Decision support models for Alzheimer’s disease and mild cognitive impairment classification based on features extracted from MRI and FDG PET

Abstract

AbstractBackgroundAccurate diagnosis in early Alzheimers’ disease (AD) and Mild Cognitive Impairment (MCI) is critical for patients to start treatments. In AD and MCI patients, there is a high correlation between brain atrophy and metabolism and physical and cognitive impairments. Here we evaluate the performance of machine learning models for identifying AD and MCI subjects based on features extracted from MR and FDG PET imaging.MethodThe ADNI database was used in this study. 201 AD subjects: 80f(58‐88yo), 121m(56‐90yo); 427 MCI subjects: 163f(56‐90yo), 263m(59‐93yo); and 326 NL subjects: 169f(60‐95yo), 157m(66‐91yo) were included. T1‐weighted MRI and matched F18‐FDG PET scans were processed by PETQuant (CorTechs Labs, San Diego, USA). Segmented brain volumes normalized by intracranial volume and corresponding PET ROI SUV ratios (SUVRs) were calculated. Random Forests algorithm was used in creating machine learning models. One‐third of randomly selected data were used for testing while the rest were used for training. The performance of the models and importance of their features were evaluated.ResultThree models’ performances were evaluated. The first one used both PET SUVRs and MRI‐based volumes, and average model accuracy, recall for identifying AD, MCI and normal were 0.77 and 0.77. The top three features were right isthmus cingulate SUVR, left amygdala volume, and left hippocampus volume. The second model only used MRI, and average accuracy and recall for identifying AD,MCI and normal were 0.77 and 0.76. The top three features were left amygdala volume, left hippocampus volume, and right inferior temporal volume. The third model used PET SUVRs only, and average accuracy and recall were 0.72 and 0.72. The top three features were SUVRs from right isthmus cingulate, left medial‐parietal, and right entorhinal cortex.ConclusionThese models showed a good classification performance among AD, MCI and normal subjects. The model using MRI and PET showed the best performance and was slightly better than using MRI data alone. The MRI model performed better than using the PET data alone. This suggests MRI volumetric measures provide a better classification power than SUVRs extracted from FDG‐PET images.