Cross-cohort dementia identification using transfer learning with FDG-PET imaging

Abstract

Machine learning techniques have been extensively adapted to dementia identification with PET imaging in the last decade. Despite the promising results reported by these studies, the accurately labeled PET brain scans used to train machine learning models are generally difficult to obtain in real clinic environments. To tackle this challenge, we proposed a dementia classification method inspired by transfer learning. The main idea is to train a machine learning model using an accurately labeled source image cohort and an unlabeled target image cohort jointly, and then use this model to label the unlabeled target cohort. We demonstrated the effectiveness of the knowledge transfer in dementia classification tasks by comparing the proposed method to several other methods on public and private image cohorts.