Optimal transport- and kernel-based early detection of mild cognitive impairment patients based on magnetic resonance and positron emission tomography images

Ziyu Liu, Travis S. Johnson, Wei Shao, Min Zhang, Jie Zhang, Kun HuangAlzheimer 2022 · DOI

Abstract

Our method is the current state of the art based on benchmark comparisons. This method is a necessary technological stepping stone to widespread clinical usage of MRI-based early detection of AD.

Introduction

Mild Cognitive Impairment (MCI) represents the transitional phase between normal cognitive aging and Alzheimer’s Disease (AD). Distinguishing between Early MCI (E-MCI) and Late MCI (L-MCI) is critical for timely clinical intervention. However, this task is challenging due to subtle physiological differences between these stages and the "small $n$, large $p$" problem inherent in neuroimaging datasets, where high-dimensional feature spaces (voxels or regions) are paired with limited sample sizes.

Methodology

The researchers developed a sophisticated three-stage pipeline to address these challenges, integrating multi-modal imaging data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).

Overview of the proposed machine learning framework for MCI detection.

Feature Selection via rMLTFL

To identify disease-essential biomarkers, the authors employed a robust multi-label transfer feature learning (rMLTFL) framework. This method identifies features from MRI and PET that are relevant across multiple related tasks, such as AD vs. Normal Control (NC), to ensure the model focuses on stable signals rather than noise.

Illustration of the robust multi-label transfer feature learning (rMLTFL) decomposition.

Domain Adaptation with Optimal Transport

The core innovation lies in using Optimal Transport (OT) for transfer learning. Instead of training solely on MCI data, the model leverages the clearer decision boundaries of the AD vs. NC task. Using Sinkhorn Distance, the "blurry" MCI samples are mapped into the "high-contrast" AD/NC feature space. This "reverse" mapping strategy makes the subtle differences between E-MCI and L-MCI more apparent to the classifier.

Bootstrap Aggregation (BAg)

To enhance stability and prevent overfitting, the framework utilizes bootstrap aggregation. Multiple sub-models are trained on different data subsets, with final predictions determined through a majority voting strategy.

Results and Discussion

The proposed method achieved state-of-the-art performance, reaching an accuracy of 75.00% and a peak AUC of 0.78 using FDG-PET features. It outperformed over 14 benchmark methods, including standard SVMs and complex transfer learning algorithms like TrAdaBoost.

Key insights include:

  • Geometry Matters: OT captures the underlying geometric structure of data distributions more effectively than simple feature alignment, "warping" the MCI distribution to align with AD/NC landmarks.
  • Multi-modal Synergy: While FDG-PET often provided higher individual accuracy, the combination of PET and VBM-MRI features, guided by smart feature selection, offered the best stability.

Conclusion

This framework demonstrates that leveraging well-defined source domains through Optimal Transport can significantly improve the detection of subtle disease stages. Future work will explore incorporating genotype data and testing the model's ability to predict conversion from stable to progressive MCI.

Summary updated

Abstract

Our method is the current state of the art based on benchmark comparisons. This method is a necessary technological stepping stone to widespread clinical usage of MRI-based early detection of AD.
BibTeX
@inproceedings{liu2022optimal,
  author = {Ziyu Liu and Travis S. Johnson and Wei Shao and Min Zhang and Jie Zhang and Kun Huang},
  title = {{Optimal transport- and kernel-based early detection of mild cognitive impairment patients based on magnetic resonance and positron emission tomography images}},
  booktitle = {Alzheimer},
  year = {2022},
  doi = {10.1186/s13195-021-00915-3},
}
Figure