Early Alzheimer's Disease Detection Using Functional MRI, Modified Census Transform, and Hierarchical Classification

Abstract

Alzheimer’s disease (AD) is the most prevalent neurodegenerative disorder worldwide, accounting for approximately 60–80% of dementia cases among older adults. Early detection of AD is particularly important because pathological brain changes may appear years or even decades before clinical dementia onset. This prodromal stage, commonly referred to as mild cognitive impairment (MCI), is marked by subtle memory and attention deficits, while patients largely retain functional independence. Consequently, identifying individuals at this stage may enable timely therapeutic interventions to delay disease progression and improve quality of life. Although several biomarkers have been proposed for AD diagnosis, including cerebrospinal fluid analysis and blood biomarkers, functional magnetic resonance imaging (fMRI) offers a promising noninvasive alternative for detecting disease-related functional alterations.

This study proposes a computer-aided framework for early detection of Alzheimer’s disease using fMRI data. The methodology integrates slice selection, feature extraction, and machine learning classification to identify the brain regions that best characterize disease progression. First, the ReliefF algorithm ranks and selects the most discriminative fMRI slices. Next, the Modified Census Transform (MCT) and a set of statistical descriptors extract features that capture local functional patterns associated with AD. These features are then used to train and evaluate two classification schemes: a multiclass classifier that discriminates among healthy controls, prodromal subjects, and patients diagnosed with Alzheimer’s dementia, and a hierarchical classifier composed of two consecutive binary stages.

Experimental results showed that the multiclass approach achieved classification accuracies ranging from 78.7% with a single slice to 89.5% when 16 selected slices were analyzed. To improve discrimination among clinically similar stages, we implemented a hierarchical framework: the first stage differentiates healthy subjects from those with Alzheimer’s-related pathology, and the second stage differentiates prodromal patients from those diagnosed with Alzheimer’s dementia. This hierarchical strategy significantly improved performance, achieving accuracies from 93.5% with a single slice to 99.6% with 16 selected slices.

These results demonstrate that combining ReliefF-based slice selection, Modified Census Transform feature extraction, and hierarchical classification yields an accurate, robust, and computationally efficient framework for early detection of Alzheimer’s disease. Furthermore, the findings indicate that a small number of carefully selected slices can achieve highly competitive performance, highlighting the proposed methodology's potential as a non-invasive biomarker and a computer-aided diagnostic tool for early Alzheimer’s disease assessment.