Learning using privileged information improves neuroimaging-based CAD of Alzheimer's disease: a comparative study

The neuroimaging-based computer-aided diagnosis (CAD) for Alzheimer's disease (AD) has shown its effectiveness in recent years. In general, the multimodal neuroimaging-based CAD always outperforms the approaches based on a single modality. However, single-modal neuroimaging is more favored in clinical practice for diagnosis due to the limitations of imaging devices, especially in rural hospitals. Learning using privileged information (LUPI) is a new learning paradigm that adopts additional privileged information (PI) modality to help to train a more effective learning model during the training stage, but PI itself is not available in the testing stage. Since PI is generally related to the training samples, it is then transferred to the learned model. In this work, a LUPI-based CAD framework for AD is proposed. It can flexibly perform a classifier- or feature-level LUPI, in which the information is transferred from the additional PI modality to the diagnosis modality. A thorough comparison has been made among three classifier-level algorithms and five feature-level LUPI algorithms. The experimental results on the ADNI dataset show that all classifier-level and deep learning based feature-level LUPI algorithms can improve the performance of a single-modal neuroimaging-based CAD for AD by transferring PI. Graphical abstract Graphical abstract for the framework of the LUPI-based CAD for AD.