Research

Physics-driven medical image reconstruction

Every MRI exam involves a trade-off between scan time and image quality. Long scans limit how many patients can be seen, are uncomfortable, and invite motion artifacts. Shorter scans leave gaps in the measured k-space data, and these gaps must be filled in during reconstruction. We develop deep learning methods that do this by pairing learned image priors with an explicit model of the imaging physics, so that each reconstruction stays faithful to what the scanner actually measured.

Our recent work spans diffusion priors that adapt to each scan at test time, diffusion bridges constrained in the Fourier domain, zero-shot reconstruction that needs no fully-sampled training data, and state-space models that unroll physics-driven updates. We carry the same principles over to low-dose CT, where the aim is to lower radiation dose without giving up diagnostic quality. Together, these methods work toward faster and more comfortable exams with images that clinicians can rely on.

Standard diffusion compared with the Fourier-constrained diffusion bridge (FDB) for MRI reconstruction
FDB, IEEE Transactions on Medical Imaging, 2026

Multi-modal medical image synthesis

Clinical protocols often call for several MRI contrasts, or for both MRI and CT, because each one reveals a different aspect of the same tissue. In practice, some of these scans end up missing because of limited scan time, cost, radiation dose or patient tolerance. We develop generative models that synthesize the missing contrasts and modalities from the scans at hand.

Our models range from conditional GANs and vision transformers to adversarial diffusion models, recursive diffusion bridges and selective state-space models. Several of them learn from unpaired or undersampled data, which reduces the need for large curated datasets. Reliable synthesis can shorten imaging protocols, spare patients additional scans, and give downstream analysis a complete set of images to work with.

Architecture of SelfRDB, a self-consistent recursive diffusion bridge for medical image translation
SelfRDB, Medical Image Analysis, 2025

Foundation and vision-language models

Foundation models learn broad representations from large and diverse collections of data, and can then be adapted to many clinical tasks with little task-specific supervision. This is an attractive route for medical imaging, where labeled examples are costly and each new task has traditionally required a model of its own.

We develop foundation models for medical imaging and study how to train and adapt them under realistic constraints. Our work includes generative priors trained jointly across institutions without pooling patient data, physics-informed transformers for image reconstruction, and parameter-efficient adaptation of vision-language models that link images to the language of clinical reports. The same ideas extend beyond imaging, to models of single-cell gene expression. The aim is models that carry knowledge across tasks, sites and modalities, rather than being rebuilt for every new problem.

ACE-LoRA adapts frozen image and text encoders of a medical vision-language model with LoRA layers and a graph-attentive context enhancement module, trained on chest X-rays and radiology reports
ACE-LoRA, arXiv preprint, 2026

Federated and privacy-preserving learning

The data needed to train robust models are spread across hospitals, and patient privacy makes pooling them difficult. Federated learning lets institutions train models together while every image stays on site. We develop federated methods for medical imaging that cope with the real differences between sites, in scanners, protocols and patient populations.

Our work includes generative image priors shared across institutions for MRI reconstruction, personalized models that handle a different set of contrasts at each site for MRI synthesis, and approaches that let each site keep its own model architecture. These methods allow models to benefit from the data of many hospitals without the data ever leaving them.

Personalized federated learning across sites for multi-contrast MRI synthesis (pFLSynth)
pFLSynth, Medical Image Analysis, 2024

Medical image analysis

Medical images hold rich diagnostic information, and reading them consistently at scale places a growing burden on clinicians. We develop deep learning methods that extract this information for disease classification, lesion analysis and image retrieval, with a focus on the subtle and co-occurring findings seen in clinical practice.

Our work includes transformer architectures for multi-label disease classification on chest X-rays, metric learning for retrieving similar cases from large image archives, and deep clustering for skin lesion analysis. We also use vision-language models to align images with clinical findings. Such tools can support clinicians with faster triage, easier comparison with prior cases, and more consistent reads.

HydraViT architecture: CNN spatial encoder, transformer context encoder and multi-branch output for multi-label chest X-ray classification
HydraViT, Biomedical Signal Processing and Control, 2025

Magnetic particle imaging and imaging systems

Magnetic particle imaging (MPI) is an emerging modality that maps the distribution of magnetic nanoparticles with high sensitivity and without ionizing radiation, with potential uses in vascular imaging, cell tracking and cancer imaging. Its image quality depends on a calibrated system matrix, which is slow to measure and has to be measured again when the setup changes.

We develop deep learning methods that recover the system matrix from a fraction of the calibration measurements, and that reconstruct images with learned priors while keeping them consistent with the measured signal. These steps make the modality faster to calibrate and more practical to use. The same principles carry over to other calibration-dependent systems, such as focal plane array imaging.

TranSMS architecture: parallel transformer and convolutional modules fused and upsampled with data consistency for super-resolution MPI system matrix calibration
TranSMS, IEEE Transactions on Medical Imaging, 2022

Functional brain imaging and computational neuroscience

We use functional MRI and machine learning to study how the human brain represents visual, auditory and linguistic information during natural behavior, and how attention reshapes these representations across the cortex. Mapping this organization is a step toward linking brain activity to perception and cognition, in health and in disease.

On the methods side, we develop models for analyzing fMRI time series, including window-based transformers, graph neural networks with improved temporal sensitivity, and diffusion-based counterfactual explanations that reveal what drives a model's decisions. Clinical collaborations include fMRI neurofeedback and connectivity studies in Parkinson's disease.

Datasets from our neuroscience experiments can be explored in the Brain Viewer.

DreaMR generates counterfactual fMRI samples with diffusion models to explain classifier decisions
DreaMR, IEEE Transactions on Medical Imaging, 2024

Earlier projects

Compressed sensing for accelerated MRI

Recently developed compressed sensing (CS) theory has opened up great avenues for accelerating MRI exams. The CS theory foresees that a compressible image can be sensed from a significantly smaller number of data samples than the Nyquist rate. To recover an image from these reduced set of samples, a linear transform of the image should be well represented by few coefficients. Furthermore, missing data samples should introduce aliasing artifacts with noise-like incoherent structure. MRI images exhibit similar statistics to natural images, and they are often highly compressible. However, the optimal sparsifying transforms and data sampling patterns for distinct body regions are not fully known.

We have ongoing efforts to apply CS to various MRI applications including MR angiography, MRI stem-cell tracking, and neuroimaging. We are developing novel automated parameter tuning algorithms that find dataset-specific reconstruction parameters to maximize image quality. These algorithms will empower practical use of CS reconstructions in clinical environments where manual tuning is impossible.

Compressed sensing for accelerated MRI
Steady-state acquisition methods for rapid MRI

MRI is a powerful modality that depicts the morphology and function of biological tissues noninvasively. Despite its increasingly important role in both clinical and research environments, image quality and spatiotemporal resolution in many MRI applications are limited by prohibitively long scan times. Thus, there has been a growing need for accelerated MRI methods.

To overcome the speed limitations of conventional MRI, we are developing a class of rapid pulse sequences based on steady-state free-precession (SSFP) imaging. SSFP sequences offer high signal-to-noise ratios, but can yield suboptimal tissue contrast and image artifacts due to signal loss. We tailor intelligent reconstruction algorithms to improve tissue contrast and artifact suppression in SSFP images to clinical standards.

Steady-state acquisition methods for rapid MRI
Neural representation of visual objects and actions

Previous studies of visual category representation have identified a limited number of regions in the human brain that are highly responsive for specific categories. Well-known examples of these specialized regions include the fusiform face area (for faces) and the extrastriate body area (for bodies). These results have led to the hypothesis that category representation is mediated by highly localized and dedicated regions. However, humans can recognize thousands of unique visual object and action categories in natural scenes. It is therefore unlikely that a distinct brain region can be dedicated to each category.

To reveal the details of category representation across the entire brain, we are using natural movie stimuli that contain thousands of object and action categories. We first record brain activity with functional MRI while participants passively view these movies. We then employ a powerful voxel-wise modeling framework to quantitatively describe the selectivity of each voxel to each category. Our results imply that many brain regions represent information about a broader set of categories than previously assumed. We are currently examining the details of category representation across both visual and non-visual brain regions.

Neural representation of visual objects and actions
Attentional modulation of neural representations

Although real-world scenes are cluttered with many different objects, humans are extremely adept at finding target objects in natural environments and shifting their attention rapidly between different targets. The precise neural mechanisms that mediate this remarkable ability are yet unknown. Previous studies have reported relatively simple mechanisms that increase the quality of brain activity evoked by attended objects, without affecting the way that information is represented in each brain region. Yet, because there are a limited number of cortical neurons, it seems unlikely that all brain regions retain fixed representations irrespective of behavioral demands.

To study the nature of neural representations during visual search, we use complex natural movies and a powerful computational modeling technique to describe the relationship between visual information and brain activity. Our preliminary results suggest a novel attentional mechanism that reallocates neural resources across the brain to maximize sensitivity for the target and improve target detection under demanding conditions. You can explore our datasets here. We are currently studying the generality and flexibility of this mechanism for vision and for other sensory modalities.

Attentional modulation of neural representations