Boosting the diagnostic power of amyloid-β PET using a data-driven spatially informed classifier for decision support

Ashwin Venkataraman, Wenjia Bai, Alex Whittington, James F. Myers, Eugenii A. Rabiner, Anne Lingford‐Hughes, Paul M. Matthews, for the Alzheimer’s Disease Neuroimaging InitiativeAlzheimer 2021 · DOI

Abstract

The diagnostic classification accuracy of amyloid PET was improved using an automated data-driven spatial classifier. Our classifier highlights the importance of considering the spatial variation in Aβ PET signal for optimal interpretation of scans. The algorithm now is available to be evaluated prospectively as a tool for automated clinical decision support in research settings.

The Challenge of Amyloid PET Interpretation

Amyloid-β (Aβ) plaques are a primary pathological hallmark of Alzheimer’s Disease (AD). While Positron Emission Tomography (PET) allows clinicians to visualize these plaques in living patients, interpreting the results is often complex. Current standards rely on either subjective visual reads or a "composite SUVr"—a single numerical average of amyloid density across the entire brain.

The limitation of a global average is that it can wash out the nuances of where amyloid is depositing. In early-stage patients, significant accumulation in specific regions may be "diluted" by low signal elsewhere, leading to false negatives. This research introduces a spatially informed model designed to capture the unique spatial "fingerprint" of amyloid deposition to boost diagnostic sensitivity.

A Data-Driven Spatial Approach

Researchers moved away from pre-defined anatomical atlases, opting instead for a data-driven methodology using voxel-wise clustering. By analyzing hundreds of scans from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), they identified four distinct spatial clusters ($k=4$) that best represent the progression of amyloid pathology.

Workflow for training the spatially informed classifier and its application in automated decision support.

Instead of a single global value, the classifier utilizes six biologically grounded features: the volume of positive voxels in each of the four identified clusters, the total volume of positive voxels in the grey matter, and the mean grey matter SUVr. These features are processed by an optimized Quadratic Support Vector Machine (SVM) to provide automated decision support.

Key Results and Performance Gains

The spatially informed classifier significantly outperformed industry standards across several metrics:

  • Diagnostic Accuracy: Achieved 81% accuracy for classifying AD vs. non-AD using PET data alone.
  • Area Under the Curve (AUC): Reached 0.91, a substantial improvement over the 0.82 achieved by standard composite SUVr methods.
  • Sensitivity: The model proved 15% more sensitive at detecting AD than the composite SUVr and 39% more sensitive than simple voxel counts.

One of the most critical insights from this study is that the spread (volume) of amyloid clusters is often a more sensitive marker of disease progression than the absolute intensity of the PET signal. Specifically, the researchers identified a "Cluster 4" (involving prefrontal and occipital regions) that acts as an inflexion point, potentially marking the transition from mild impairment to clinical Alzheimer’s.

Clinical Significance and Future Directions

Unlike "black-box" deep learning models, this SVM-based approach is explainable, using just six features that are easy for clinicians to interpret. This transparency makes it a viable candidate for integration into clinical workflows and clinical trials.

While the study currently relies on cross-sectional data from the ADNI cohort, future work will focus on longitudinal validation to confirm these spatial trajectories over time. The algorithm has been made available for further research evaluation, offering a promising tool for identifying the right patients earlier and more accurately measuring the efficacy of amyloid-clearing therapies.

Summary updated

Abstract

The diagnostic classification accuracy of amyloid PET was improved using an automated data-driven spatial classifier. Our classifier highlights the importance of considering the spatial variation in Aβ PET signal for optimal interpretation of scans. The algorithm now is available to be evaluated prospectively as a tool for automated clinical decision support in research settings.
BibTeX
@inproceedings{venkataraman2021boosting,
  author = {Ashwin Venkataraman and Wenjia Bai and Alex Whittington and James F. Myers and Eugenii A. Rabiner and Anne Lingford‐Hughes and Paul M. Matthews and for the Alzheimer’s Disease Neuroimaging Initiative},
  title = {{Boosting the diagnostic power of amyloid-β PET using a data-driven spatially informed classifier for decision support}},
  booktitle = {Alzheimer},
  year = {2021},
  doi = {10.1186/s13195-021-00910-8},
}
Figure