RoRankNet: Modeling Ordinal Progression in Alzheimer’s Disease with Plackett-Luce Ranking and Hard Instance Learning

Alzheimer’s disease (AD) is a chronic neurodegenerative disorder characterized by progression from normal control (NC) to mild cognitive impairment (MCI), and ultimately to AD. These stages reflect ordinal deterioration of specific brain tissues. Timely and accurately identifying these stages is crucial for effective treatment. While existing multi-classification methods using structural magnetic resonance imaging show promise, they often overlook the inherently ordinal progressive relations. Addressing this, our study introduces RoRankNet, a robust ranking network designed to model the ordinal progression of AD. Specifically, we propose a distance-based Plackett-Luce ranking module to effectively represent the distance relations of these stages, enhancing the alignment between extracted features and disease progression. Additionally, we present a hard instance rectification module to identify and correct challenging instances, enhancing classification robustness. Our extensive experiments demonstrate RoRankNet’s superiority in classifying NC, MCI, and AD. Notably, our approach not only preserves the ordinal nature in feature space but also excels in localizing discriminative biomarkers essential for accurate stage prediction in clinical practice. Source codes will be publicly available soon.