Taming Foundation Model with Invariance-Oriented Pre-Training for Broad-Spectrum EEG Analysis across Signal-Level, Brain-State, and Brain-Health Tasks

Yulong Dou, Han Wu, Guo Chen, Fangmao Ju, Zhiming Cui, Dinggang Shen

Under Submission 2026

Electroencephalography (EEG) is a widely used window into human brain function, but most EEG models remain tied to a one-dataset-one-model supervised paradigm. Here we present INCEPT, an invariance-oriented EEG foundation model trained on over 11,000 hours of unlabelled clinical EEG. (view more ) Electroencephalography (EEG) is a widely used window into human brain function, but most EEG models remain tied to a one-dataset-one-model supervised paradigm. Recent EEG foundation models offer a route toward reusable representations, but most remain reconstruction-centered, assuming that EEG content predictable from local context is necessarily transferable neural information. Here we present INCEPT, an invariance-oriented EEG foundation model trained on over 11,000 hours of unlabelled clinical EEG. Rather than prioritizing signal recovery alone, INCEPT learns representation-level stability across correlated EEG observations, separating stable neural structure and essential subject-sensitive information from the nuisance variability that dominates scalp recordings while preserving subject-, state- and condition-discriminative information. We evaluate INCEPT on a broad-spectrum benchmark of ten datasets spanning three levels of post-acquisition EEG analysis: signal-level assessment, brain-state decoding, and brain-health evaluation. INCEPT ranks first among recent EEG foundation models on 26 of 30 linear-probing metrics and 24 of 30 fine-tuning metrics, and also surpasses strong task-specific specialist encoders across diverse downstream settings. Objective ablations and representation analyses further show that invariance-oriented pre-training improves transfer and organizes subject-sensitive neural representations beyond reconstruction alone. These results establish invariance learning as a promising principle for building reusable EEG foundation models. (view less )

[Paper] [Code]

COMET: Controllable Orthodontic Video Editing with Multi-view Estimation

Lanzhuju Mei, Yulong Dou, Han Wu, Guo Chen, Zhiming Cui, Dinggang Shen

Under Submission 2026

Orthodontic treatment visualization is critical for patient decision-making and expectation management during lengthy and complex dental alignment procedures. We introduce COMET, a diffusion-based framework for controllable orthodontic video editing with multi-view estimation, which predicts structurally consistent post-treatment outcomes from pre-treatment videos and provides intuitive dynamic visualizations of expected results. (view more ) Orthodontic treatment visualization is critical for patient decision-making and expectation management during lengthy and complex dental alignment procedures. Existing methods predict static 2D images, and patients struggle to envision dynamic treatment outcomes from fixed-pose photographs alone. Two difficulties stand in the way of video-based prediction. First, paired pre- and post-treatment photographs are rarely available in real-world settings, so there is little direct supervision for learning orthodontic transformations. Second, the fine geometry of the teeth must stay consistent while head pose, viewpoint, and tooth visibility change from frame to frame. To address these difficulties, we introduce COMET, a diffusion-based framework for Controllable Orthodontic video editing with Multi-view EsTimation, which predicts structurally consistent post-treatment outcomes from pre-treatment videos and provides intuitive dynamic visualizations of expected results. COMET integrates an adapted ControlNet into a pretrained video diffusion model and augments it with two conditioning modules: (1) an Orthodontic Predictor that learns treatment-induced tooth movements from paired clinical 3D dental models and transforms the pre-treatment image in the first frame into a post-treatment image, and (2) a Multi-view Generator that estimates tooth contours across viewpoints so that the dental structures stay consistent across frames. In a user study with 18 participants, including three dentists, COMET receives the highest ratings on all criteria, and it improves tooth integrity from 0.74 to 0.87 over the best competing method on real-world videos. COMET provides a practical tool for offline orthodontic visualization and patient consultation. (view less )

A Unified Latent Diffusion for High-Fidelity Any-to-Any Brain Modality Synthesis

Yulong Dou*, Guo Chen*, Chenfan Xu, Yulin Wang, Zhe Xu, Zhiming Cui, Dinggang Shen (* equal contribution)

6th Deep Generative Models Workshop at MICCAI (DGM4MICCAI) 2026 (Long Oral)

Comprehensive neurological diagnosis relies on the synergy of multimodal data from MRI, CT, and PET to capture distinct anatomical and pathological markers. To address incomplete imaging suites and the limitations of paired training data, we present BrainDiff, a unified latent diffusion framework for flexible any-to-any brain modality synthesis. (view more ) Comprehensive neurological diagnosis relies on the synergy of multimodal data from MRI, CT, and PET to capture distinct anatomical and pathological markers. However, clinical reality often results in incomplete imaging suites due to time constraint, high costs, or patient contraindication. Although cross-modality synthesis offers a potential solution, existing methods are limited by their reliance on strictly paired training data. This constraint prevents them from utilizing abundant single-modality data and from learning the distinct characteristics of each imaging domain, resulting in limited generative fidelity. To address these challenges, we present BrainDiff, a unified latent diffusion framework for flexible any-to-any brain modality synthesis. Our method employs a universal VQ-VAE to project all modalities into a shared latent space, where a prompt-guided diffusion model synthesizes corresponding images conditioned on descriptive prompts (e.g., patient attributes, imaging parameters). During cross-modality synthesis, a ControlNet module is incorporated to provide guidance that preserves anatomical structure of the source-modality image. Extensive multi-institutional experiments demonstrate that BrainDiff achieves state-of-the-art performance, delivering high-fidelity and anatomically consistent results for any-to-any generation. Code is available at https://github.com/douyl/BrainDiff. (view less )

[Code]

CMD-AD: Cross-Modality Distillation from fMRI-EEG for Single-Modality Alzheimer's Disease Diagnosis

Xin Lin*, Yuxiao Liu*, Wenfei Yu*, Yulong Dou, Weiqi Li, Minhui Tan, Yuanzhe He, Zhiming Cui^#, Dinggang Shen^# (* equal contribution, ^# corresponding author)

International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2026

Alzheimer’s disease (AD) disrupts both the spatial organization and temporal dynamics of brain functional connectivity, which can be jointly characterized by fMRI and EEG. We propose CMD-AD (Cross-Modality Distillation for AD diagnosis), a three-stage framework that distills fused spatiotemporal knowledge into single-modality encoders to improve their performance. (view more ) Alzheimer’s disease (AD) disrupts both the spatial organization and temporal dynamics of brain functional connectivity, which can be jointly characterized by fMRI and EEG. fMRI-derived functional connectivity networks (FCNs) encode multiscale spatial topology, and EEG-derived FCNs capture multiband temporal coupling patterns, providing complementary information for AD diagnosis. However, paired fMRI-EEG acquisitions are rarely available in clinical practice, and existing fusion methods require both modalities at inference, limiting real-world deployment. To address this challenge, we propose CMD-AD (Cross-Modality Distillation for AD diagnosis), a three-stage framework that distills fused spatiotemporal knowledge into single-modality encoders to improve their performance. In Stage 1, hierarchical graph attention aggregation (HGAA) encoders are independently pretrained on each modality to extract multiscale fMRI spatial features and multiband EEG temporal features. In Stage 2, scale-wise bidirectional cross-modality fusion (BCMF) integrates fMRI and EEG representations at each parcellation scale, producing fused spatiotemporal targets. In Stage 3, teacher-student distillation transfers the fused spatiotemporal targets back to each modality-specific encoder, enabling accurate single-modality diagnostic performance without paired inputs. Evaluated on 8 multicenter datasets, CMD-AD improves fMRI-only accuracy from 73.51% to 75.50% and EEG-only accuracy from 63.52% to 65.02%, with AUC increasing from 80.88% to 83.70% and from 71.63% to 73.77%, respectively. These results demonstrate that cross-modality training can effectively enhance single-modality deployment for AD diagnosis. Our code is available at https://github.com/yizhixiaolinya/Code-MICCAI2026_fMRI-EEG.git. (view less )

[Code]

OrthoFlow: Decoupled Landmark Prediction and Image Synthesis for Orthodontic Treatment Outcome Visualization

Leyuan Wang*, Yulong Dou*, Jiancheng Yang, Guangying Song, Zhiming Cui (* equal contribution)

International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2026

Predicting post-treatment facial profiles from lateral cephalograms is clinically valuable for orthodontic planning and patient communication, yet remains challenging due to the complex interplay between skeletal, dental, and soft-tissue changes. We propose OrthoFlow, a two-stage decoupled generative framework that separates geometric prediction from image synthesis. (view more ) Predicting post-treatment facial profiles from lateral cephalograms is clinically valuable for orthodontic planning and patient communication, yet remains challenging due to the complex interplay between skeletal, dental, and soft-tissue changes. Conventional approaches either predict discrete clinical indices without visual output, or apply direct image-to-image translation that fails to preserve anatomical consistency. We propose OrthoFlow, a two-stage decoupled generative framework that separates geometric prediction from image synthesis. In the first stage, a Flow Matching model with a Graph-Transformer hybrid backbone predicts post-treatment landmark displacements conditioned on textual treatment plans. In the second stage, a conditional Flow Matching model synthesizes high-fidelity lateral cephalograms guided by the predicted anatomical contours. Experiments on a large-scale orthodontic dataset demonstrate that OrthoFlow achieves superior geometric accuracy and visual fidelity compared to existing methods, validated by both quantitative metrics and clinical evaluation by orthodontists. Our code and dataset are available at https://github.com/ShanghaiTech-IMPACT/OrthoFlow. (view less )

[Code]

BrainParc: Unified Lifespan Brain Parcellation from Structural MR Images

Jiameng Liu, Feihong Liu, Kaicong Sun, Zhiming Cui, Tianyang Sun, Zehong Cao, Jiawei Huang, Shuwei Bai, Yulin Wang, Yulong Dou, Kaicheng Zhang, Caiwen Jiang, Yuyan Ge, Han Zhang, Feng Shi, Dinggang Shen

Nature Computational Science 2026

Accurate brain parcellation from structural MRI (sMRI) across the human lifespan is essential for advancing neuroimaging and neuroscience studies. However, existing methods often struggle to generalize due to intensity and contrast variations across brain maturation, aging, and differences in MRI acquisition protocols, limiting their clinical and research utility. (view more ) To address this, we present BrainParc, a unified parcellation framework that leverages anatomical information invariant to intensity and contrast, enabling accurate, robust, and longitudinally consistent parcellation across heterogeneous dataset without the need for fine-tuning. We curated the largest known sMRI dataset to date, comprising over 93300 T1w MRI data from 19 globally distributed cohorts spanning ages 0-100 years, to train and rigorously evaluate BrainParc. Extensive experiments on both internal and external datasets demonstrate that BrainParc significantly outperforms state-of-the-art methods in delineating 106 brain regions, achieving average Dice scores of 95.48% internally and 94.10% externally. BrainParc consistently shows superior performance across diverse populations and imaging conditions, both quantitatively and qualitatively. Beyond anatomical segmentation, we show that BrainParc enables reliable tracking of brain development and facilitates early diagnosis of neurological disorders, underscoring its potential as a robust and generalizable tool for large-scale neuroimaging studies and clinical translation. (view less )

[Paper] [Code]