs Analysis

Diffuse gliomas are the most common primary brain cancer in adults, with the most aggressive and common form glioblastoma (GBM) having a median survival of only 15 months. Despite advances, current clinical imaging protocols lack the precision to track microscopic tumor infiltration that later becomes recurrence. Determining the precise site of future recurrence on preoperative imaging could allow improved treatments such as boosted radiation dose to regions at risk for recurrence. Deformable image registration (DIR) enables the alignment of longitudinal MRI scans to achieve this task, but existing DIR approaches suffer from limited accuracy and unreliable verification, hindering clinical adoption. Our lab has set out to implement our expertise in DIR methodologies to address these limitations. We aim to develop a verifiable and accurate DIR method for diffuse glioma MRIs by incorporating AI-based blood vessel segmentation and bifurcation matching into a new high-precision DIR approach. By improving registration accuracy, this project can improve the precision of GBM treatment planning, enhance recurrence detection, detect tumor progression, and ultimately improve patient outcomes.

The comprehensive workflow and examples of processing brain GBM MRIs, segmenting blood vessels, and identifying the corresponding vessel bifurcation points.

Page updated

Google Sites

Report abuse