NeuroDrift: A Continuous-Time Multimodal 3D Brain Foundation Model for Alzheimer's Disease Research

NeuroDrift is an open-source 3D brain foundation model project initiated by RishiShah99. It integrates technologies such as continuous-time flow, multimodal fusion, and causal reasoning, supporting drug intervention simulation and age-span prediction, thus providing a new tool for neurodegenerative disease research. The project is currently in Phase 0, aiming to build a deep learning system that can understand brain changes over time. Its browser-based interactive interface can intuitively display intervention effects.

• Original author/maintainer: RishiShah99
• Source platform: GitHub
• Original link: https://github.com/RishiShah99/neurodrift
• Release date: 2026-06-09
• Current phase: Phase 0 (Procurement and infrastructure construction) The core goal of the project is to create a system that understands brain changes over time, focusing on research into neurodegenerative diseases such as Alzheimer's disease.

NeuroDrift's technical architecture includes nine innovations:

1. 3D volumetric latent space and continuous-time flow: Direct 3D operations to capture continuous changes;
2. Large-scale multi-cohort training data: Planned to use 150,000 MRI images covering ages 9-90;
3. Multimodal fusion: Supports multiple imaging modalities, with modality dropout to enhance practicality;
4. Causal reasoning: Uses IPW and Mendelian randomization for counterfactual simulation;
5. Drug category guidance: 64-dimensional treatment embedding to distinguish intervention types;
6. Calibrated uncertainty estimation: Random interpolator provides probabilistic output;
7. Interpretable mechanism: SAE identifies key biomarker directions;
8. Browser-based 3D rendering: Hierarchical Gaussian decoder enables smooth interaction;
9. Fully open-source: Releases weights, models, and APIs under the MIT license.

• Development tools: uv (Python environment), Node22 (web demo), custom fleet system for GPU cluster scheduling;
• Data pipeline: Follows BIDS standards, with emphasis on quality control;
• Model references: Draws on cutting-edge work such as C4G, Stable Diffusion3, and random interpolators.

NeuroDrift can be applied in:

• Clinical research: Understanding the progression mechanism of Alzheimer's disease;
• Drug development: Virtual screening to accelerate candidate drug evaluation;
• Clinical decision-making: Personalized disease progression prediction to assist treatment; It represents a new paradigm of AI in neuroscience: shifting from categorical diagnosis to dynamic, causal, and intervenable prediction.

• Phase 0: Data procurement and infrastructure construction (current phase);
• Phase1: Multimodal 3D VAE and life-cycle flow training (using 12,000 data samples);
• Phase2: Expansion to longitudinal data subsets;
• Phase3+: Integration of more cohort data and validation of downstream tasks; Detailed planning can be found in PLAN.md.

The challenges faced include:

• Data privacy and ethics: Strict protection of medical imaging data is required;
• Model generalization: Differences between scanners/populations affect performance;
• Reliability of causal inference, computational resource requirements, and complexity of clinical validation; Countermeasures: Responsible use terms (docs/MODEL_LICENSE.md), re-evaluation of cohort usage strategy after Phase1.

NeuroDrift integrates technologies such as continuous-time modeling and multimodal fusion, providing a powerful tool for neurodegenerative disease research. Its open-source commitment and innovations will provide experience for the field, and it is worth being closely followed by researchers interested in the intersection of AI and neuroscience.