• acquisition timing
• cardiac amyloidosis
• CNN
• dropout
• neural network
• pet
• Python
• PyTorch

Currently, PET imaging can diagnose cardiac amyloidosis (CA), with late PET images distinguishing AL – CA from ATTR and other conditions initially suspected as amyloidosis (CTRL). Ideally, we would like to identify subtypes of CA through early PET/CT images, as this would bring significant benefits for both the patient and the facility performing the exam. However, early images show no significant differences in cardiac uptake among the three groups. This thesis explores Deep Learning, specifically convolutional neural networks (CNNs), to detect CA in early PET images. The goal is to determine the optimal acquisition time for differential diagnosis using trained neural networks. We analyzed the performance of deterministic and dropout neural networks on a dataset of 61 patients, focusing on static images acquired at 5, 15, and 30 minutes after injecting [18F]–Florbetaben. We also explored the best classification approach: direct multi-class or cascaded binary classification (AL vs. NOT AL, ATTR vs. NOT ATTR). To mimic clinical practice, we assessed model performance at the patient level, using majority voting for classification. Dropout neural networks allow for classification that is more accurate by statistically discarding irrelevant slices. Through rigorous experimentation, we demonstrated Deep Learning's ability to extract information from static PET images at different times, contributing to a non-invasive diagnostic tool for suspected CA. The framework used is PyTorch.