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

Early detection of Alzheimer's disease (AD) is important because of its progressive impact on cognitive function. This study presents a CNN-LSTM-based framework for three-class AD classification from structural MRI, with the primary contribution being a post-hoc information-theoretic evaluation strategy rather than a new network architecture. Experiments were conducted using 827 ADNI subjects, including normal controls (NC: 340), mild cognitive impairment (MCI: 307), and AD (180). To mitigate data scarcity and improve training diversity, GAN-based augmentation was applied only to the training data, while validation and test subjects were kept separate. In addition to conventional metrics, trained models were evaluated using Renyi mutual information, Renyi divergence, and Henze-Penrose divergence to quantify information preservation, representation stability, and distributional alignment. Under a subject-level evaluation protocol, the CNN-LSTM model achieved 96.7% accuracy and outperformed evaluated benchmark architectures under the same protocol. The information-theoretic measures provided complementary evidence for comparing model behavior beyond accuracy, particularly regarding information retention and output-distribution alignment. Overall, the findings suggest that post-hoc information-theoretic analysis can support more transparent assessment of MRI-based AD classification models. However, external validation on independent multi-center datasets is required before clinical deployment can be considered.