Early identification was essential in efficient the management of Alzheimer's disease (AD), which is one of the biggest causes of dementia. Conventional Medical Imaging (CMI) methods such as Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) scans are used to understand various aspects of health. However, they lack the ability to capture the dynamic progression of Alzheimer's disease without incorporating sequential data series. This research introduces a novel method that has been designed to address these limitations. By combining spatial-temporal analysis with dynamic sequence processing, the work presents the "Attention Enhanced CNN+LSTM" architecture. The proposed dual pipeline architecture combinesLSTeM (Long ShortTerm enhanced Memory) for LSTM processing with an Attention ResNet for CNN processing. This work designs a novel "CS-Attention Block" with "Aggregate Weighted Pooling" for the enhancement of ResNet-50 and integration of LSTeM; further, a QKV (Query-Key-Value) attention mechanism for feature fusion is also proposed. For experiments, the Alzheimer’s disease Neuroimaging Initiative (ADNI) dataset containing longitudinal images provides the MRI and PET scan image data used for the analysis of the model. Results: The proposed model showed better performance for early AD 12 months before clinical diagnosis. The model also achieved an accuracy of 0.9151, and the AUC reached a value of 0.9784. This result was further improved to be improved by0.9302 for the metric accuracy and 0.9913 for the metric AUC ahead of the 6-month prediction. The model fuses spatial-temporal features and processes the sequences to predict AD by addressing the limitations of existing models.

Alzheimer's Disease Deep Learning Accuracy AUC LSTeM ResNet