Novel multi-task learning for Alzheimer's stage classification using hippocampal MRI segmentation, feature fusion,...

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Alzheimer's disease (AD) diagnosis relies on clinical assessment and neuroimaging, but accurate staging of progression remains challenging.
• Hippocampal integrity on MRI is a key biomarker for AD, necessitating advanced analytical methods for precise classification.
• Existing methods often lack interpretability or struggle to integrate diverse data sources effectively.

Purpose of the Study:

• To develop and validate a comprehensive, interpretable framework for multi-class classification of AD progression stages.
• To integrate radiomic, deep learning, and clinical features from hippocampal MRI for enhanced diagnostic accuracy.
• To establish a robust and clinically actionable system for AD diagnostics.

Main Methods:

• A retrospective multi-center study analyzed 2956 patients across four AD stages using T1-weighted hippocampal MRI.
• Standardized segmentation (MedT) and feature extraction (radiomic, deep learning) were performed, followed by feature fusion, harmonization, and selection (LASSO).
• Classification employed machine learning models (XGBoost), with interpretability assessed via SHAP, nomogram, and decision curve analysis (DCA).

Main Results:

• MedT achieved superior hippocampal segmentation (Dice=92.03%).
• Fused features with XGBoost demonstrated the highest classification performance (accuracy=92.8%, AUC=94.2%).
• Key predictors included MMSE, hippocampal volume, and APOE ε4; the nomogram showed clinical utility via DCA.

Conclusions:

• Integrating radiomics, deep learning, and clinical data from hippocampal MRI enables accurate and interpretable AD stage classification.
• The proposed framework is robust, generalizable, and clinically actionable, offering a scalable solution for AD diagnostics.
• This approach enhances diagnostic capabilities for Alzheimer's disease progression.