Development of a serverless,interactive application for Alzheimer’s disease detection and visualization using MRI images

Abstract

Background

Alzheimer’s disease requires early detection for effective intervention with disease-modifying treatments, yet significant implementation barriers persist in clinical practice, including limited computational infrastructure and the gap between research model performance and practical deployment in resource-constrained settings.

Objectives

To develop and evaluate a computer-aided diagnosis system for classifying cognitive states (AD, MCI, and CN) from structural MRI, with automated preprocessing and cost-effective cloud deployment suitable for resource-constrained healthcare facilities.

Design

Computer-aided diagnosis system integrating neural architecture search with serverless cloud infrastructure.

Methods

A multi-view MRI analysis model was optimized through neural architecture search, incorporating Universal Inverted Bottleneck blocks and Kolmogorov–Arnold Networks. Automated MRI preprocessing using FSL was deployed through cloud-based serverless functions for scalable image processing. Evaluation used the ADNI dataset (1687 individuals: 368 AD, 625 MCI, and 694 CN). A web application was developed providing patient management, MRI visualization, and automated diagnostic prediction.

Results

The model achieved 86.7% accuracy and 0.900 AUC in three-class classification with 1.7 million parameters. High specificity was observed across all classes (CN: 91.0%, MCI: 91.8%, AD: 97.3%), with 100% CN-AD specificity ensuring no AD cases were misclassified as cognitively normal. Operational costs were approximately 0.028 USD per diagnosis for typical hospital workloads.

Conclusion

This system provides a cost-effective approach for early Alzheimer’s diagnosis accessible to resource-constrained environments. Despite challenges in MCI classification, the combination of neural architecture search with serverless deployment demonstrates progress toward clinically deployable automated AD detection. Future work should focus on prospective clinical validation and integration of interpretability features.