Deep learning improves utility of tau PET in the study of Alzheimer's disease

Abstract

Introduction: Positron emission tomography (PET) imaging targeting neurofibrillary tau tangles is increasingly used in the study of Alzheimer's disease (AD), but its utility may be limited by conventional quantitative or qualitative evaluation techniques in earlier disease states. Convolutional neural networks (CNNs) are effective in learning spatial patterns for image classification.

Methods: 18F-MK6240 (n = 320) and AV-1451 (n = 446) PET images were pooled from multiple studies. We performed iterations with differing permutations of radioligands, heuristics, and architectures. Performance was compared to a standard region of interest (ROI)-based approach on prediction of memory impairment. We visualized attention of the network to illustrate decision making.

Results: Overall, models had high accuracy (> 80%) with good average sensitivity and specificity (75% and 82%, respectively), and had comparable or higher accuracy to the ROI standard. Visualizations of model attention highlight known characteristics of tau radioligand binding.

Discussion: CNNs could improve tau PET's role in early disease and extend the utility of tau PET across generations of radioligands.

FIGURE 1

Summary overview of project pipeline. A, Image preprocessing steps, involving (1) motion correction (with FSL), (2) registration of each image to a template generated using ANTS, (3) creation of three time acquisition windows (80–100, 85–105, and 90–110 minutes post‐injection) from each scan's available windows (80–110) for the MK‐6240 dataset (the 80–100 minute time window was used for AV‐1451), (4) rotation of each time window image by 7, 14, and 21 degrees along the sagittal plane for data augmentation (see Supplementary Methods 2.3.2 for details), (5) averaging and internal normalization of uptake values. B, 2D image generation for input into 2D inception model. The orientation of each coronal slice is shown (R = right, L = left, S = superior, I = inferior) and numbered here to show slice order from rostral to caudal. All images subsequently shown are the same orientation. We elected to generate five such images for each subject, with differing coronal slices used. C, Determination of binary label with either clinical status (MK‐6240 dataset and AV‐1451 dataset) or cognitive test result when formal determination unavailable (MK‐6240 dataset). D, Each cycle of 5‐fold cross‐validation, which involves input of train set images (pink) into the model returning a scalar prediction of likelihood of binary impairment status, followed by model weight adjustment based on accuracy of the prediction (using batch gradient descent), followed by testing external validity of the model on an independent validation set (green). The model with the highest accuracy after 30 epochs is then tested on a holdout test set (yellow). AD, Alzheimer's disease; CN, cognitively normal; MCI, mild cognitive impairment; SRT‐DFR, Selective Reminding Test, Delayed Free Recall

FIGURE 2

ROC curves for 2D and 3D models versus composite/entorhinal cortex (EC) SUVR. Shown by radioligand. Training our neural network models involved using either a single (“singular”) radioligand or pooling together both radioligands (“combined”). Predictions from our “combined” 2D/3D model configurations are assigned to the appropriate comparison group. 2D, two‐dimensional input model; 3D, three‐dimensional input model; EC, entorhinal cortex; ROC, receiver operating characteristic; SUVR, standardized uptake value ratio

FIGURE 3

Select heat maps for our MK‐6240 model with selected subjects. Negative values represent regions of positive predictive importance for impairment, whereas positive areas represent areas of negative predictive importance. The scale to the right of images represents proportional change from baseline prediction due to occlusion of specified region (i.e., more “important” areas have a larger absolute value and are brighter/darker on the maps). We highlight a specific slice for each example subject to the right. The probability of impairment as predicted by the model interpretation of the image is shown along with amyloid and actual impairment status to the right. The orientation of these heat map images (representing the 3 × 3 coronal slice images fed into the model) is the same as explicated in Figure 1. The leftmost column of images represents the input image, the middle column represents the generated sensitivity maps, and the paired images in the rightmost column are a representative comparison slice for each image and heatmap for each subject. The first image represents a true positive prediction, with relevance conferred by the sensitivity analysis to cortical binding in medial temporal and interestingly a contralateral parietal area. For these models, differential preference for different sides of an image (despite bilateral radioligand deposition) seems to be an important criterion. The second image represents another true positive prediction in an individual with low EC SUVR but high SUVR binding elsewhere, placing high diagnostic importance on a right midline parietal region. This suggests that the proposed method may have additional uses in non‐conventional AD subtypes. The third image represents an amyloid‐positive participant who had incidental tau deposition in a temporoparietal region without complaints of memory impairment and normal performance on cognitive testing. Notably, this area of incidental tau signal seems to have been identified by the model as having negative predictive value. The fourth image represents an amyloid negative without impairment, who exhibits high off‐target binding load. Interestingly, the neural network is able to identify a large portion (though not all) of this binding as non‐relevant to the classification task. AD, Alzheimer's disease; EC, entorhinal cortex; SUVR, standardized uptake value ratio