Assistant Professor
Director of the Computational Core
A.A. Martinos Center for Biomedical Imaging.
Massachusetts General Hospital
Harvard Medical School
Massachusetts Institute of Technology
Contact: adalca@mit.edu
Our group develops machine learning methods and open source systems for medical image analysis. We focus on technically interesting problems motivated by clinical and scientific needs through a cross-institutional group spanning MGB, Harvard, and MIT.
We develop adaptable robust models, and are building AI agents to support complex imaging workflows. I also direct the Computational Core at the Martinos Center, a major computational infrastructure for biomedical imaging and AI across MGB.
I work with companies and other organizations to assess AI technologies and help shape long-term technical direction, particularly where medical imaging, clinical workflows, and emerging AI methods intersect. I am open to selected advisory roles, industry collaborations, and conversations around startups or research spinouts.
Our group continually explores new technical, clinical, and scientific directions. We highlight several current
areas of focus below.
See the full publications list for a more complete view.
Most existing imaging AI tools solve only the narrow task they were trained for, making them difficult to apply off the shelf to new problems or complex end-to-end workflows. We build general-purpose AI systems that adapt to varied imaging tasks from simple prompts, enabling segmentation, registration, measurement, and detection without task-specific retraining. Methods such as UniverSeg, Tyche, ScribblePrompt, MultiVerSeg, VoxelPrompt, and Pancakes show how prompting and in-context learning can turn a single model into an adaptable assistant for clinical and research needs.
Traditional models often fail outside the exact modality, contrast, or resolution they were trained on, limiting their use in new settings. We developed a framework for building procedural simulations of anatomy, pathology, contrast, resolution, and artifacts. These simulations train AI models that generalize across scanners, sites, and populations, and are now a central component of our model-training workflow. Tools such as SynthSeg, SynthMorph, SynthStrip, and Anatomix demonstrate how synthetic diversity can produce robust medical imaging algorithms.
Existing image registration methods require substantial effort and manual tuning, and can be difficult to adapt to new imaging setups or clinical needs. We established core approaches in learning-based image registration and continue to develop fast, accurate methods that combine learning with rigorous deformation modeling. Tools such as VoxelMorph, SynthMorph, HyperMorph, and MultiMorph enable modality-invariant alignment, large-deformation handling, and substantial acceleration for clinical and population-scale imaging studies.