PET Head Motion Estimation Using Supervised Deep Learning with Attention

TL;DR

This work addresses the challenge of head motion in brain PET imaging by introducing DL-HMC++, a supervised deep-learning model with cross-attention that predicts rigid head motion from one-second PET cloud images. The method surpasses multiple state-of-the-art baselines in both motion estimation accuracy and motion-corrected image quality, demonstrated across two scanners (HRRT, mCT) and four tracers ($^{18}$F-FDG,$^{18}$F-FPEB, $^{11}$C-UCB-J, $^{11}$C-LSN3172176) in a cohort of over 280 scans. Central to DL-HMC++ is a cross-attention mechanism that aligns reference and moving PCIs, paired with Deep Normalization and Fusion blocks, yielding accurate six-parameter motion estimates ${\hat{\theta}}=[t_x,t_y,t_z,r_x,r_y,r_z]$. The model supports event-by-event MOLAR reconstruction and shows strong cross-tracer generalization, suggesting practical deployment in clinical PET without external motion-tracking hardware. Overall, the approach significantly reduces motion artifacts and improves quantitative accuracy, enabling broader clinical adoption of PET motion correction.

Abstract

Head movement poses a significant challenge in brain positron emission tomography (PET) imaging, resulting in image artifacts and tracer uptake quantification inaccuracies. Effective head motion estimation and correction are crucial for precise quantitative image analysis and accurate diagnosis of neurological disorders. Hardware-based motion tracking (HMT) has limited applicability in real-world clinical practice. To overcome this limitation, we propose a deep-learning head motion correction approach with cross-attention (DL-HMC++) to predict rigid head motion from one-second 3D PET raw data. DL-HMC++ is trained in a supervised manner by leveraging existing dynamic PET scans with gold-standard motion measurements from external HMT. We evaluate DL-HMC++ on two PET scanners (HRRT and mCT) and four radiotracers (18F-FDG, 18F-FPEB, 11C-UCB-J, and 11C-LSN3172176) to demonstrate the effectiveness and generalization of the approach in large cohort PET studies. Quantitative and qualitative results demonstrate that DL-HMC++ consistently outperforms state-of-the-art data-driven motion estimation methods, producing motion-free images with clear delineation of brain structures and reduced motion artifacts that are indistinguishable from gold-standard HMT. Brain region of interest standard uptake value analysis exhibits average difference ratios between DL-HMC++ and gold-standard HMT to be 1.2 plus-minus 0.5% for HRRT and 0.5 plus-minus 0.2% for mCT. DL-HMC++ demonstrates the potential for data-driven PET head motion correction to remove the burden of HMT, making motion correction accessible to clinical populations beyond research settings. The code is available at https://github.com/maxxxxxxcai/DL-HMC-TMI.

Figures (7)

• Figure 1: DL-HMC++ network architecture. (Top) A shared encoder extracts imaging features from a pair of moving and reference PET cloud images. Then, the extracted features are fed into the cross-attention module to learn the correlation of anatomical features. Deep Normalization and Fusion (DNF) blocks refine the attention features both before and after concatenation. Finally, concatenated attention features are fed into a multi-layer perceptron Regression block to predict motion. (Bottom) Details of the cross-attention module.
• Figure 2: HRRT motion prediction results with $^{18}$F-FDG and $^{11}$C-UCB-J tracers. Each column represents a unique subject from the test set. Rows show rigid transformation parameters (from top to bottom: translation in x, y, z directions and rotation about the x, y, z axes) from gold-standard Vicra (orange), DL-HMC (yellow) and our proposed method (blue) on HRRT. Red boxes indicate time intervals of interest for DL-HMC++ performance.
• Figure 3: Grad-CAM saliency map visualization. Sagittal view from five different time frames of the HRRT testing set during 30 min (1,800 s) PET acquisition. Our proposed DL-HMC++ method more accurately localizes the head anatomy compared to DL-HMC without attention.
• Figure 4: mCT motion prediction results with $^{18}$F-FPEB and $^{11}$C-LSN3172176 tracers. Each column represents a unique subject from the test set. Rows show rigid transformation parameters (from top to bottom: translation in x, y, z directions and rotation about the x, y, z axes) from gold-standard Vicra (orange), DL-HMC (yellow) and our proposed method (blue) on mCT. Red boxes indicate time intervals of interest for DL-HMC++ performance.
• Figure 5: 3D PET Cloud Image (PCI) Dimensions. Example one-second HRRT PET cloud images of different dimensions and resolutions: (top) 32$^3$ voxels; (middle) 64$^3$ voxels; and (bottom) 96$^3$ voxels.