TriDo-Former: A Triple-Domain Transformer for Direct PET Reconstruction from Low-Dose Sinograms

Jiaqi Cui, Pinxian Zeng, Xinyi Zeng, Peng Wang, Xi Wu, Jiliu Zhou, Yan Wang, Dinggang Shen

Introduction

Positron Emission Tomography (PET) is a cornerstone of modern diagnostic imaging, providing critical metabolic information for oncology, neurology, and cardiology. However, the procedure requires the injection of radioactive tracers, posing potential health risks to patients. While reducing the tracer dose (Low-Dose PET or LPET) minimizes radiation exposure, it significantly degrades the signal-to-noise ratio in the raw data, known as sinograms. This degradation typically results in reconstructed images characterized by severe noise, blurring, and the loss of fine structural details.

Traditional deep learning approaches to this problem generally follow two paths: indirect reconstruction (denoising an already reconstructed image) or direct reconstruction (mapping raw sinograms to final images). Indirect methods often fail because the initial reconstruction process—such as Filtered Back Projection (FBP) or Ordered Subset Expectation Maximization (OSEM)—can permanently discard high-frequency information. Direct methods, while more promising, have historically relied on Convolutional Neural Networks (CNNs). CNNs are limited by their local receptive fields, making them less effective at capturing the global structural dependencies of an organ and often leading to "spectral bias," where high-frequency textures and sharp boundaries are smoothed over.

To address these challenges, we present TriDo-Former, a triple-domain transformer architecture designed for direct, high-quality PET reconstruction from low-dose sinograms.

Methodology

TriDo-Former utilizes a cascaded transformer framework that operates across three distinct domains: the sinogram domain, the spatial image domain, and the frequency (spectral) domain. The architecture is divided into two primary stages.

1. Sinogram Enhancement Transformer (SE-Former)

The first stage focuses on denoising the raw data before it is transformed into an image. Unlike standard Vision Transformers (ViTs) that split data into 2D patches, the SE-Former employs a physics-informed 1D patching strategy. By treating the sinogram as a series of 1D projection view angles (rows), the model preserves the inherent geometric structure of the projection data. This prevents noise from leaking across different projection angles, ensuring a cleaner input for the reconstruction stage.

2. Spatial-Spectral Reconstruction Transformer (SSR-Former)

The second stage performs the complex mapping from the denoised sinogram to the 2D PET image. This stage utilizes a U-shaped architecture integrated with Swin Transformer blocks, which allow the model to capture long-range spatial relationships across the entire image.

Crucially, the SSR-Former incorporates a Global Frequency Parser (GFP). The GFP applies a Discrete Fourier Transform (DFT) to the feature maps, allowing the model to operate in the frequency domain. By using learnable filters to adjust specific frequency components, the GFP specifically targets the restoration of high-frequency edges and textures that are often lost in low-dose scans. The features are then returned to the spatial domain via an Inverse DFT, resulting in images that are both structurally accurate and visually sharp.

Experimental Results

Dataset and Evaluation

The TriDo-Former was evaluated using a clinical human brain dataset comprising 16 subjects (8 Normal Control and 8 Mild Cognitive Impairment). The model was trained to map 3-minute "low-dose" scans to 12-minute "standard-dose" ground truth images.

Quantitative Performance

TriDo-Former established a new state-of-the-art (SOTA) for direct PET reconstruction. It achieved a Peak Signal-to-Noise Ratio (PSNR) of 24.912 dB for NC subjects and 25.288 dB for MCI subjects, representing a ~0.6 dB improvement over previous leading methods like LCPR-Net. The Structural Similarity Index (SSIM) reached 0.987, indicating near-perfect structural fidelity to standard-dose scans.

Clinical Utility

To validate the real-world impact of the reconstruction quality, the authors utilized a pre-trained Alzheimer’s disease classifier. Images reconstructed by TriDo-Former achieved a diagnostic accuracy of 88.6%, remarkably close to the 90% accuracy achieved using real standard-dose images. In comparison, traditional OSEM reconstruction only achieved 79.2% accuracy, highlighting the clinical necessity of advanced reconstruction frameworks.

Discussion and Efficiency

One of the most significant advantages of TriDo-Former is its efficiency. Despite the complexity of the triple-domain approach, the model is more lightweight than many CNN-based competitors. With only 38M parameters and 16.05 GFLOPs, it is significantly more efficient than models like 3D-cGAN (127M parameters) or LCPR-Net (77M parameters), making it a more viable candidate for integration into clinical hardware.

Conclusion

TriDo-Former demonstrates that by explicitly modeling the sinogram, spatial, and frequency domains, transformers can overcome the limitations of local-bias in CNNs for PET reconstruction. While current work focuses on 2D slices, future iterations aim to extend this architecture into native 3D transformers to further enhance volumetric consistency. This research marks a significant step toward safer, lower-dose PET imaging without sacrificing diagnostic precision.

Summary updated

BibTeX
@inproceedings{cui2023trido,
  author = {Jiaqi Cui and Pinxian Zeng and Xinyi Zeng and Peng Wang and Xi Wu and Jiliu Zhou and Yan Wang and Dinggang Shen},
  title = {{TriDo-Former: A Triple-Domain Transformer for Direct PET Reconstruction from Low-Dose Sinograms}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43999-5\_18},
}
Figure