Fig 1.
Comprehensive diagram of neuroimaging study.
Upper section details data processing techniques including format conversion and normalization, while the lower section depicts the neural network structure with layers like 3D Convolutional Blocks and LeakyReLU. The diagram also includes information of the ADNI Cohort we used, with specific scan dimensions.
Table 1.
Demographic information of ADNI fMRI cohort.
Table 2.
Comparison of classifier performance.
Table 3.
Statistical comparison of model performance on resting and PW datasets.
Fig 2.
Comparative ROC curves for neural network models.
Receiver Operating Characteristic (ROC) curves comparing the performance of different neural network models, including Our CNN, Gupta’s CNN, Baseline RNN, and Baseline MLP, across different datasets. Each curve represents the trade-off between the True Positive Rate (TPR) and False Positive Rate (FPR) for a specific model, providing insights into their diagnostic ability in distinguishing between classes.
Fig 3.
Comparison of max-pooled and normalized feature maps from the first convolutional layer of our proposed model for AD prediction.
On the left, feature maps from a participant diagnosed with Alzheimer’s Disease (AD) show distinct patterns of activation in regions associated with the condition. On the right, feature maps from a cognitively normal (NC) participant exhibit different attention to activations, highlighting differences leveraged by the CNN for diagnostic classification.
Fig 4.
Visualization of SHAP values on grayscale fMRI scans.
The left image depicts the spatial distribution of influential voxels for a patient with Alzheimer’s Disease, and the right image for a NC subject. The overlays highlight the voxels that most significantly contribute to the CNN model’s predictive differentiation between AD presence and absence. These visualizations indicate potential regions impacted by AD and the model’s classification strength.
Fig 5.
Multi-angled fMRI scans highlighting regions of hyperactivation in the hippocampus of an Alzheimer’s Disease patient.
The color overlay indicates areas of hippocampus region, which are known to related to AD. The left panel shows a sagittal view, the center panel depicts a coronal view, and the right panel illustrates an axial view, together providing a comprehensive 3D perspective of hippocampal engagement.
Fig 6.
SHAP Maximum Intensity Projection (MIP) for Alzheimer’s Disease prediction.
The heatmap illustrates the concentration of SHAP values across various brain slices, with brighter colors indicating higher importance in the model’s predictive assessment. This MIP view consolidates the most significant voxels, offering a comprehensive perspective on the regions critical for AD classification.