An EfficientNet-Based Method for Interpretable Early Detection of Alzheimer

Alzheimer's disease is a degenerative impairment that unhurriedly impairs remembrance, cognition, and behavior, ultimately leading to a loss of independence. This study focuses on detecting Alzheimer's across four categories, NonDementia, VeryMildDementia, MildDementia, and ModerateDementia that represent different levels of cognitive decline. Accurate early prediction is essential, as it enables timely intervention, which may help slow or even halt symptom progression, thereby improving patients' quality of life. Applying machine learning and deep learning in Alzheimer's detection presents a cost-effective solution, reducing the reliance on manual analysis and providing a scalable, efficient diagnostic tool. However, In prediction models, we still face a lack of interpretability, as healthcare providers must understand and trust the results. Previous research has utilized algorithms like Support Vector Machines, Random Forest, k-Nearest Neighbors for Alzheimer's classification; nonetheless, these models often lack transparency. This research makes use of the Alzheimer's Disease Neuroimaging Initiative or ADNI dataset, applying EfficientNet for feature extraction owing to its superior precision and efficiency in computation. A Dense layers is used as the classifier, chosen for its seamless integration within the neural network framework and compatibility with interpretability methods. GradCAM++ is applied to enhance interpretability, highlighting specific brain regions that influence each decision. Through our design and applied methods, this research achieves a high classification accuracy of 97.84%, reflecting an improvement of nearly 2% over previous approaches. This approach shows promise in supporting healthcare professionals with early diagnosis and the planning of targeted interventions.