Automated detection and classification of patients with Alzheimer’s disease on FDG PET neuroimaging using convolutional recurrent neural networks

Introduction: Alzheimer’s disease (AD) is a leading cause of dementia accounting for up to 80% of all dementia cases worldwide (1). Early detection and diagnosis of AD are highly important clinical needs for preventing and delaying disease progression in patients with AD. Brain ¹⁸F-fludeoxyglucose (FDG) PET neuroimaging has been shown to be useful in the early detection and diagnosis of AD as well as for distinguishing various neurogenerative diseases (2). Patients with mild cognitive impairment (MCI), a cognitive disorder characterized by memory impairment, have an increased chance of developing AD (3). We aimed to develop a deep learning (DL) approach based on convolutional recurrent neural networks to perform classification of patients with AD, MCI, and cognitively normal (CN) healthy controls using FDG PET neuroimaging data.

Methods: Data from 146 participants were extracted from the publicly available Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (4). FDG PET images and Alzheimer’s Disease Assessment Scale-Cognitive (ADAS-Cog) scores were extracted from baseline, 6 months after baseline, and 1 year after baseline. ADAS-Cog is a commonly used cognitive outcome measure where higher scores indicate higher cognitive disease severity for patients with AD (5). Additional clinical measures, including age and gender, were extracted. The data consisted of 35 patients with AD, 69 patients with MCI, and 42 CN healthy controls. The patient data were randomly partitioned into training, validation, and test sets using a 70%/10%/20% split where each dataset consisted of 99, 18, and 29 patients, respectively. Maximum intensity projections of the FDG PET images and the clinical ADAS-Cog scores at baseline, 6 months after baseline, and 1 year after baseline were used as inputs to the DL approach for the classification task (Figure 1a). A convolutional long short-term memory (LSTM) network, a type of recurrent neural network, extracted relevant spatiotemporal imaging features from the input FDG PET images, and a LSTM network extracted the relevant temporal features from the input ADAS-Cog scores. Those extracted features were combined with the input clinical measures of age and gender in a fully connected network to yield the predicted classification. The DL approach was trained and optimized on the training set for 200 epochs with a batch size of 4 using a multiclass categorical cross-entropy loss function and a stochastic gradient decent optimization algorithm based on adaptive moment estimation (Adam). Early stopping was applied on the validation set and dropout regularization was used during training to prevent overfitting. The approach was evaluated on the test set where the overall accuracy, precision, recall, F1 score, receiver operating characteristic (ROC) curve analysis, and the confusion matrix were used to assess performance.

Results: The DL approach yielded an overall accuracy of 0.86 (95% confidence interval: 0.74, 0.99) on the test set. The ROC curves for each class on the test set are shown in Figure 1b. The approach yielded area under the ROC (AUROC) curve values of 0.99, 0.96, and 0.99 for classifying participants in the CN, MCI, and AD classes, respectively. The approach had an overall AUROC of 0.98 across all classes on the test set. The approach yielded precision values of 1.00, 0.83, and 0.86, recall values of 0.67, 0.94, and 0.86, and F1 scores of 0.80, 0.88, and 0.86, for the CN, MCI, and AD classes, respectively, on the test set. The approach yielded an overall precision of 0.87, an overall recall of 0.86, and an overall F1 score of 0.86 across all classes on the test set. The confusion matrix for the approach on the test set is shown in Figure 1c.

Conclusions: The developed DL approach holds promise for early detection and diagnosis of AD and was able to distinguish between patients with AD, MCI, and CN healthy controls using baseline FDG PET neuroimaging data and longitudinal clinical data as inputs.

• Download figure
• Open in new tab
• Download powerpoint