Domain specific transfer learning and classifier chains in Alzheimer's disease detection using 3D convolutional neural networks

This study examines different configurations of deep convolutional neural networks (CNNs) and the effect of using domain-specific transfer learning for distinguishing Alzheimer’s Disease and Mild Cognitive Impairment from normal controls. The data used to train our models was provided by ADNI and included 1,118 3D FDG-PET scans in total. We train a binary and a multiclass classifier, as well as chains of binary classifiers, for consecutive multiclass classification. Two chains were trained with different orders: chain A classified cognitively normal (CN) vs. non-CN, followed by Alzheimer’s disease (AD) vs. mild cognitive impairment (MCI). Classifier chain B classified AD vs. non-AD first, followed by MCI vs. CN. All classifiers were trained with and without the use of domain-specific transfer learning, using weights from Med3D. All models achieve comparable performance to the state-of-the-art. Classifier chain A even achieved superior performance with an accuracy of 96%, F1 score of 95% and AUROC of 99%. Using domain-specific transfer learning resulted in worse performance among the majority of the models, producing decreases in accuracy of up to 55%. These results show the potential of binary classifier chains and open some questions about the use of domain-specific transfer learning. • Binary classifiers outperform multiclass 3D CNNs in performance. • Chaining binary classifiers for multiclass scenarios improves performance. • The use of domain-specific transfer learning should be evaluated critically.