Robust MR-free Grey Matter Extraction in Amyloid PET/CT Studies with Deep Learning

Quantification of amyloid PET studies is most accurate if regions of interest (ROIs) are not affected by the presence of cerebrospinal fluid. Patients with high amyloid load often have great atrophy, therefore, the use of atlas-based ROIs, instead of patient specific anatomy, can underestimate amyloid load, leading to a bias. Traditionally, this can be overcome only using MR anatomical sequences, which are burdensome and might not be ideal to be performed for each patient in the clinical routine. In this work, we propose to overcome this issue by using a method based on deep learning. As CT scans provide anatomical information, even at the very low doses used for PET attenuation correction, we propose the use of such a scan, together with the PET one, for a U-NET based segmentation. The approach achieves a median DICE score of 77% on a validation cohort of N=20 patients, even when using only N=14 patients in the training dataset. A dedicated data augmentation strategy is used, and the individual contribution of each modality is analyzed. We find that the joint effect of PET and CT is beneficial (median DICE: PET only 73.0%, CT only 74%). A near perfect correlation with MR-based quantification was also found.