Spatial Normalization to Improve Deep Learning-based Head Motion Correction in PET

Head motion during brain positron emission tomography (PET) studies reduces image quality, lowers concentrations in high-uptake regions, and causes mis-estimation in tracer kinetic modeling. Previously, we proposed a data-driven motion correction method, Deep Learning-based Head Motion Correction (DL-HMC), to predict rigid motion for brain PET. However, since different subjects have different head positions for their reference images, using one fixed reference frame for each subject can result in suboptimal model performance when predicting continuous motion. To improve the accuracy of motion prediction, we propose using spatial normalization (SN) in DL-HMC to mitigate the negative effects of varying reference frame positions. We tested our proposed SN-DL-HMC method on a multi-subject dataset and compared its performance with other methods. Our results demonstrate that SN-DL-HMC outperforms other methods in predicting continuous head motion, and should therefore be considered a standard procedure in our DL-HMC pipeline.