Evaluación de un modelo de aprendizaje automático basado en redes neuronales para el diagnóstico temprano de la enfermedad de Alzheimer

Abstract

Alzheimer's disease, as the leading cause of dementia, poses a critical challenge in its early detection, where timely interventions can delay its progression. This issue is addressed by evaluating deep learning models based on neural networks to classify magnetic resonance imaging (MRI) scans into different stages of the disease using data from the OASIS and ADNI databases. The importance lies in the need for precise and automated tools to improve early diagnosis, particularly in stages such as mild cognitive impairment. The methodology included constructing a dataset from preprocessed images and applying it to the EfficientNet B7 and ResNet50 architectures. These were trained with advanced techniques such as data augmentation and validated across scenarios involving controlled data, modified images, and new data. Performance metrics such as precision, recall, specificity, F1-score, and ROC-AUC curves were analyzed. The results showed that the EfficientNet B7 architecture outperformed ResNet50 in precision, sensitivity, and specificity, especially in classifying early stages of Alzheimer's disease. EfficientNet B7 demonstrated greater generalization ability, achieving high precision with preprocessed and new images, while ResNet50 showed limitations when working with heterogeneous data. This highlights the importance of modern architectures in solving complex problems like early Alzheimer’s detection, although it also evidences that relying solely on controlled datasets like OASIS and ADNI may limit applicability in real clinical scenarios.