ALZHEIMER’S DISEASE CLASSIFICATION
The influence of registration and brain extraction on the learned features by heat mapping was investigated and a novel relevance-guided training method was compared, focusing on brain tissue, yielded highest classification accuracies and confirmed histopathologically relevant regional iron deposition.
Abstract
When using deep neural networks to separate Alzheimer’s disease patients (n=119) from normal controls (n=131) by using MR images, heat mapping revealed that the image preprocessing is introducing misleading features used by the classifier. Therefore we systematically investigated the influence of registration and brain extraction on the learned features by heat mapping. Results were compared to a novel relevance-guided training method, focusing on brain tissue. The relevance-guided configurations yielded highest classification accuracies and also confirmed histopathologically relevant regional iron deposition. KeywordsDeep convolutional neural networks, heat mapping, relevance guidance, Alzheimer’s disease