Brain imaging × interpretable AI
We study how foundation models, normative inference, and rigorous validation can make fMRI useful for psychiatric research and therapeutic discovery.
Move across the signal to inspect time
Programme
01 / Research
How can a model learn from population-scale brain data and still say something calibrated, interpretable, and useful about one person?
Masked prediction can turn long, unlabeled fMRI sequences into transferable representations of functional networks, future brain states, and clinical variables.
Our direction is to compare ROI-, surface-, and voxel-level encoders that retain slow BOLD dynamics while separating biology from scanner, site, and protocol effects.
“Can a pretrained representation remain useful when the cohort, scanner, and clinical endpoint all change?”
The frontier is moving beyond single-modality classifiers toward compact tokens that jointly encode structural MRI and heterogeneous fMRI time series.
We explore structure–function priors, shared hub tokens, and modality-aware fusion as a route to generalisation across neurodevelopmental and psychiatric questions.
“What should be shared across modalities—and what information must remain modality-specific?”
Population reference models can show where an individual task-evoked response departs from expectation, retaining heterogeneity erased by case–control averages.
We focus on calibrated deviation maps, transdiagnostic symptom dimensions, test–retest uncertainty, and multisite reference cohorts rather than binary diagnosis alone.
“Can deviation maps become stable enough to follow an individual through time and treatment?”
Emerging multimodal work uses pretrained representations to reduce data scarcity in pharmaco-fMRI and predict response from resting-state networks and clinical variables.
We treat this as evidence generation: quantify uncertainty, require external validation, and support mechanism-of-action hypotheses without replacing clinical judgment.
“Which representation can travel from retrospective imaging to a prospective therapeutic study?”
These external references define the scientific frontier that informs the programme; they are distinguished from completed community work.
02 / Evidence line
Each project becomes a traceable evidence line—from signal quality to a decision that can be challenged.
Preprocessing, physiology and motion QC, cohort-aware alignment
Dynamic connectivity, graph states, voxel and ROI tokens
Nested CV, site-held-out testing, calibration and ablation
Stable attribution, deviation maps and quantified uncertainty
Clinical endpoints and reproducible evidence packages
Research principle
Complex models only matter when the evidence can travel beyond the lab.
03 / Community
Explore how brain dynamics, clinical translation, robust inference, and discovery fit around the same problem.
Research lead
Research conducted with MIT and during a Google internship focused on high-dimensional fMRI preprocessing, ML/DL model evaluation, and interpretable brain–symptom relationships for neuroscience, psychiatric assessment, and drug-discovery support.
Concept profiles are editorial personae used to map the expertise this community is designed to assemble. Tomohisa Sasaki is the lead researcher profile.
04 / Collaborate
Share the research problem, the evidence you already have, and the decision it needs to support. We will use that to judge whether the right disciplines can be connected.
Selected research anchors