Classification of amyloid status using machine learning with histograms of oriented 3D gradients

Area of Science:

• Neuroimaging
• Machine Learning
• Biomarker Discovery

Background:

• Quantitative assessment of brain amyloid burden is crucial for diagnosing neurodegenerative diseases.
• Standardized uptake value ratios (SUVRs) from PET imaging improve reliability but depend on tracers and regions of interest.
• Existing methods lack tracer independence and require specific region selection.

Purpose of the Study:

• To develop a novel machine learning approach for amyloid status classification that is independent of positron emission tomography (PET) tracers and regions of interest.
• To improve the accuracy and reduce classification uncertainty in brain amyloid detection.

Main Methods:

• A machine learning model was developed using feature vectors extracted from 3D gradient orientation histograms of amyloid PET images.
• The method was optimized on 18F-florbetapir datasets and validated on separate 18F-florbetapir, 11C-PiB, and 18F-florbetaben datasets.
• Performance was compared against SUVRs and a voxel intensity-based machine learning method.

Main Results:

• The proposed method demonstrated the largest mean distances between subjects and the classification boundary, indicating higher confidence classifications.
• It achieved the highest classification accuracy across all three tracers (18F-florbetapir, 11C-PiB, 18F-florbetaben), consistently exceeding 96%.
• The approach proved independent of tracer type and specific regions of interest.

Conclusions:

• The developed machine learning method offers a robust and accurate approach for brain amyloid status classification from PET imaging.
• This tracer-independent method enhances diagnostic reliability and reduces classification ambiguity in amyloid detection.
• The findings suggest a significant advancement in quantitative amyloid imaging analysis for neurodegenerative disease research.