Classification of amyloid PET images using novel features for early diagnosis of Alzheimer's disease and mild cognitive impairment conversion

Yu Yan,Edward Somer,Vicente Grau

DOI: https://doi.org/10.1097/MNM.0000000000000953

Abstract:Background: New PET tracers could have a substantial impact on the early diagnosis of Alzheimer's disease (AD), particularly if they are accompanied by optimised image analysis and machine learning methods. Fractal dimension (FD) analysis, a measure of shape complexity, has been proven useful in MRI but its application to fluorine-18 amyloid PET has not yet been demonstrated. Shannon entropy (SE) has also been proposed as a measure of image complexity in DTI imaging, but it is not yet widely used in radiology. Materials and methods: In this study, one volumetric FD method and one volumetric SE method were applied to fluorine-18-flutemetamol and fluorine-18-florbetapir 3D amyloid images from 65 and 281 participants, respectively, including healthy volunteers, and patients with probable Alzheimer's disease (pAD) or mild cognitive impairment (MCI). Results: The group average FD of white matter surface and SE of white matter volume for healthy volunteers were higher than for pAD patients. Both FD and SE are effective in the identification of MCI patients who progress to pAD during the 2-year follow-up (ground truth). Finally, we developed a support vector machine multimodal classification framework using both PET and MRI features, which showed higher accuracy compared to traditional standard uptake value ratio or using PET alone. The classification accuracy for flutemetamol and florbetapir is 88.9 and 83.3%, respectively, for MCI progression, which is competitive with existing literature. Conclusion: The results presented in this study demonstrate the potential of FD and SE methods for the analysis of brain PET scans in early AD diagnosis and in the prediction of MCI-AD conversion.