[MICCAI/arXiv] Whole MILC: Generalizing Learned Brain Dynamics Across Datasets and Disorders

TL;DR

Whole MILC is a novel self-supervised framework that learns the "language" of brain dynamics from large-scale healthy fMRI data (HCP) to solve clinical classification tasks with very few samples. By maximizing mutual information between local time windows and global brain states, it achieves SOTA performance in identifying Schizophrenia, Autism, and Alzheimer’s, even when pre-trained on an entirely different population.

Problem & Motivation: The Data Scarcity Trap

In neuroimaging, we face a fundamental paradox: fMRI captures incredibly rich spatio-temporal data ($m$), but clinical datasets for specific disorders typically involve only a few dozen subjects ($n$). This $m\gg n$ scenario leads to massive overfitting for standard Deep Learning models.

Traditionally, researchers bypass this by "handcrafting" features—for example, reducing a 4D fMRI scan into a static "Functional Network Connectivity" (FNC) matrix. While useful, this approach discards the dynamics—the very temporal signatures that might hold the key to early diagnosis. The authors of Whole MILC ask: Can we learn the essence of brain dynamics from unlabeled data and transfer that knowledge to tiny clinical datasets?

Methodology: Mutual Information as a Teacher

The "Whole MILC" (Mutual Information Local to Context) approach treats the relationship between a part (a time window) and the whole (the entire sequence) as the supervisory signal.

1. The Architecture

The model consists of:

• CNN Encoder: Compresses a local window of ICA time courses into a latent vector $z_t$.
• BiLSTM & Attention: Aggregates the entire sequence into a global context vector $c$.
• Separable Critic: A function $f(z,c)$ that measures how well a local window fits a global sequence.

2. The Learning Objective

The model uses the InfoNCE loss. For a given brain sequence, it tries to make the local window $z_t$ highly predictable of its own global context $c$, while distinguishing it from the contexts of other subjects.

Figure 1: The MILC architecture showing the local path (CNN) and global path (BiLSTM + Attention) feeding into the Mutual Information objective.

By training on the Human Connectome Project (HCP) dataset (healthy adults), the encoder learns the fundamental "physics" of how brain networks fluctuate over time.

Experiments: Crossing the "Generalization Gap"

The most impressive aspect of this work is its zero-shot-to-fine-tuning capability. The model was pre-trained on healthy adults but tested on:

1. Schizophrenia (FBIRN & COBRE datasets)
2. Autism (ABIDE - Children/Adolescents)
3. Alzheimer’s (OASIS - Elderly)

SOTA Comparison

As shown in the results below, the pre-trained MILC models (FPT - Frozen, UFPT - Unfrozen) consistently dominate the Not Pre-trained (NPT) baselines, especially when the training size is extremely small (e.g., only 15-30 subjects).

Figure 2: Performance (AUC) across different disorders. Note the significant gap between MILC-based models and the NPT baseline at low sample sizes.

Why it Works: Model Introspection

Beyond classification, Whole MILC allows for Saliency Mapping. By looking at which spatio-temporal regions the attention mechanism focuses on, researchers can pinpoint specific ICA components and time points that are most "suspicious" for a diagnosis, providing a path toward interpretable AI in medicine.

Critical Analysis & Conclusion

Whole MILC proves that unlabeled healthy data is a goldmine for clinical AI. The fact that a model trained on healthy 20-year-olds can improve Alzheimer's detection in 70-year-olds suggests that the underlying dynamics of brain networks share a universal structure.

Limitations:

• The model currently relies on ICA (Independent Component Analysis) as a preprocessing step. While effective, this still imposes a linear constraint on the spatial features.
• Future work needs to explore "End-to-End" dynamics directly from raw 4D fMRI voxels.

Takeaway: If you are working with high-dimensional time-series and small labels, don't train from scratch. Use mutual information to learn a global-to-local representation on your largest available unlabeled pool first.