Abstract
Magnetic resonance (MR) imaging is a powerful technique for non-invasive in-vivo imaging of the human brain. We employed a recently validated method for robust cross-sectional and longitudinal segmentation of MR brain images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort. Specifically, we segmented 5074 MR brain images into 138 anatomical regions and extracted time-point specific structural volumes and volume change during follow-up intervals of 12 or 24 months. We assessed the extracted biomarkers by determining their power to predict diagnostic classification and by comparing atrophy rates to published meta-studies. The approach enables comprehensive analysis of structural changes within the whole brain. The discriminative power of individual biomarkers (volumes/atrophy rates) is on par with results published by other groups. We publish all quality-checked brain masks, structural segmentations, and extracted biomarkers along with this article. We further share the methodology for brain extraction (pincram) and segmentation (MALPEM, MALPEM4D) as open source projects with the community. The identified biomarkers hold great potential for deeper analysis, and the validated methodology can readily be applied to other imaging cohorts.
Introduction
Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI) are characterized by progressive neurodegeneration, yet identifying these changes early and objectively remains a clinical challenge. While automated tools can extract "biomarkers" from MRI scans, many high-performing models lack interpretability or suffer from poor reproducibility due to proprietary methods and small datasets.
This study addresses these hurdles by processing 5,074 MR images from 1,069 subjects in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort. The authors provide a massive, quality-checked, and interpretable dataset of brain structures, alongside the open-source tools used to create it, bridging the gap between complex machine learning and clinical utility.
Methodology: Robust Whole-Brain Segmentation
The researchers employed a pipeline focused on automated whole-brain segmentation using several key tools:
- Brain Extraction (
pincram): Isolates brain tissue from the skull. - Cross-Sectional Segmentation (MALPEM): Uses "multi-atlas label propagation" to divide the brain into 138 distinct anatomical regions based on 30 expert-labeled maps.
- Longitudinal Segmentation (MALPEM4D): Tracks changes over 12 and 24 months while ensuring temporal consistency to prevent noise from being mistaken for atrophy.
To ensure results were due to disease rather than natural demographics, the authors used multiple linear regression to correct volumes for age, gender, and head size.
Key Insights and Results
The study demonstrates that biologically meaningful features, such as the volume of specific regions, can perform as well as highly complex "black-box" features.
Discriminative Biomarkers
While the hippocampus is a traditional focus of AD research, this study found that the amygdala was the most discriminative individual structure for distinguishing AD from healthy controls (80% accuracy). Additionally, the enlargement of the ventricles (fluid-filled spaces) proved to be a more sensitive marker for disease progression than the shrinking of brain tissue in several classification tasks.
Longitudinal Analysis and MCI Conversion
Predicting which MCI patients will progress to AD is notoriously difficult. The authors found that while baseline volumes provide 68% accuracy, incorporating the rate of atrophy over 24 months significantly boosts accuracy to 78%.
Classification Performance
- AD vs. Healthy Controls: Achieved a balanced accuracy of 89% using all structural volumes in a Support Vector Machine (SVM) classifier.
- Atrophy Rates: Confirmed that AD patients lose hippocampal volume at a rate of ~4.8% per year, compared to only ~1.1% in healthy elderly subjects.
Conclusion and Future Directions
The release of 5,074 quality-checked segmentations provides a "gold standard" for the research community, particularly for training data-hungry deep learning models. While the study relies on clinical diagnoses rather than histopathological ground truth, the validated methodology and shared morphometry database offer a powerful resource for future neurodegeneration research.
Summary updated