TL;DR
This paper introduces MDPET, a novel unified adversarial network that simultaneously corrects motion blur and denoises low-dose PET images. By integrating motion estimation and denoising into a single end-to-end framework, MDPET achieves superior image quality and more accurate motion estimation compared to existing methods, which is crucial for improving diagnostic accuracy in low-dose PET imaging.
What Problem Does This Paper Address?
Positron Emission Tomography (PET) is a vital medical imaging technique. However, to minimize radiation exposure, low-dose PET scans are often employed. These low-dose scans are inherently noisy, and when combined with respiratory motion (which is common in thoracic and abdominal imaging), the resulting images can be significantly blurred and difficult to interpret. Respiratory gating is used to mitigate motion blur by dividing the scan into phases, but this further reduces the number of events per phase, exacerbating noise. Existing methods for motion correction and denoising are often separate, and applying them to noisy, low-dose gated PET data can lead to inaccurate motion estimation and poor final image quality. This paper aims to address this by developing a unified approach that can effectively handle both motion and noise simultaneously in low-dose gated PET.
Methods
The core of the proposed method is MDPET, a unified adversarial network. It comprises two main components:
- Temporal Siamese Pyramid Network (TSP-Net): This module is responsible for motion estimation. It uses shared-weight Siamese Pyramid Networks (SP-Nets) as basic units, each designed to learn coarse-to-fine features from pairs of low-dose gated images. These SP-Nets are then integrated with a Bidirectional Convolutional Long Short-Term Memory (BiConvLSTM) layer. This recurrent component allows the network to learn temporal motion patterns across different respiratory gates, improving motion estimation accuracy by considering the entire respiratory cycle. The TSP-Net predicts transformation fields that align the low-dose gated images to a reference gate (typically end-expiration).
- Denoising Network: After the low-dose gated images are aligned by the TSP-Net, they are averaged to create a motion-compensated image. This averaged image, which still contains noise due to the low-dose nature of the original data, is then fed into a denoising network. The authors adapt a UNet architecture for this purpose, known for its effectiveness in image denoising.
- Adversarial Training: The entire network is trained in an end-to-end fashion using a combination of registration loss (to supervise motion estimation), denoising loss (to supervise the denoising output), and an adversarial loss. A discriminator is used to encourage the denoising network to produce realistic, high-quality images.
What's novel is the unification of these two critical tasks (motion correction and denoising) into a single adversarial framework, allowing for mutual benefit and improved performance over separate, sequential processing.
Datasets
The study utilized a dataset of 28 pancreas 18F-FPDTBZ PET/CT studies from the Yale PET Center. These studies were acquired using Siemens Biograph mCT scanners and an AZ-733V respiratory gating system. The data was processed to generate 6 respiratory gates per study. High-dose images were simulated by using 100% of the listmode data, while low-dose images were created using only 1.5% of the data. The authors performed four-fold cross-validation, with 21 studies for training and 7 for testing in each fold.
Key Insights and Ideas
- Unified Framework Advantage: The most significant insight is that jointly learning motion estimation and denoising in an end-to-end adversarial network leads to better results than separate, sequential approaches. The denoising component can provide feedback to the motion estimation, helping it to be more robust to noise.
- Temporal Motion Features Matter: Incorporating a BiConvLSTM in the motion estimation module (TSP-Net) significantly improves accuracy by leveraging temporal information across respiratory gates, which is crucial for capturing complex respiratory motion patterns.
- Noise-Robust Motion Estimation: Traditional methods struggle with noisy low-dose PET data, leading to increased registration errors. MDPET's ability to estimate motion directly from noisy images, aided by its unified denoising capability, is a key advantage.
- Anatomical Detail Preservation: MDPET not only reduces noise but also better preserves fine anatomical details, such as liver veins, which are often lost in averaged images from less effective motion correction methods.
Key Results
- Motion Estimation: MDPET achieved superior motion estimation performance compared to established methods like VoxelMorph (VM), Non-Rigid B-spline Registration (NRB), and their previous Siamese Adversarial Network (SAN). For example, it significantly reduced Normalized Mean Absolute Error (NMAE) for gates with large motion displacement (e.g., Gate 1 and Gate 6).
- Denoising: The unified denoising output from MDPET demonstrated superior performance with a mean NMAE of 0.088, SSIM of 0.966, and PSNR of 32.28, outperforming two-stage processing methods.
- Speed: MDPET achieved a fast GPU inference time of 0.54 seconds per study, significantly faster than iterative methods like NRB and more efficient than methods requiring multiple inferences per gate.
- Ablation Studies: Ablation studies confirmed the benefits of BiConvLSTM for motion estimation and adversarial learning for denoising.
Limitations and Future Work
- Ground Truth Distinction: While MDPET significantly improves image quality, the denoised results are still not as distinct as the original high-dose ground truth.
- Denoising Network Interchangeability: The authors note that the UNet denoising network can be replaced with other advanced architectures to potentially further improve image quality.
- Perceptual Loss for 3D: Incorporating perceptual loss, which is effective in 2D, could be beneficial but requires adaptation for 3D imaging data.
- Intra-Gate Motion: The current work primarily addresses inter-gate motion. Future work could explore methods to mitigate intra-gate motion blur, which can affect both motion estimation and the final image quality. Event-by-event listmode correction is suggested.
- Data Augmentation and Reconstruction: Further investigation into data augmentation strategies and the impact of different image reconstruction parameters (e.g., number of iterations) is warranted.
- Clinical Applications: The authors suggest potential applications in abdominal imaging, continuous bed motion (CBM) for dynamic PET, and deviceless low-dose gating reconstruction.
Summary updated