Comparison of Three Automated Approaches for Classification of Amyloid-PET Images

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Biomarker Discovery

Background:

• Automated amyloid-PET image classification enhances diagnostic confidence in clinical settings.
• Comparing automated methods is crucial for optimizing Alzheimer's disease diagnosis.

Purpose of the Study:

• To compare three automated amyloid-PET classification methods: ROC analysis, machine learning (ML), and deep learning (DL).
• To evaluate performance based on training data size, radiotracers, and patient cohorts.

Main Methods:

• Utilized 276 [¹¹C]PiB and 209 [¹⁸F]AV45 PET images from ADNI and local cohorts.
• Developed 68 ML models using regional SUVr values and one DL network.
• Compared global cut-points (ROC), regional ML, and 3D DL approaches.

Main Results:

• ML classification demonstrated high accuracy comparable to ROC, with superior convergence and reduced training data needs.
• Naïve Bayes was the top-performing ML algorithm; maximum SUVr cut-points improved accuracy over mean SUVr.
• DL networks excelled in definite cases but struggled with equivocal cases; rainbow-scale imaging improved agreement and accuracy.

Conclusions:

• Machine learning offers a robust alternative to ROC analysis for amyloid-PET classification, showing better generalization.
• Image scaling methods and algorithm choice significantly impact classification performance.
• Further ML advancements hold promise for more accurate Alzheimer's disease diagnosis.