Detecting Alzheimer Disease on FDG PET Images Using a Similarity Index Based on Mutual Information

Mutual information is an image similarity metric often used for the robust registration of multimodality images. The aim of this study is to investigate the use of a simple to implement similarity computation method based on a mutual information index for the automated detection of Alzheimer's disease from FDG PET studies. 102 healthy and 95 Alzheimer's disease FDG PET patient images from the online Alzheimer's disease Neuroimaging Initiative (ADNI) database were used to develop and test the system. Images were preprocessed for enabling comparison. An index was computed for each new image based on its degree of similarity to images belonging to AD patients versus healthy control patients. Classification was made based on the value of this index. The leave-one-out method was used for performance evaluation. Performance was evaluated using Receiver Operating Characteristic (ROC) curves. The diagnostic reliability given by the area under the curve (AUC) was determined as $0.857\pm 0.0261$. The results suggest that a mutual information based image similarity method can potentially be useful as a second opinion computer aided diagnostic (CAD) system providing verification to visual and black box approaches. The system does not need training with new data and does not require the computation of image features.