Abstract
To acquire high-quality positron emission tomography (PET) images while reducing the radiation tracer dose, numerous efforts have been devoted to reconstructing standard-dose PET (SPET) images from low-dose PET (LPET). However, the success of current fully-supervised approaches relies on abundant paired LPET and SPET images, which are often unavailable in clinic. Moreover, these methods often mix the dose-invariant content with dose level-related dose-specific details during reconstruction, resulting in distorted images. To alleviate these problems, we propose a two-stage Semi-Supervised SPET reconstruction framework, namely S3PET, to accommodate the training of abundant unpaired and limited paired SPET and LPET images. Our S3PET involves an unsupervised pre-training stage (Stage I) to extract representations from unpaired images, and a supervised dose-aware reconstruction stage (Stage II) to achieve LPET-to-SPET reconstruction by transferring the dose-specific knowledge between paired images. Specifically, in stage I, two independent dose-specific masked autoencoders (DsMAEs) are adopted to comprehensively understand the unpaired SPET and LPET images. Then, in Stage II, the pre-trained DsMAEs are further finetuned using paired images. To prevent distortions in both content and details, we introduce a dose knowledge decouple module to disentangle the respective dose-specific and dose-invariant knowledge of LPET and SPET, and a dose-specific knowledge learning module to transfer the dose-specific information from SPET to LPET, thereby achieving high-quality SPET reconstruction from LPET images. Experiments demonstrate that our S3PET achieves state-of-the-art performance quantitatively and qualitatively.
Introduction
Positron Emission Tomography (PET) is a cornerstone of medical diagnosis, but high-quality imaging requires injecting patients with significant doses of radioactive tracers, posing health risks. While "low-dose" PET (LPET) is safer, the resulting images are often noisy and blurry, making them difficult for clinicians to interpret accurately.
Deep learning has shown promise in reconstructing standard-dose (SPET) quality from LPET scans. However, most current models are fully supervised, requiring massive datasets of perfectly paired LPET and SPET scans—data that is notoriously difficult to collect in clinical settings. Furthermore, existing models often struggle to distinguish between the patient's actual anatomy (dose-invariant content) and the noise/blur caused by the low radiation (dose-specific details), leading to reconstructed images with structural distortions.
Methods
The S3PET framework is a two-stage semi-supervised approach designed to maximize the utility of both abundant unpaired clinical data and limited paired datasets:
- Stage I: Unsupervised Pre-training: The model uses two independent Dose-specific Masked Autoencoders (DsMAEs). One is trained on unpaired SPET images and the other on unpaired LPET images. By masking parts of the image and forcing the network to reconstruct them, the encoders learn deep, dose-specific feature representations without needing labels.
- Stage II: Supervised Dose-aware Reconstruction: Using a small set of paired data, the model is fine-tuned with two specialized modules:
- Dose Knowledge Decouple (DKD): This module disentangles image embeddings into "content" tokens (anatomy) and "dose-specific" tokens (noise/detail level).
- Dose-specific Knowledge Learning (DKL) & Token Swap: The core innovation where the model swaps the dose-specific tokens of an LPET image with those learned from an SPET image. This effectively "upgrades" the noise profile of the low-dose scan while keeping the patient's anatomy intact.
S3PET utilizes a hybrid architecture where Transformer-based encoders capture global context for feature disentanglement, while CNN-based decoders handle the dense pixel-level reconstruction to maintain sharp edges and local textures.
Results and Discussion
The model was evaluated on the Ultra-low Dose PET Imaging Challenge (UDPET) dataset, featuring LPET images simulated at 1% and 2% of the original dose. S3PET achieved state-of-the-art (SOTA) results across all primary metrics:
- PSNR (Peak Signal-to-Noise Ratio): Reached 25.159 dB, outperforming the previous leading semi-supervised method (SemiPET) by 0.44 dB.
- SSIM (Structural Similarity): Achieved 0.955, indicating very high structural fidelity to the ground truth.
Visual comparisons demonstrate that S3PET produces images with much clearer boundaries and significantly fewer artifacts than GAN-based approaches, which often produce "blotchy" or overly smoothed results. By treating the reconstruction as a "token swap" rather than simple denoising, the model avoids the common pitfall of distorting anatomical structures.
Conclusion
S3PET demonstrates that semi-supervised learning can effectively bridge the gap between low-dose safety and standard-dose image quality. By decoupling anatomical content from dose-specific noise, the framework achieves high-fidelity reconstruction even with limited paired data. Future work will explore native 3D architectures to improve volumetric consistency and validate the model on real-world, non-simulated low-dose acquisitions.
Summary updated