Abstract: Medical image classifcation using convolutional neural networks (CNNs) is promising but often requires extensive manual tuning for optimal model defnition. Neural architecture search (NAS) automates this process, reducing human intervention signifcantly. This study applies NAS to [18F]-Florbetaben PET cardiac images for classifying cardiac amyloidosis (CA) sub-types (amyloid light chain (AL) and transthyretin amyloid (ATTR)) and controls. Following data preprocessing and augmentation, an evolutionary cell-based NAS approach with a fxed network macro-structure is employed, automatically deriving cells’ micro-structure. The algorithm is executed fve times, evaluating 100 mutating architectures per run on an augmented dataset of 4048 images (originally 597), totaling 5000 architectures evaluated. The best network (NAS-Net) achieves 76.95% overall accuracy. K-fold analysis yields mean±SD percentages of sensitivity, specifcity, and accuracy on the test dataset: AL subjects (98.7±2.9, 99.3±1.1, 99.7±0.7), ATTR-CA subjects (93.3±7.8, 78.0±2.9, 70.9±3.7), and controls (35.8±14.6, 77.1±2.0, 96.7±4.4). NAS-derived network performance rivals manually determined networks in the literature while using fewer parameters, validating its automatic approach’s efcacy
Keywords: Neural architecture search · AutoML · Nuclear medicine · [18-F]-Florbetaben · Cardiac amyloidosis
File: https://link.springer.com/article/10.1007/s10278-024-01275-8