Artificial Intelligence in the Analysis of PET Scans of the Human Brain

Abstract

Artificial intelligence (AI) is being intensely studied, evaluated, and applied in healthcare and especially in medical imaging. Having shown performance equaling that of experienced radiologists in tasks such as detecting pneumonia on chest X-rays and identifying cancerous long nodules on X-ray, AI is poised to radically optimize many areas of medical practice from early detection of disease to prediction of progression and personalization of therapeutic strategy. Artificial intelligence extends classical statistical techniques and machine learning, both of which characteristically involve manually establishing imaging features hypothesized to modulate a certain outcome. With AI, predictive features are automatically established in a data-driven fashion, which in turn implies that raw unprocessed data can be fully utilized, human bias can be avoided, and previously unrealized disease mechanisms potentially can be discovered. Here we discuss applications of AI in PET imaging for image reconstruction, attenuation correction without CT, dose reduction, automatic identification of pathology, and differentiation of disease progression. One of the costs of more automated analyses and better accuracy with AI compared to classical machine learning is larger volumes of training data; however, the field is rapidly evolving, and we discuss possible mitigations as well as other directions for valuable future applications of AI in PET imaging.