Optimized deep learning-based multi-stage Alzheimer’s disease characterization with temporal modeling and soft-NMS for neuroimaging

Alzheimer’s disease (AD) represents an important and growing global health challenge, characterized by progressive neurodegeneration and cognitive decline. Early and accurate diagnosis, especially distinguishing between different stages such as cognitively normal (CN), mild cognitive impairment (MCI), and AD, is crucial for timely intervention, patient management, and the development of effective therapies. However, current diagnostic paradigms, relying on clinical assessments, neuropsychological tests, and conventional neuroimaging analysis, often suffer from limitations including subjectivity, high cost, limited accessibility, and insufficient sensitivity to subtle changes in the early stages or during progression. Furthermore, many existing automated approaches focus on binary classification (AD vs. CN), neglecting the critical intermediate MCI stage and the temporal dynamics of disease progression. This paper introduces a novel, automated, multi-stage AD classification framework leveraging enhanced deep learning techniques applied to structural Magnetic Resonance Imaging (MRI) data. Our integrated approach addresses the shortcomings of previous methods by combining an improved ResNet50 architecture for enhanced feature extraction, a Bidirectional Gated Recurrent Unit (Bi-GRU) to model temporal disease progression from sequential MRI scans, and a faster R-CNN framework with Soft Non-Maximum Suppression (Soft-NMS). This final component provides refined localization of AD-related neuropathological changes by learning to identify these pathological regions implicitly, without requiring any manual annotations, thereby mitigating issues with overlapping region proposals. Comprehensive experimental validation was performed on the widely used Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, using 25,024 structural MRI scans from 6,100 unique participants encompassing subjects with CN, MCI, and AD. The proposed framework demonstrated exceptional performance, achieving state-of-the-art accuracy in distinguishing AD from CN at 98.2%, AD from MCI at 95.3%, and MCI from CN at 93.5%. In particular, it also achieved high accuracy (87.9%) in the challenging three-way multiclass classification task (AD vs. MCI vs. CN). Ablation studies confirmed the significant contribution of each integrated component. Furthermore, the inclusion of Grad-CAM visualizations enhances model interpretability by highlighting brain regions critical to classification decisions.