Deep Learning-Based Respiratory Motion Correction for Positron Emission Tomography

Positron emission tomography (PET) plays a central role in oncological imaging due to its ability to visualize metabolic processes, making it valuable for tumor detection, staging, and treatment monitoring. However, in thoracic and abdominal regions, image quality is often compromised by artifacts caused by respiratory motion during ongoing data acquisition. These motion-related artifacts reduce spatial resolution, impair quantification accuracy, and can compromise diagnostic accuracy. In the past, different strategies have been developed to compensate for respiratory motion, including incorporating motion models into image reconstruction, motion-deblurring techniques, and gate-based registration techniques, where the acquisition data is partitioned into bins, known as “gates”, corresponding to different phases of the breathing cycle. More recently, deep learning has emerged as a promising approach for addressing various medical imaging tasks, ranging from segmentation to image reconstruction. Deep learning techniques have also been successfully applied to image registration tasks and might have the potential to offer superior solutions for respiratory motion correction. Correspondingly, the aim of this work was the development and evaluation of a deep learning-based procedure for respiratory motion correction for clinical PET data and to assess its strengths and potential weaknesses in comparison to established approaches. In particular, the investigation focused on oncological PET, with the objective of providing flexible and accurate motion correction while avoiding local image corruption caused by spurious local deformations unrelated to the actual patient motion. To address this task, an unsupervised CNN-based framework was developed for fully automated gate-to-gate image registration of clinical PET investigations divided into multiple respiratory gates. The registration network was designed to take pairs of gated images and predict deformation fields that warp the individual gates to align them with a mutual reference gate. After registration, the aligned gates were averaged to yield the motion-corrected image. The network was trained in an unsupervised manner, i.e., without requiring ground truth data, by optimizing a loss function based on linear voxel-wise correlation as the image similarity metric. The loss function was combined with two regularization terms: one ensuring smooth deformations and the other encouraging volume preservation in high-intensity regions. This design aimed to predict anatomically plausible deformations that would effectively align all gates and thereby correct the motion artifacts while avoiding spurious distortions that could compromise clinical interpretation. The choice of reference gate has a crucial influence on the quality of the obtained results. Among the considered options, the mid-inspiration gate was found to be the most stable and suitable reference gate. Additionally, a cascading scheme was implemented and evaluated, in which the network was applied repeatedly with varying matrix and voxel sizes of the respective input images. The predicted deformation fields of the cascading stages were combined and applied to the original input images to minimize interpolation artifacts. This approach improved motion correction, particularly in cases with large respiratory motion amplitudes, since residual motion artifacts could be resolved by the repeated network applications without the risk of unrealistic deformations. Evaluation of the obtained results encompassed both quantitative and qualitative criteria. Quantitative evaluation included gate-based assessment of registration accuracy, as well as analysis of the predicted deformation vector fields and the respective Jacobian determinants. Motion correction performance was assessed using a combination of image-based metrics and qualitative visual inspection. Furthermore, the final framework was compared to the commercial motion correction…