The Impact of Motion Correction and Partial Volume Correction on Classification of 11C-PiB SUVR Images with Supervised Machine Learning Models

Objectives: To determine the impact of image processing methods on binary classification of 11C-PiB data with supervised machine learning models. Methods Three pre-processing pipelines were applied to 11C-PiB scans (acquired on a Siemens mMR): no motion correction (nMC), motion correction (MC) and motion correction plus iterative Yang partial volume correction (PVC). SUVR (cerebellum norm.) images in native T1 space were created for all processed images. SUVR from 26 brain regions (features) were recorded for all images and binary classification was performed with two supervised machine learning models: linear discriminant analysis (LDA) and support vector machines (SVM). For each pre-processing pipeline the models were trained and tested on a set of 40 scans; Alzheimer's disease (AD) = 20, no cognitive impairment (NCI) = 20. Data was divided into 1000 different train/test combinations (30/10) with equal group numbers (nAD, nNCI = 15/5) and optimised models were calculated using leave-one-out cross validation. The impact of pre-processing methods on scan classification was assessed by calculating the accuracy and the area under the ROC curve (AUC) from classifying the unseen test data with the trained models. All possible combinations of 1 to 6 features were assessed for both LDA and SVM. Results For LDA, nMC data demonstrates the highest accuracy and AUC. A positive correlation between accuracy and number of features was observed in the LDA models: r2 = 0.9 (p = 0.01), 0.8 (p = 0.01) and 0.8 (p = 0.01) for nMC, MC and PVC respectively. For nMC data the ventral diencephalon (DC) was found consistently in maximum accuracy combinations of 2 to 6 features and the brain stem was also found consistently for maximum accuracy combinations of 3 to 6 features. For MC data the caudal anterior cingulate was found consistently in maximum accuracy combinations of 2 to 6 features. For PVC data the brain stem and ventral DC were found in maximum accuracy combinations of 5 and 6 features. The maximum accuracy achieved with the LDA models was higher than the maximum accuracy achieved with SVM models for nMC (LDA 6 features = 91%, SVM 4 features = 86%), MC (LDA 5 features = 89%, SVM 3 features = 85%) and PVC (LDA 6 features = 86%, SVM 5 features = 85%) data. For the SVM models the maximum accuracy was observed at 4, 3 and 5 features for nMC, MC and PVC respectively. For greater number of features accuracy decreased. For MC data the hippocampus and subcortical white matter w