Convolutional Neural Networks for Multiclass Predictions of Alzheimer's Disease Progression
Alzheimer’s disease (AD) is a neurodegenarative disorder that is increasingly affecting the worldwide population. There are currently no disease-modifying treatments available, however, promising drugs are being tested in multi-cohort clinical trials. Accurate early-stage forecasts of the development of AD progression are helpful in improving cohort selection. In this study, I investigate the performance of a multi-input convolutional neural network (CNN) for predictions of AD progression. I specifically focus on the generalisability of these forecasts to previously unseen individuals. The multiclass predictions distinguish between cognitively normal (CN) patients, patients with a mild cognitive impairment (MCI), and AD patients. The model is based on grey matter density maps of brain MRI scans and two non-image features, i.e. current diagnosis and follow-up time. One of the main challenges of predicting future diagnosis is to outperform a simple model that always predicts current diagnosis. I therefore introduce a custom loss function that differentiates between non-converters, converters to MCI, and converters to AD. Furthermore, the Grad-CAM tool is used to visualise model activations in the brain and the external performance of the methods is validated on a separate dataset. This study shows that the progression of Alzheimer’s disease can be predicted quite accurately based on a single MRI scan and corresponding current diagnosis. When converters are of interest, a custom loss function is beneficial for tackling the imbalance in the data. A model that uses such a custom loss function generalises quite well to predictions further in the future and to an external dataset. The Grad-CAM results show that model activations in the brain are not always consistent, however, the cerebellum appears to be an important brain region for MRIbased predictions of Alzheimer’s disease progression.