Unsupervised learning to perform respiratory motion correction in PET imaging

Introduction: Due to the long acquisition times required for positron emission tomography (PET) imaging, respiratory motion can have a substantial effect on the quality of the resulting images. This is especially true when evaluating tumors that are located in areas that are more prone to movement, such as those close to the diaphragm. To correct for this issue, coincidences detected only during a particular phase of the breath (e.g. the quiescent phase) are typically used to generate gated images. However, by only using a small portion of the data collected during the entire imaging process, the number of detections in the gated image compared to a non-gated one is lower and the images are noisier. To solve this problem, one either has to increase the scan time or inject a higher activity of the radiopharmaceutical. Both solutions are not ideal, as longer scan times decrease patient comfort and throughput, while increasing the activity subjects the patient to more harm from the additional radiation. Another method currently used in the literature is to perform a registration between the different phases of the breathing motion using mutual information, which utilizes all of the detections acquired during the scan. In this project we aim to explore an unsupervised machine learning approach to perform these registrations.

Methods: A training dataset of 270 patients was generated using the 4D extended cardiac-torso (XCAT) anthropomorphic digital phantom with (2×4×4) mm³ voxels, exhibiting varying inhalation levels for different anatomies as well as a variety of other patient parameters. Training the machine learning network was accomplished by providing the model with two PET frames from different breathing phases. The network then predicted the pixel-wise shift in 3D from one frame to the other. By comparing the shifted frame to the target frame, the model learned to produce more accurate shifts, making the learning process unsupervised. As a result, the trained model is able to receive different gated PET images and groups them all into a motion-corrected single bin; providing a final image without the blurring effects that were initially observed. By utilizing the XCAT phantom, accessing the ground-truth motion corrections was trivial, making the validation of the network performance on similar data robust and straight-forward. For a variety of different breathing amplitudes, improvements in tumor recovery were quantified by computing the intersection over union of the tumors as well as the enclosed activity before and after the corrections were applied. Furthermore, the pixel-wise differences in the shifts between the predictions and the ground-truths were calculated.

Results: Comparing the uncorrected and corrected image to the ground truth distribution, the network was found to provide relative improvements of 73% and 81% for the intersection over union and total activity, respectively, when correcting for a diaphragm shift of 21 mm. Improvements in tumor recovery were found to be consistent across the tested range of breathing amplitudes. For a diaphragm shift of 21 mm, the median absolute residual error in the flow predictions was less than 1.5 mm, which is smaller than the smallest voxel dimension of the phantom.

Conclusions: Our results suggest that our proposed method has the potential to accurately correct for breathing motion artifacts in PET. Furthermore, this method is unsupervised, avoiding the need for human intervention when transferring the method to the clinical setting.

• Download figure
• Open in new tab
• Download powerpoint