An unsupervised deep learning framework for respiratory motion correction in PET


An unsupervised deep learning framework for respiratory motion correction in PET

Rosin, B.; Nikulin, P.; Hofheinz, F.; Maus, J.; Braune, A.; Kotzerke, J.; van den Hoff, J.

Breathing related patient motion is a serious source of image blurring in oncological whole body PET. Respiratory gating allows to subdivide the aquisition data into temporal bins ("gates") depending on the breathing cycle which are basically motion-free but exhhibit increased noise levels. Image registration of all gates to a reference gate and subsequent averaging is possible but traditional registration algorithms are frequently slow or of limited accuracy.
Deep learning methods have recently attracted much interest and have been successfully applied to image registration tasks. Their flexibility and execution speed are especially attractive in the present context. We have therefore implemented and evaluated an unsupervised deep learning framework for the registration of gated PET images.

Image volume pairs consisting of a fixed gate and a second moving gate serve as input to a convolutional neural network
which predicts a deformation vector field (DVF) mapping the moving image to the fixed image. The network is trained unsupervised by optimizing a similarity
metric between the registered image pairs as well as an additional regularization loss.
Fifty-two gated PET/CT scans (8 gates) were available for training (N=42; 2352 gate pairs) and testing (N=10).
Normalized cross correlation (NCC) was used as a measure of registration accuracy.
The motion corrected images were compared to the respective ungated image, single reference gate
and a commercially available motion correction method ("OncoFreeze", Siemens). Lesion SUVmax and noise levels in the liver were determined.

Our method achieved visually very satisfactory motion compensation and consequently improved NCC for all test scans. Motion artifacts were substantially reduced while maintaining the noise level of the ungated images. The detailed numerical results will be reported.

The proposed framework is suitable for efficient reduction of motion artifacts in PET and is competitive to the OncoFreeze method.

Keywords: motion correction; CNN; unsupervised; registration; PET; gates

  • Lecture (Conference)
    61. Jahrestagung der Deutschen Gesellschaft für Nuklearmedizin, 19.-22.04.2023, Leipzig, Deutschland

Permalink: https://www.hzdr.de/publications/Publ-38203