Explainable AI Fuses MRI and Texture Features to Stage Alzheimer’s Disease

Researchers have developed an explainable deep learning model that combines structural MRI scans with handcrafted texture descriptors to stage Alzheimer’s disease, reporting accuracy above 99 percent on two benchmark datasets. The work, published in Applied Intelligence, addresses a persistent problem in multimodal medical AI: fusing heterogeneous data sources while keeping the model’s reasoning interpretable for clinicians.

The system, called an Explainable-Fusion Transformer, integrates convolutional vision transformers processing MRI images with features extracted using established texture descriptors—local binary patterns (LBP), histograms of oriented gradients (HOG), and local phase quantization (LPQ). A novel attention-based fusion mechanism merges these two streams, allowing the model to weigh information from each modality when assigning disease stages.

A central contribution of the study is a systematic comparison of attention designs. The team evaluated four cross-attention variants—standard cross-attention, Linformer, Performer, and MERS—against a self-attention baseline, using consistent benchmarking across both datasets. The choice of attention mechanism mattered: the model achieved its best result of 99.06 percent accuracy on a community-curated Hugging Face cohort using cross-attention, while on the clinically standardized ADNI repository it reached 99.69 percent accuracy with the Performer variant.

Interpretability was validated with two complementary techniques tailored to each branch of the model. Grad-CAM was applied to the MRI branch, and Integrated Gradients to the histogram branch. The resulting attributions highlighted temporal and ventricular brain regions—areas biologically consistent with Alzheimer’s disease progression—and produced coherent saliency over the textural features, which the authors say suggests the two modalities carry complementary information.

The authors frame the modality-specific interpretability of the design as a principled starting point for an interpretable screening and triage decision-support tool, one that could enable rapid case prioritization for specialist review in heterogeneous, multi-site clinical settings. However, they are explicit that these applications remain contingent on site-specific external validation and calibration, and the abstract notes the findings are pending such validation. The reported accuracies come from benchmark evaluations on the two named cohorts rather than from deployed clinical use, so performance in real-world, prospectively collected data has not yet been established.

Subject of Research: Explainable multimodal deep learning for Alzheimer's disease staging using MRI and texture features

Article Title: Explainable cross-attention fusion with convolutional vision transformers for multi-view Alzheimer’s disease staging

Article References: Holguin-Garcia, S. A., Cardona-Pérez, L. M., Guevara-Navarro, E., Garcia-Herrera, S., Delgado-Martínez, M. C., Ocampo Osorio, F., Ramirez Rodriguez, F. A., Isaza, G., Ruz, G. A., Tabares-Soto, R., & Bravo-Ortiz, M. A. (2026). Explainable cross-attention fusion with convolutional vision transformers for multi-view Alzheimer’s disease staging. Applied Intelligence, 56(16), Article 530. https://doi.org/10.1007/s10489-026-07444-2

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07444-2

Keywords: Alzheimer's disease, explainable AI, vision transformers, cross-attention, MRI, multimodal fusion, Grad-CAM, Integrated Gradients, ADNI, deep learning, medical imaging, disease staging