Deriving Physiological Information from PET Images Using Machine Learning

Abstract Machine learning (ML) algorithms have become popular in recent years and found increasing utility the field of medical imaging, specifically positron emission tomography (PET) imaging. The interest ML PET imaging for study neurodegenerative diseases stems from potential these techniques to analyze predict physiological parameters biomarkers such as total volume distribution (V $$_{\text {t}}$$ t ) organ or a structure be explored. In this paper, we investigated whether V [ $$^{18}$$ 18 F]-FEPPA radiotracer, an indicator neuroinflammation, could estimated directly non-invasive way, given activity radiotracer brain tissue. used several regression models different regions where 31 were defined each 24 patients with Parkinson disease 20 healthy subjects, train four tree-based models. predicted reference values compared by Bland-Altman analysis model’s performance was evaluated mean absolute error (MAE). best result obtained XGBoost model MAE 2.6. results indicate that are average very close bias 0.23 "Image missing" 2.82. Significant main effect genotype on both caudate putamen been preserved Vt (p < 0.05). paired t-test difference between is not statistically significant 6 out 8 groups. proposed provide efficient tool values, hallmark neuroinflammation believed trigger Parkinson’s development.