Extraction of Features from Patch Based Graphs for the Prediction of Disease Progression in AD

There has been significant interest in approaches that utilize local intensity patterns within patches to derive features for disease classification. Several existing methods explore the patch relationship between different subjects for classification. Other methods utilize the patch relationship within the same subject to aid classification. In this paper, we proposed a new approach to extract different types of features by utilizing the patch relationships within and between subjects. Specifically, features are first extracted by exploiting the patch relationship between subjects. Then, the relationship among patches within the same subject is modeled as a network and features are derived from the constructed network. Finally, these two different types of features are integrated into one framework for identifying mild cognitive impairment (MCI) subjects who will progress to Alzheimer’s disease. Using the standardized ADNI database, the proposed method can achieve an area under the receiver operating characteristic curve (AUC) of 81.3 % in discriminating patients with stable MCI and progressive MCI in a 10-fold cross validation, demonstrating that the integration of patch relationship between and within subjects can aid the prediction of MCI to AD conversion.