An information-theoretic evaluation framework for CNN-LSTM-based Alzheimer's disease classification from structural...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Early detection of Alzheimer's disease (AD) is crucial for managing its progressive cognitive decline.
• Structural Magnetic Resonance Imaging (MRI) offers valuable insights for AD diagnosis.
• Existing classification models require robust evaluation methods beyond traditional accuracy metrics.

Purpose of the Study:

• To develop and evaluate a CNN-LSTM framework for three-class AD classification (Normal Control, Mild Cognitive Impairment, AD) using structural MRI.
• To introduce and assess a novel post-hoc information-theoretic evaluation strategy for AD classification models.
• To quantify model performance using metrics like Renyi mutual information, Renyi divergence, and Henze-Penrose divergence.

Main Methods:

• A Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) framework was designed for AD classification.
• Generative Adversarial Networks (GANs) were employed for data augmentation to address data scarcity.
• Models were evaluated using conventional metrics and advanced information-theoretic measures on 827 ADNI subjects.

Main Results:

• The CNN-LSTM model achieved 96.7% accuracy in subject-level AD classification, outperforming benchmark architectures.
• Information-theoretic measures provided complementary insights into model behavior, including information preservation and output distribution alignment.
• GAN-based augmentation improved training diversity without compromising validation and test data integrity.

Conclusions:

• The proposed CNN-LSTM framework demonstrates high accuracy for AD classification from structural MRI.
• Post-hoc information-theoretic analysis offers a more transparent and comprehensive method for evaluating classification models.
• Further external validation on multi-center datasets is necessary before clinical implementation.