Evaluating different training techniques for a convolutional neural network that classifies Alzheimer’s disease

Sammanfattning: Effective computer diagnosis of Alzheimer’s disease could bring large benefitsto the millions of people worldwide who does or will suffer from dementia. One popular method for trying to achieve this is the training of convolutional neural networks to classify MRI brain scans. An abundance of training methods exists that aims to improve the performance of these neural networks, but their effectiveness and eventual disadvantages are not always clear. This study evaluates the performance of a CNN on the OASIS-3 neuroimaging dataset. The CNN’s task is to classify each MRI scan into one of four classes which correspond to the stage of Alzheimer’s disease. Two different training methods are compared to a training baseline. The first method evaluated is class-weighting, a technique that tries to compensate for rare classes in imbalanced datasets. The second method is data augmentation, a techniquet hat extends the dataset in an attempt to reduce overfitting and increase performance. Class-weighting was found to improve the classification performance significantly on the rarest class, while not having too large effect on the other classes. Data augmentation was not found to improve performance in general, but did improve the recall on some classes.