Thanh-Huy Nguyen · AIMA Research
A community of Vietnamese researchers doing translational AI for health. We are mentored by PhD students, postdocs and professors in the United States and Australia, and connected from three hubs in Vietnam to researchers at more than thirty institutions across North America, Europe, the Middle East, Asia and Australia.
Explore projects Join the next cohort
New · AIMA 2026
Our record & the AIO 2026 cohort
47 papers since 2024 at CVPR, ECCV, MICCAI, ISBI, Medical Image Analysis and more
I
AIMA Research started in 2024 to open a path into research for Vietnamese students. Today it is a community of researchers who want their AI to leave the benchmark and reach the clinic. Members work in small research groups, each led by a mentor who has done this before, and learn to think and work as independent researchers with a worldwide mindset.
Every group is led by a PhD student, postdoc or professor at a US university or hospital. Mentors train, steer and review; members own the research.
AI agents and vibe-coding make code cheap. Good questions, careful experiments and honest evaluation are still the hard part, and that is what we train.
Researchers in AI and medicine at universities and hospitals across the United States, Europe, the Middle East, Asia and Australia, who share projects, data expertise and opportunities.
II
From your first paper-reading session to leading your own study.
Apply to a cohort and join a research group of 4–6 members, matched to one of our research themes and led by a mentor.
Read papers critically, build reproducible PyTorch pipelines, handle medical images, and work with AI coding agents without losing rigor.
Own a research question, run the experiments and defend the results in weekly meetings. Mentors guide; they do not do it for you.
Workshop, then conference, then journal. Along the way you build a record for graduate school and a network that lasts.
III
Our strongest line of work is clinically translatable segmentation of organs, tissues and tumors. Around it we study human signals, report generation and, most recently, cognitive and hybrid AI.
IV
Twelve starting points for new research groups, one or more per theme. They are prototypes: each group refines the question, data and methods with its mentor before starting.
Segment the liver, its vessels and every focal lesion from multi-phase CT. Then turn the masks into numbers surgeons use: Couinaud segments, lesion size and the volume of liver left after resection.
You will learn 3D U-Net and nnU-Net, multi-phase fusion, lesion-wise evaluation.
The pancreas is small, thin and shaped differently in every patient. Find it with a coarse-to-fine model, segment it, and flag small cystic lesions early, when they matter most.
You will learn cascaded localisation, small-object segmentation, class imbalance.
Lobes, airways, heart and nodules from a single chest CT, with a chest X-ray branch for settings without CT. Airways test whether a model respects thin, connected structures.
You will learn multi-structure segmentation, topology-aware losses, nodule detection.
Brain tumors in children look and behave differently from adult ones, and far less data exists. Transfer what adult models know, then adapt them to the pediatric sub-regions that guide treatment.
You will learn multi-sequence MRI, transfer learning, learning from small datasets.
A model that scores well on one dataset often fails on another scanner. Test organ and tissue models across sites, detect the shift, estimate uncertainty and flag cases for a clinician instead of failing silently.
You will learn external validation, domain generalisation, calibration and uncertainty.
Small, faint tumors are the ones models miss and clinicians care about. Detect, segment and measure lesions in liver, lung and kidney, and follow them from one scan to the next.
You will learn detection plus segmentation, lesion-level metrics, longitudinal registration.
Every abdominal CT already contains information on muscle and fat. Segment it automatically to screen for sarcopenia and metabolic risk, using scans that were taken for another reason.
You will learn tissue segmentation, slice selection, linking imaging biomarkers to outcomes.
One model for CT, MRI, X-ray, ultrasound and pathology, prompted with a click or a sentence such as “segment the liver”. Compare it honestly with SAM, MedSAM and specialist models.
You will learn foundation models, prompting, training on many datasets at once.
Eye trackers record where a radiologist looks while reading. Use that gaze as free supervision to teach models where to attend, and predict gaze to learn how experts search an image.
You will learn eye-tracking data, attention supervision, human–AI collaboration.
Draft realistic radiology reports in which every finding points to a mask or a measurement. Catch hallucinated findings, and evaluate with clinical metrics rather than word overlap.
You will learn vision-language models, grounding, clinical evaluation of text.
Experts glance at the whole image, focus on what looks wrong, then reason with priors and history. Build agents that read images in the same order and can explain each step.
You will learn agentic vision-language models, visual reasoning, ideas from cognitive science.
Neural networks sometimes produce anatomy that cannot exist. Pair them with a knowledge graph of anatomical rules (what touches what, what contains what) to catch those errors and explain decisions.
You will learn neuro-symbolic methods, knowledge graphs, constraint-based refinement.
The images are synthetic illustrations made from simple shapes and noise, not patient data. The brain figure uses a public MRI slice. Datasets named on the cards are examples to start from.
Open to AIO 2025 students only (deadline January 28, 2026; interviews January 29 – February 1).
Robustness, stability and failure modes of SAM, MedSAM and SAM2 on 2D medical images under controlled noise. Target: a CVPR workshop benchmark, then a Q1 journal.
Project descriptionAn occlusion-robustness benchmark for SAM-based models, using surgical tools, probes and implants as occluders. Target: a CVPR workshop, then a Q1 journal.
Project descriptionHow brain segmentation degrades as labeled data shrinks, across classical and advanced architectures. Target: an international conference, then a Q2 journal.
Project descriptionLiver and liver-tumor CT segmentation on LiTS with progressively fewer labels. Target: an international conference, then a Q2 journal.
Project descriptionCNNs versus Vision Transformers for breast-ultrasound classification under limited labels. Target: an international conference, then a Q2 journal.
Project descriptionWhite-blood-cell classification from blood-smear images with few labels, CNN versus ViT. Target: an international conference, then a Q2 journal.
Project descriptionV
Mentors lead our research groups and run the community day to day: PhD students, postdocs and faculty at universities and medical centers in the United States and Australia. Each appears here as an animal alter ego, drawn for this page.
Assistant Professor, North Carolina A&T State University
Postdoc, MD Anderson Cancer Center
Postdoc, Yale University
Postdoc, Mayo Clinic
Research Fellow, Northwestern University
Research Associate, Carnegie Mellon University
PhD Student, UNC Charlotte
PhD Student, Duke University
PhD Student, Duke University
PhD Student, Florida International University
PhD Student, UNC Chapel Hill
PhD Student, University of Technology Sydney
Last but not least, thank you to every mentee who has worked alongside us and contributed so much to this community. We could not have come this far without you.
VI
Where the people behind AIMA work: our mentors, and the advisors, professors and collaborators who support our groups and are not listed by name. Almost everyone started in Vietnam. From three hubs there, the community now reaches 33 institutions in 8 countries on 4 continents.
Loading the map…
Hanoi, Da Nang and Ho Chi Minh City, where most of us set out from, linked to every city in the network. Hover or tap a dot to see the institutions there. Which hub a city links to is illustrative.
VII
Want to join the next cohort? It will be announced here and on our Facebook page. If one of the projects above excites you, write to us with the project name, a short note on your background, and what you want to learn.
