NeuroXAI-Caps: an explainable CNN–capsule network for early Alzheimer’s diagnosis

Document Type

Article

Source of Publication

Neural Computing and Applications

Publication Date

4-10-2026

Abstract

Alzheimer’s disease is the primary and irreversible neurodegenerative disorder, affecting more than 55 million people, and yet to date, there is no effective early diagnostic solution. Though conventional CNNs have proven to be efficient in feature extraction, they inherently lose crucial spatial hierarchies via pooling operations, which reduces their sensitivity to subtle neuroanatomical deviations indicative of early Alzheimer’s disease. This paper introduces NeuroXAI-Caps, an innovative explainable hybrid deep learning paradigm that effectively synergizes CNNs with Capsule Network (CapsNet) in order to offer pose-equivariant vector encoding of features from T1-weighted MRI and thus enable fine-grained stage classification. Using a benchmark MRI dataset comprising 11,279 images across four cognitive stages, the proposed model showed superior predictive capability: 96% training accuracy, 93% validation accuracy, and 90% test accuracy, along with consistently high precision, recall, and F1-scores. Through 5-fold Stratified Cross-Validation, NeuroXAI-Caps achieved mean accuracy = 0.99, F1 = 0.97, Cohen’s κ = 0.997, and AUROC = 0.995, validating its robust generalization and reliability. External evaluation on the OASIS dataset yielded 99.69% accuracy and 100% sensitivity across all impairment stages, thus underpinning clinical scalability. For transparency, Grad-CAM and LIME were embedded to visualize discriminative regions (hippocampal and ventricular structures) that strengthen neuro-anatomical validity and interpretability. This lightweight architecture comprising 4.08 M parameters with 9.6 GFLOPs achieved an average inference latency of 0.73 ms and LAAI = 1.23, thus confirming real-time applicability. Taken together, NeuroXAI-Caps bridges the long-standing gap among diagnostic accuracy, interpretability, and clinical deployability. This approach yields a cost-effective, non-invasive, along explainable framework for early Alzheimer’s screening and precision neurodiagnostics.

ISSN

0941-0643

Publisher

Springer Science and Business Media LLC

Volume

38

Issue

8

Disciplines

Computer Sciences

Keywords

Alzheimer’s Disease, Artificial Intelligence in Healthcare, Capsule Networks (CapsNet), CNN, Early Diagnosis, Explainable AI (XAI)

Scopus ID

105036594176

Indexed in Scopus

yes

Open Access

no

Share

COinS