Impact of Defacement on Deep Learning-based Amyloid Prediction with Multi-contrast MRI

Motivation: Defacement has become standard in neuroimaging to protect privacy but may inadvertently remove valuable information for predicting amyloid PET positivity from MRI scans. Goal(s): To assess whether defacement impacts deep learning models' ability to predict amyloid positivity from MRI data. Approach: Using non-processed, defaced, and skull-stripped T1w and T2-FLAIR images, we trained EfficientNet-B1 and SEResNet50 models to predict amyloid PET positivity, comparing performance across different inputs. Results: Skull-stripped images achieved the highest accuracy, outperforming both non-processed and defaced images. While facial information may not enhance predictions, defacement reduced accuracy, as non-defaced images outperformed defaced ones. This warrants further exploration with advanced models. Impact: Defacing MRI scans for privacy inadvertently removes information that may aid amyloid PET positivity prediction, as shown by decreased model performance on defaced images compared to non-defaced ones.