Unsupervised inter-frame motion correction for whole-body dynamic PET using convolutional long short-term memory in a...

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Subject motion in whole-body dynamic PET scans causes inter-frame mismatch, degrading parametric imaging quality.
• Conventional non-rigid registration methods for motion correction are computationally intensive and time-consuming.
• Deep learning offers potential for fast and accurate motion correction but requires investigation for dynamic PET and whole-body applications.

Purpose of the Study:

• To develop an unsupervised deep learning framework for automatic correction of inter-frame body motion in dynamic PET imaging.
• To evaluate the framework's performance in motion estimation and its impact on quantitative parametric imaging.

Main Methods:

• Developed a motion estimation network using a convolutional neural network with a convolutional long short-term memory layer.
• Utilized dynamic temporal and spatial features for motion estimation.
• Dataset comprised 27 subjects with 90-min FDG whole-body dynamic PET scans, evaluated using motion simulation and 9-fold cross-validation.

Main Results:

• The proposed network achieved the lowest motion prediction error compared to traditional and deep learning baselines.
• Demonstrated superior spatial alignment of parametric Ki and Vb images, significantly reducing parametric fitting errors.
• Inference time was approximately 460 times faster than conventional registration methods after training.

Conclusions:

• The unsupervised deep learning framework effectively corrects inter-frame motion in whole-body dynamic PET scans.
• The method enhances image quality, reduces errors, and offers significant speed advantages for clinical translation.
• Improved motion correction has the potential to enhance downstream analysis, aiding in distinguishing malignant from benign lesions.