Comparing Amyloid Imaging Normalization Strategies for Alzheimer's Disease Classification using an Automated Machine Learning Pipeline

Abstract

Amyloid imaging has been widely used in Alzheimer's disease (AD) diagnosis and biomarker discovery through detecting the regional amyloid plaque density. It is essential to be normalized by a reference region to reduce noise and artifacts. To explore an optimal normalization strategy, we employ an automated machine learning (AutoML) pipeline, STREAMLINE, to conduct the AD diagnosis binary classification and perform permutation-based feature importance analysis with thirteen machine learning models. In this work, we perform a comparative study to evaluate the prediction performance and biomarker discovery capability of three amyloid imaging measures, including one original measure and two normalized measures using two reference regions (i.e., the whole cerebellum and the composite reference region). Our AutoML results indicate that the composite reference region normalization dataset yields a higher balanced accuracy, and identifies more AD-related regions based on the fractioned feature importance ranking.

Figure 1:

Overall flowchart of the STREAMLINE pipeline.

Figure 2:

Prediction performance comparison for different algorithms: (a) Boxplot for balanced accuracy comparison. (b) Boxplot for ROC AUC score comparison.

Figure 3:

Feature importance heat maps for different datasets and methods. (a-d) represent the feature importance scores after normalization, fraction, weighted normalization, and weighted fraction, respectively. In each figure, each column lists the ROIs, and each row corresponds to one of the thirteen machine learning models with original data, whole cerebellum normalized data, and composite reference region normalized data (from top to bottom).

Figure 4:

The sum of the feature importance rank for each algorithm and rescaling method. (a-c) represent the results for original, whole cerebellum normalized, and composite reference region normalized datasets respectively. In each figure, the colors represent different AD-related regions, and the top of each bar is the sum of FI rankings for all regions under the different methods. The first thirteen bars represent the machine learning algorithms, and the “Mean” bar was the average values for the previous thirteen variables. Bars labeled ”All” represent the ROI’s overall rank for rescaling FI based on all algorithms. “All_n”, “All_n_w”, “All_f”, “ALL_f_w” represented the normalized, weighted and normalized, fractioned, normalized and fractioned FI ranking respectively. Red box highlights the two best performed dataset/rescaling method combinations

Figure 5:

Feature importance results of two best performed datasets/FI rescaling methods combinations. (a-b) are the ROIs rank plots based on the feature importance. Each bar represents an ROI, and different colors within the bar represent the algorithms. Higher bars represent the higher feature importance values. (c-d) the brain maps of the top five regions. Figures from left to right are the coronal, sagittal, and axial planes of the brain. (a) demonstrates the normalized feature importance with original data, and (c) plots its top five ROIs in the brain. (b) represents the fractioned FI with composite reference region normalized data, and (d) plots its top 5 regions in the brain.