We design new machine learning algorithms for medical imaging tasks to contend with current challenges such as data paucity, sensitivity to distribution shifts, and explainability. Note that many of these self/semi/un-supervised models, representation learning methods, and explainable AI techniques are applied to related healthcare problem areas such MRI acquisition, MRI and CT image analysis, and opportunistic analysis of CT scans.

We develop techniques to help turn medical images into medical insights that can be used for downstream prediction of health outcomes. Specifically, we seek to build segmentation tools to automatically detect potential biomarkers of disease activity for varying anatomies and volumetric imaging techniques.

Magnetic resonance imaging (MRI) is an excellent diagnostic imaging modality, however, more widespread use of MRI is limited by its long acquisition duration and costs. We develop new techniques to accelerate acquiring MRI scans without compromising diagnostic value and to convert MRI from a qualitative to quantitative technology.

Diagnostic abdominal CT scans can answer specific clinical questions, however, most information in these data-rich 3D scans is not evaluated because it is invisible to the human eye or too time consuming to analyze manually. We have developed a paradigm of opportunistic CT where automated analysis of bone, fat, and muscle biomarkers is combined with the existing electronic health record (EHR) to screen patients for chronic disorders with high accuracy.