AIMA Research ✦ AIO 2026 cohort ✦ Project call
Three questions about the mind & the reading room.
Decode what the brain sees. Build AI that adapts to the radiologist, not the other way round. Teach the next generation where to look. Each project is a small research group with mentors at US universities.
The 2026 projects
Pick the question you cannot stop thinking about.
Twelve seats, split evenly: four students per project, six months from first paper to submission.
Brain decoding: from fMRI & EEG to images and video
Reconstruct what a person sees from their brain activity, and find out how much of the picture really came from the brain.
4 seatsRead → IIA self-adapting, radiologist-centered AI workflow for CT & MRI screening
An agentic system that triages, drafts and measures, learns from every correction, and knows when to hand the case back.
4 seatsRead → IIIAI-guided eye-gaze training for medical students & junior radiologists
Learn from how experts look at an image, then guide learners where and how to look for abnormalities.
4 seatsRead →Project I · Neuro-AI
Brain decoding: from fMRI & EEG to images and video
In three years, brain decoding has gone from blurry shapes to photo-like images and smooth video. It has also learned that a strong diffusion prior can paint a convincing picture that the brain never saw. Our group will build on the best recent methods and ask how much of each reconstruction truly came from the brain.
Seeing, recording, decoding
1 What they see
a real photograph
2 The viewer
3 Recorded signal
4 Decoder
brain encoder → shared latent
5 Reconstruction
Photos: CC0, Wikimedia Commons. Steps 3–5 are illustrations of how decoders work, not output from a real recording; the reconstructions are softened versions of the photo made for this page.
Three years of the field, on one map
Thirty-one papers from 2023 to 2026, most at CVPR, ICCV, ECCV, NeurIPS, ICML, ICLR, AAAI and Nature-family journals. Filter by signal and output; hover or tap a paper for its one-line summary and link.
Eight open research gaps
Each gap is backed by the papers above and sized for four students and one to four GPUs. The group picks one or two with its mentors in month one.
01Cross-subject decoding with little or no data
+
Shared-subject models cut new-subject data from 40 hours to about one3–6, and zero-shot decoding is emerging7. But zero-shot still trails fine-tuning, and almost all results come from the same 7–8 NSD subjects24.
DirectionA leave-one-subject-out benchmark of alignment methods (linear, transfer matrix, LoRA, adversarial) at 0, 15 and 60 minutes of new-subject data, plus a new hyperalignment-adapter baseline.
02Faithfulness: is it the brain, or the prior?
+
A 30–50-dimensional bottleneck recovers most reconstruction quality9, and text-guided NSD models behave like category classifiers plus diffusion “hallucination”8. Common metrics reward plausible pictures, not faithful ones.
DirectionA faithfulness audit suite (bottleneck, shuffled-signal and noise controls, out-of-distribution test sets, category-leakage checks) used to re-score 4–5 open-source decoders. Mostly inference, so it needs little compute.
03The EEG fidelity gap
+
EEG is cheap and portable, but reconstructions are semantically right and low-level wrong11,12. MEG decodes high-level features well over time13. Older EEG datasets carry a block-design confound18.
DirectionEEG-to-image on THINGS-EEG225 with a pretrained EEG foundation encoder23 and a separate low-level branch, reporting low- and high-level metrics separately. Trains on a single RTX 4090.
04Low-level detail versus meaning
+
Most strong decoders split the problem into a low-level layout path and a semantic path2,16, but the trade-off between them is tuned by hand.
DirectionA controlled study of the low-level/semantic Pareto front per brain region (early visual cortex versus higher areas) on NSD, with region-specific adapters.
05Temporal consistency in video decoding
+
fMRI is slow and the standard video set has only three subjects27. Motion can be invented by the video generator; permutation tests are needed to show it was decoded10,15–17.
DirectionAn explicit motion-decoding head with motion-shuffled baselines and temporal metrics on CC2017. Needs an A100 for video diffusion.
06Real-time, portable decoding
+
07Generalizing beyond one dataset
+
08Brain to language, responsibly
+
Reproduce and audit
Reproduce two open decoders (for example MindEye2 and ATM) and run the faithfulness controls from gap 02 on them.
Close one gap
Propose a method for the chosen gap: few-data cross-subject decoding, or low-level fidelity for EEG.
Evaluate honestly, then publish
Test on a second dataset, release code, and submit to a NeurIPS/ICLR workshop or a MICCAI-track venue.
Data
You will learn
- fMRI and EEG preprocessing
- Contrastive alignment and diffusion priors
- Cross-subject adaptation (LoRA, alignment)
- Evaluating generative models honestly
Good fit if you
- Like neuroscience as much as generative models
- Are comfortable with PyTorch and GPUs
- Enjoy asking “is this result real?”
Data & signals
Preprocessing pipelines for NSD, THINGS and EEG; subject splits.
Baselines
Reproduces open decoders and keeps results reproducible.
Method
Designs the alignment, encoder or adapter for the chosen gap.
Evaluation
Owns metrics, faithfulness controls and figures.
Project II · Human-centered clinical AI
A self-adapting, radiologist-centered AI workflow for CT & MRI screening
We use AI agents to automate high-skill work on our own devices every day, yet the radiologist’s workflow has barely changed. We want to build a trustworthy agentic system that works with the radiologist in the loop: it triages, drafts and measures, learns from every correction, and knows when to step back.
The machine and the radiologist, working together
Chest CT of a solitary fibrous tumour of the pleura: Sean Novak, Wikimedia Commons, CC BY-SA 4.0. The outline, labels and report lines are illustrative.
Why today’s AI does not close the gap
Across 140 radiologists and 15 chest X-ray tasks, the effect of AI assistance varied widely from one reader to the next. Experience, subspecialty and prior AI use did not predict who would benefit, and wrong AI predictions made readers worse29.
- Radiologists underweight AI advice and treat it as independent of their own read. Routing each case to either the human or the AI can do better than giving everyone the same assistance30.
- Incorrect AI suggestions pulled mammography readers at every experience level toward the wrong answer31.
- Explanations raised physician trust whether the AI was right or wrong32.
radiologists: the effect of the same AI assistance differed widely from reader to reader, and could not be predicted from experience.
Yu et al., Nature Medicine 2024 ↗What makes this possible now
Deferral models learn when to hand a case to the clinician33. Clinicians and vision-language models already co-write reports34. LLM agents can orchestrate imaging tools without retraining35. What is missing is a system that adapts to each reader over time and is evaluated on whether readers actually do better.
An agentic first read
An agent that orchestrates open segmentation and detection tools on public CT/MRI to triage cases, measure findings and draft structured reports, with calibrated uncertainty.
Learning from every correction
Turn accept, edit and reject actions into feedback: per-reader calibration, deferral (“this one is yours”), and online updates that avoid learning a reader’s mistakes.
Measure trust, not just accuracy
Simulated readers first, then a small pilot with mentor radiologists: accuracy, reading time, over-reliance when the AI is wrong, and how that changes across sessions.
Data & tools
- Public CT/MRI screening and segmentation datasets
- Open segmentation and foundation models as agent tools
- An LLM or VLM planner, run locally where possible
You will learn
- Agentic AI and tool use
- Uncertainty, calibration and learning to defer
- Human-AI study design and clinical evaluation
Good fit if you
- Want to build systems people actually use
- Like software engineering as much as modelling
- Care about what clinicians need
Agent & tools
Planner, tool calling and the first-read pipeline.
Vision models
Segmentation, detection, measurement and uncertainty.
Adaptation
Feedback learning, per-reader calibration and deferral.
Interface & study
Reading interface, simulated readers and the pilot study.
Project III · AI for medical education
AI-guided eye-gaze training for medical students & junior radiologists
AI now matches or beats less-experienced readers on several imaging tasks. Instead of only giving learners the answer, we want AI that teaches them how to find it: software that learns from where expert radiologists look, then guides a student’s gaze toward the regions and the search pattern an expert would use.
Learning where to look
Chest radiograph of the same pleural tumour: Sean Novak, Wikimedia Commons, CC BY-SA 4.0. Scanpaths are illustrative: numbered circles are fixations (larger means longer), the glow is the gaze heatmap.
The evidence
- Deep-learning systems outperformed non-radiologist physicians and radiologists on chest radiographs, and improved every reader as a second opinion36,37.
- A stand-alone mammography AI was non-inferior to the average of 101 radiologists38.
- LLM help raised residents’ brain-MRI diagnostic accuracy far more than neuroradiologists’39, and wrong AI suggestions misled inexperienced readers the most31.
The gap we target
Gaze already helps train better models42–45, and public chest X-ray gaze datasets exist40,41. Yet a systematic review found little evidence that teaching novices to “search like an expert” improves diagnosis46. Whether AI-guided gaze training works is an open question, and we will test it.
Model the expert
Train on expert-radiologist gaze only to predict, for a new image, the heatmap and the ordered scanpath an expert would follow.
Guide the learner
Software that tracks a learner’s gaze (webcam or low-cost tracker), compares it with the expert model, and suggests the next region to check and what it has missed.
Does it teach?
A pilot with medical students: guided versus unguided practice, measuring detection, coverage and whether the skill stays once guidance is turned off.
Data
You will learn
- Eye-tracking data and saliency / scanpath models
- Vision transformers for medical images
- Designing an education study
Good fit if you
- Are a medical student or have a clinical interest
- Like building interactive tools
- Want to study how people learn
Gaze data
REFLACX / EGD-CXR processing, fixations and scanpaths.
Expert model
Heatmap and scanpath prediction from expert gaze.
Guidance app
Real-time gaze capture and the guidance interface.
Education study
Study design, metrics and analysis with mentors.
Six months
From first paper to submission.
All three groups share one rhythm, with weekly group meetings and a monthly review with mentors.
Onboard
Read 10–15 key papers, set up data access and compute, write a one-page proposal.
Reproduce
Match a published baseline. Share a reproducibility report.
Propose
First version of the method, defended in a mentor review.
Iterate
Ablations, failure analysis, fix what the review found.
Validate
External data or pilot study; freeze results.
Submit
Paper, code release and a talk for the AIMA community.
Twelve seats · four per project
Tell us which question
you want to answer.
Write to us with the project number, a short note on your background, and what you want to learn. Selection details for the AIO 2026 cohort will be announced by AIMA.
References
- Takagi & Nishimoto. High-resolution image reconstruction with latent diffusion models from human brain activity. CVPR 2023. link
- Scotti et al. Reconstructing the Mind’s Eye: fMRI-to-image with contrastive learning and diffusion priors. NeurIPS 2023. arXiv
- Scotti, Tripathy et al. MindEye2: Shared-subject models enable fMRI-to-image with 1 hour of data. ICML 2024. PMLR
- Wang et al. MindBridge: A cross-subject brain decoding framework. CVPR 2024. arXiv
- Gong et al. MindTuner: Cross-subject visual decoding with visual fingerprint and semantic correction. AAAI 2025. link
- Dai et al. MindAligner: Explicit brain functional alignment for cross-subject visual decoding from limited fMRI data. ICML 2025. PMLR
- Wang et al. ZEBRA: Towards zero-shot cross-subject generalization for universal brain visual decoding. NeurIPS 2025. OpenReview
- Shirakawa et al. Spurious reconstruction from brain activity. Neural Networks 2025. DOI
- Mayo et al. BrainBits: How much of the brain are generative reconstruction methods using? NeurIPS 2024. arXiv
- Lu et al. Animate your thoughts: Decoupled reconstruction of dynamic natural vision from slow brain activity (Mind-Animator). ICLR 2025. arXiv
- Song et al. Decoding natural images from EEG for object recognition (NICE). ICLR 2024. arXiv
- Li et al. Visual decoding and reconstruction via EEG embeddings with guided diffusion (ATM). NeurIPS 2024. arXiv
- Benchetrit, Banville & King. Brain decoding: toward real-time reconstruction of visual perception. ICLR 2024. arXiv
- Kneeland et al. ENIGMA: A unified lightweight EEG-to-image model for multi-subject visual decoding. NeurIPS 2025 Workshop. link
- Chen, Qing & Zhou. Cinematic Mindscapes: High-quality video reconstruction from brain activity (MinD-Video). NeurIPS 2023. arXiv
- Gong et al. NeuroClips: Towards high-fidelity and smooth fMRI-to-video reconstruction. NeurIPS 2024. arXiv
- Wang et al. NEURONS: Emulating the human visual cortex improves fidelity and interpretability in fMRI-to-video reconstruction. ICCV 2025. CVF
- Li et al. The perils and pitfalls of block design for EEG classification experiments. IEEE TPAMI 2021. DOI
- Xia et al. UMBRAE: Unified multimodal brain decoding. ECCV 2024. arXiv
- Qiu et al. MindLLM: A subject-agnostic and versatile model for fMRI-to-text decoding. ICML 2025. PMLR
- Tang et al. Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience 2023. DOI
- Horikawa. Mind captioning: Evolving descriptive text of mental content from human brain activity. Science Advances 2025. DOI
- Jiang et al. Large brain model for learning generic representations with tremendous EEG data in BCI (LaBraM). ICLR 2024. arXiv
- Allen et al. A massive 7T fMRI dataset to bridge cognitive neuroscience and artificial intelligence (NSD). Nature Neuroscience 2022. DOI
- Gifford et al. A large and rich EEG dataset for modeling human visual object recognition (THINGS-EEG2). NeuroImage 2022. DOI
- Hebart et al. THINGS-data: fMRI, MEG and behavioral data on object representations. eLife 2023. DOI
- Wen et al. Neural encoding and decoding with deep learning for dynamic natural vision (CC2017). Cerebral Cortex 2018. DOI
- Alljoined: EEG datasets for natural-image decoding, including consumer-grade Alljoined-1.6M. arXiv · arXiv
- Yu, Moehring, Banerjee, Salz, Agarwal & Rajpurkar. Heterogeneity and predictors of the effects of AI assistance on radiologists. Nature Medicine 2024. DOI
- Agarwal et al. Combining human expertise with artificial intelligence: Experimental evidence from radiology. NBER Working Paper 31422, 2023. link
- Dratsch et al. Automation bias in mammography: The impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology 2023. DOI
- Prinster et al. Care to explain? AI explanation types differentially impact chest radiograph diagnostic performance and physician trust in AI. Radiology 2024. DOI
- Dvijotham et al. Enhancing the reliability and accuracy of AI-enabled diagnosis via complementarity-driven deferral to clinicians (CoDoC). Nature Medicine 2023. DOI
- Tanno et al. Collaboration between clinicians and vision–language models in radiology report generation. Nature Medicine 2025. DOI
- Fallahpour et al. MedRAX: Medical reasoning agent for chest X-ray. ICML 2025. arXiv
- Hwang et al. Development and validation of a deep learning–based automated detection algorithm for major thoracic diseases on chest radiographs. JAMA Network Open 2019. link
- Nam et al. Development and validation of deep learning–based automatic detection algorithm for malignant pulmonary nodules on chest radiographs. Radiology 2019. DOI
- Rodriguez-Ruiz et al. Stand-alone artificial intelligence for breast cancer detection in mammography: Comparison with 101 radiologists. JNCI 2019. DOI
- Performing best when needed least: Reader experience shapes accuracy gains in LLM-assisted brain MRI differential diagnosis. Radiology 2026. DOI
- Karargyris et al. Creation and validation of a chest X-ray dataset with eye-tracking and report dictation for AI development (EGD-CXR). Scientific Data 2021. DOI
- Bigolin Lanfredi et al. REFLACX, a dataset of reports and eye-tracking data for localization of abnormalities in chest x-rays. Scientific Data 2022. DOI
- Bhattacharya et al. GazeRadar: A gaze and radiomics-guided disease localization framework. MICCAI 2022. link
- Bhattacharya et al. RadioTransformer: A cascaded global-focal transformer for visual attention-guided disease classification. ECCV 2022. arXiv
- Wang et al. Follow my eye: Using gaze to supervise computer-aided diagnosis. IEEE TMI 2022. arXiv
- Ma et al. Eye-gaze-guided vision transformer for rectifying shortcut learning. IEEE TMI 2023. DOI
- van der Gijp et al. How visual search relates to visual diagnostic performance: A narrative systematic review of eye-tracking research in radiology. Advances in Health Sciences Education 2017. DOI