Automated Neural Architecture Search for Cardiac Amyloidosis Classification from [18F]-Florbetaben PET Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Convolutional Neural Networks (CNNs) show promise for medical image classification but require extensive manual tuning.
• Automating CNN model definition through Neural Architecture Search (NAS) can significantly reduce human intervention.

Purpose of the Study:

• To apply NAS to [18F]-Florbetaben Positron Emission Tomography (PET) cardiac images for classifying cardiac amyloidosis (CA) subtypes.
• To automatically derive optimal CNN architectures for CA classification, reducing manual effort.

Main Methods:

• An evolutionary cell-based NAS approach with a fixed macro-structure was employed.
• The NAS algorithm explored 5000 architectures on an augmented dataset of 4048 cardiac PET images.
• The process involved data preprocessing, augmentation, and five independent execution runs.

Main Results:

• The best NAS-derived network (NAS-Net) achieved 76.95% overall accuracy.
• K-fold analysis demonstrated high sensitivity and specificity for AL amyloidosis and controls, with varying results for ATTR-CA.
• NAS-derived networks performed comparably to manually designed networks while utilizing fewer parameters.

Conclusions:

• NAS is an effective automated approach for developing CNNs for cardiac amyloidosis classification using PET imaging.
• The study validates the efficacy of NAS in optimizing deep learning models for complex medical image analysis tasks.
• The NAS-derived model offers a competitive alternative to manually engineered networks, with potential for improved efficiency.