• deep learning
• neural architecture search
• cardiac amyloidosis
• AutoML
• evolutionary algorithms
• CNN
• Amyloid light chain (AL)
• Amyloid transthyretin (ATTR)

Designing and finding a suitable neural network can be a challenging and time-consuming task. Often modifications to existing architectures must be made, relying on the experience and knowledge of the researcher. In recent years, Neural Architecture Search (NAS) has been developed to automate the search for the best performing neural network, thus reducing human intervention. However, no attempt has yet been made to apply this technique to the classification of nuclear medicine images. The present work attempts to fill this gap by implementing an Evolutionary Algorithm.
First, the algorithm was validated using an online dataset of medical images to compare the performance of the best discovered network with that of the manually created one. This dataset consists of paediatric chest X-rays of normal and pneumonia subjects.
Then, a set of cardiac amyloidosis images (consisting into 3D static PET acquisition 15 minutes after the injection of the [18-F]florbetaben) was used. The gold standard for diagnosis of this disease is cardiac biopsy, a risky approach related to its invasiveness; therefore, researchers are still trying to find alternative and less invasive methods for the diagnosis of this pathology. Based on the results, PET imaging can be considered promising as it allows early diagnosis and differentiation between different forms of amyloidosis (e.g., AL, ATTR) to make therapy more efficient.