Evaluation of Convolutional Neural Network Methods for Alzheimer's Classification

Sammanfattning

Diagnostic errors in Alzheimer’s disease (AD) can be mitigated by using modern technologies like machine learning, specifically convolutional neural networks (CNNs). By analyzing MRI scans and we can assist in early detection. Given the rising prevalence of AD among the aging global population, it is critical to evaluate the efficacy of different CNN models in accurately classifying the severity of AD using MRI datasets. In this thesis, three CNN architectures were trained: ResNet152V2, DenseNet121, and EfficientNetB0 on two 2D MRI datasets to assess their performance across various metrics. The first dataset, from Kaggle, contained 6 400 images. EfficientNetB0 achieved the best performance, with test accuracies of 71%. The second dataset, from ADNI, consisted of 34,000 images and initially showed high accuracy (93-100%) across all classifiers. However, data leakage was discovered due to oversampling and applied data augmentation on overlapping train and test set, which compromised the results for the dataset. EfficientNetB0 consistently outperformed the other model architectures, but none were reliable for enough for clinical use. The study highlights AI’s potential in AD diagnosis and emphasizes the need for robust datasets and methodological improvements for future research and application.