Room: AAPM ePoster Library
Despite a clear clinical need, there are no generalizable tools to distinguish between Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI), or to identify patients who will progress from MCI to AD. We develop and evaluate a flexible, comprehensive convolutional neural network (CNN) to diagnose AD and MCI and predict disease progression using 18F FDG PET images.
Data from 928 patients from the ADNI database were used to implement a 3D-CNN, expanded from previous work. Model inputs included FDG PET images and clinical data. The model was adapted and run twice to perform two separate tasks: classification of disease diagnosis (AD, MCI, normal control (NC)) and prediction of progression of MCI to AD (stable (sMCI) versus progressing (pMCI)). To explore the interaction of the tasks, the model was also implemented as a 4-class model (AD, pMCI, sMCI, NC) performing both tasks simultaneously. Training was performed using class balancing, holding 15% of the scans for validation and testing, and optimizing accuracy. Performance was assessed using ROC analysis and accuracy.
The classification model achieved an accuracy of 67.6% and AUCs of 0.78, 0.79, and 0.69 for AD, MCI, and NC, respectively. The prediction model distinguished MCI patients who would develop AD from those with stable MCI with accuracy of 70.8% and AUC of 0.74. The 4-class model achieved an accuracy of 47.1%, with AUCs of 0.65, 0.56, 0.76, and 0.63 for AD, pMCI, sMCI, and NC, respectively. This inferior performance may suggest the presence of specific features predictive of MCI evolution which are not present in AD.
The success of this flexible 3D-CNN for the individual tasks of diagnosing AD and predicting progression of MCI demonstrates progress toward the identification of a PET-based pattern for AD tracking. The multitask model exploration provides insight toward understanding the disease’s evolution.
Not Applicable / None Entered.