Predicting Cognitive Decline with Deep Learning of Brain Metabolism and Amyloid Imaging

• MCI means mild cognitive impairment; FDG, 18F-fluorodeoxyglucose; CNN, convolutional neural network; and ADNI, Alzheimer’s Disease Neuroimaging Initiative.Source: Abbreviations · p. 3
• CDR-SB means Clinical Dementia Rating Sum of Boxes; ADAS-Cog, Alzheimer’s Disease Assessment Scale-Cognitive Subscale 13-item; FAQ, Functional Activities Questionnaire; and MMSE, Mini-Metal Status Examination.Source: Abbreviations · p. 3
• ROC means receiver operating characteristic.Source: Abbreviations · p. 3

• MCI has heterogeneous causes and decline rates, and some patients never convert to AD, making identification of likely treatment beneficiaries important [2].Source: Introduction · p. 4
• Previous studies used FDG and AV-45 PET quantitative parameters or visual assessment, but visual analysis lacked objective quantitative data and quantitative analyses required complicated processing [3-6, 7-10].Source: Introduction · p. 4
• Applied to FDG and AV-45 PET, the CNN discriminated MCI groups by cognitive outcome with minimized image processing and combined both modalities quantitatively.Source: Introduction · p. 4
• The CNN-based biomarker strongly correlated with future cognitive decline.Source: Introduction · p. 4

Methods

• The CNN was learned for diagnostic classification using AD and healthy subjects with baseline FDG and AV-45 PET scans.Source: Subjects · p. 5

Subjects

• MCI participants had baseline FDG and AV-45 scans plus three-year clinical follow-up, and were grouped by conversion to AD within three years.Source: Subjects · p. 5
• The cohort comprised 182 normal controls, 139 AD patients, and 171 MCI subjects.Source: Subjects · p. 5
• Cognition was measured with CDR-SB, ADAS-Cog, FAQ, and MMSE, with longitudinal changes assessed at one and three years.Source: Subjects · p. 5

FDG PET and AV-45 PET

• The investigators downloaded preprocessed PET images from ADNI and used them for CNN training and testing as they were.Source: FDG PET and AV-45 PET · p. 6
• FDG PET was acquired 30–60 minutes and AV-45 PET 50–70 minutes after injection; images were co-registered, frame-averaged, and standardized to 1.5 x 1.5 x 1.5 mm voxels.Source: FDG PET and AV-45 PET · p. 6
• Preprocessing omitted nonlinear spatial warping, so brain size and shape were not changed.Source: FDG PET and AV-45 PET · p. 6

Study design

• The AD/NC-trained CNN was applied directly to classify MCI converters versus nonconverters, independently of MCI PET training data.The study aimed to predict cognitive decline and identify subjects who would eventually convert to AD; Figure 1 describes 10-fold cross-validation of AD/NC training and assessment of MCI conversion accuracy and ROC.Source: Study design · p. 6, Figure 1 · p. 7
• The network’s final output was defined as ConvScore, indicating how close baseline images are to AD and proposed for use as a predictive biomarker.The output layer nodes were reassigned to MCI converter and nonconverter.Source: Study design · p. 6

Preparation for 3-dimensional deep convolutional neural network

• The study used a 3D CNN because its shift-invariant property can detect features across three-dimensional spatial data.The authors describe CNN shift invariance as allowing detection of target objects across spatial space.Source: Preparation for 3-dimensional deep… · p. 8
• FDG and AV-45 PET images were used as two channels so the network could exploit diverse features from both modalities.Source: Preparation for 3-dimensional deep… · p. 8
• After co-registration of two modalities, voxel-size standardization, and scanner-specific smoothing, image data were rescaled; each modality (FDG and AV-45) was rescaled to the range 0 to 255.Source: Preparation for 3-dimensional deep… · p. 8
• The preprocessed 160 x 160 x 96 FDG and AV-45 volumes were concatenated into a 160 x 160 x 96 x 2 input for the network.Source: Preparation for 3-dimensional deep… · p. 8

Network architecture

• The model combines three convolutional layers with ReLU and max pooling, followed by one fully connected layer.The architecture is described as an efficient multilayer CNN.Source: Network architecture · pp. 9–10
• Feature maps increase from 2 input maps to 64, 128, and 512 across the network’s depth.The first convolutional layer receives a 160 × 160 × 160 × 2 input volume.Source: Network architecture · p. 9
• Supervised training assigns output labels of one for AD and two for normal controls and minimizes a loss using stochastic gradient descent.The network was trained on AD and normal-control imaging data with 10-fold cross-validation, for 50 epochs.Source: Network architecture · p. 10
• The two softmax-node scores represent AD and normal-control scores, or MCI-converter and nonconverter scores, respectively.The AD-node value, called ConvScore, indicates how close input PET data are to AD or an MCI converter.Source: Network architecture · pp. 10–11

Training and validation of classification between AD and NC

• Ten-fold cross-validation repeatedly trained on nine randomly assigned subsets and evaluated classification performance on the held-out testing subset.Convolutional and fully connected layers were initialized with random weights.Source: Training and validation of… · p. 11
• Left-right flipping augmented the training data; each fold used 289 x 2 (or 288 x 2) image pairs for training and 32 (or 31) for validation.The flipping was based on the anatomical symmetry of the brain.Source: Training and validation of… · p. 11
• Sensitivity, specificity, and accuracy were calculated across validation subsets; AD was predicted when its softmax probability exceeded 0.5, and ROC analysis used the AD-node output ConvScore.Source: Training and validation of… · p. 11

Prediction of cognitive decline in MCI subjects

• The AD/NC-trained CNN was applied to MCI PET images, with output labels changed to MCI converter and nonconverter and converter status assigned above a 0.5 probability threshold.A patient was classified as a predictive converter when the network output probability exceeded 0.5.Source: Prediction of cognitive decline in MCI… · pp. 11–12
• Sensitivity, specificity, and accuracy were measured, and ROC analysis was performed using ConvScore.Source: Prediction of cognitive decline in MCI… · p. 12
• ConvScore was correlated with longitudinal changes in CDR-SB, ADAS-Cog, FAQ, and MMSE scores.Changes at 1- and 3-year follow-ups were calculated relative to baseline, and Pearson correlation was used.Source: Prediction of cognitive decline in MCI… · p. 12

Feature volume of interests based analysis

• Feature VOI analyses of FDG and AV-45 PET were conducted to compare conventional quantification methods with CNN-based biomarkers.Source: Feature volume of interests based… · p. 12
• FDG volumes were spatially normalized using nonlinear warping, and regional uptake was expressed relative to the pons and cerebellar vermis.Mean uptake was calculated in the right and left angular and temporal cortices and posterior cingulate cortex.Source: Feature volume of interests based… · p. 12
• For AV-45 PET, FreeSurfer-segmented cortical regions supplied regional uptake measures, with whole-cerebellum uptake as the reference.Regions included frontal, anterior and posterior cingulate, lateral parietal, and lateral temporal cortex.Source: Feature volume of interests based… · p. 12

Statistics

• McNemar’s nonparametric test compared CNN and feature VOI diagnostic and prediction accuracy.Source: Statistics · p. 12
• ROC AUCs were measured for ConvScore and feature VOI parameters and compared using DeLong’s test for correlated curves.Source: Statistics · pp. 12–13
• Pearson’s correlation assessed the association between ConvScore and longitudinal changes in cognitive measurements.Source: Statistics · p. 13
• Statistical analyses used MATLAB, with P-value < 0.05 considered significant.Source: Statistics · p. 13

Results of AD classification and MCI conversion prediction

• For AD-versus-NC classification, CNN sensitivity, specificity, and accuracy were 93.5%, 97.8%, and 96.0%, significantly exceeding VOI-based analyses.Source: Results of AD classification and MCI… · p. 14
• For predicting MCI conversion, CNN sensitivity, specificity, and accuracy were 81.0%, 87.0%, and 84.2%, respectively; CNN specificity was significantly higher than VOI-based AV-45 PET.Sensitivity and accuracy were also higher than conventional methods, but this difference did not reach statistical significance.Source: Results of AD classification and MCI… · p. 14, Table 2 · p. 15
• ConvScore AUCs were 0.98 versus 0.91 (FDG) and 0.84 (AV-45) for AD classification, and 0.89 versus 0.82 and 0.83 for MCI conversion.Differences were significant for both tasks: p < 0.001 for AD classification; p < 0.01 versus FDG and p < 0.05 versus AV-45 for MCI conversion.Source: Results of AD classification and MCI… · p. 14, Figure 2 · p. 16

Correlation of CNN-based biomarker with cognitive outcomes

• ConvScore correlated with longitudinal changes in cognitive measurements at 1 and 3 years, positively for CDR-SB, ADAS-Cog, and FAQ and negatively for MMSE.At 1 year, correlations were CDR-SB r=0.37, ADAS-Cog r=0.29, FAQ r=0.40, and MMSE r=-0.30. At 3 years, they were CDR-SB r=0.63, ADAS-Cog r=0.24, FAQ r=0.67, and MMSE r=-0.61.Source: Correlation of CNN-based biomarker… · p. 16
• ConvScore was significantly higher in MCI converters than nonconverters.Source: Correlation of CNN-based biomarker… · p. 16, Figure 4 · p. 18

Discussion

• The network automatically extracts features from spatially unnormalized baseline AD and normal-control images, enabling prediction by simply inputting a subject’s PET images.The authors describe the method as requiring minimal image processing and no feature extraction or spatial transformation for converter prediction.Source: Discussion · pp. 19, 21
• Conventional PET biomarkers required region-based or whole-cortical uptake measures and processing such as spatial normalization and MRI-based segmentation, which was not standardized and could introduce errors.The authors note potential error particularly when nonlinear transformation is applied to brains with morphological alterations.Source: Discussion · p. 19
• 84.2% accuracy for distinguishing MCI converters from nonconverters outperformed conventional feature-VOI methods, and prediction accuracy was significantly higher than other methods in ROC comparisons.The approach also outperformed feature-selection machine-learning algorithms for this differentiation task, although those studies used different imaging modalities and clinical variables.Source: Discussion · p. 19
• ConvScore fuses both PET modalities and strongly correlates with longitudinal cognitive measurements; high baseline scores indicate patients whose cognition could rapidly deteriorate.The authors suggest it could help select prodromal patients likely to benefit from early intervention in clinical trials.Source: Discussion · p. 20
• The authors identify validation in completely independent cohorts as future work and propose that the CNN interpretation system could support biomarker development across several disease areas.They mention cancer, cardiovascular disorders, and neurodegenerative diseases as potential applications.Source: Discussion · pp. 20–21
• Without left-right image-flip augmentation, the network achieved 89.4% accuracy for AD-versus-NC differentiation and 81.3% for predicting MCI converters.Augmentation increased performance but could introduce classification errors because the brain hemispheres have partly different functions; larger cohorts and deeper networks may improve the approach.Source: Discussion · p. 21