Image Intensity Normalization Benefits Deep Learning Brain PET Motion Correction

Machine learning algorithms are now widely used to perform various tasks in medical imaging research, and some studies have shown that such methods are sensitive to input data normalization. While this pre-processing step has been investigated on other imaging modalities, few studies are available on PET images intensity normalization as neural network inputs. In this work, we explore different normalization strategies on input data used to train a Deep Learning Head Motion Correction (DL-HMC) method that we previously proposed. Qualitative and quantitative analysis on testing subjects show that motion prediction benefits from normalization. Our results demonstrate that normalizing images with the intensity of the earliest frame significantly decreases motion prediction error on testing subjects compared to data that are not normalized or normalized using other strategies. Reconstructions and Standard Uptake Value analysis further confirm this observation. We validate the proposed method on data from two tracers.