Parameter Constrained Transfer Learning for Low Dose PET Image Denoising

Yu Gong, Yueyang Teng, Hongming Shan, Taohui Xiao, Ming Li, Guodong Liang, Ge Wang, Shanshan WangarXiv 2019 · arXiv

Abstract

Positron Emission Tomography (PET) is essential for clinical diagnosis but requires radioactive tracers that pose risks to patients. Low-dose PET scans reduce radiation exposure but suffer from significant noise and artifacts. This paper introduces the Parameter-Transferred Wasserstein Generative Adversarial Network (PT-WGAN), a novel framework that combines a hybrid 2D/3D generator architecture with task-specific initialization via transfer learning. PT-WGAN achieves superior denoising performance, reaching a PSNR of 55.048 and an RFSIM of 0.776, while reducing computational parameters by 50% and significantly accelerating training convergence. These results suggest a viable path toward an 80% reduction in patient radiation dose without compromising image fidelity.

Introduction

PET imaging is a cornerstone of modern oncology and cardiology, yet the necessity of radioactive tracer injections limits its application due to safety concerns. While "low-dose" PET scans are safer, they are notoriously noisy, creating artifacts that can lead to misdiagnosis. Existing denoising methods often struggle with a trade-off: 2D deep learning models are fast but ignore 3D spatial context, while pure 3D models are memory-intensive and prone to training instability. Furthermore, standard Generative Adversarial Networks (GANs) often suffer from mode collapse, making them difficult to deploy reliably in medical imaging.

Methods

The proposed PT-WGAN addresses these challenges through three primary innovations:

  1. Hybrid 2D/3D Architecture: The generator utilizes a "sandwich" structure. It begins with 3D convolutional layers to capture volumetric information, transitions to 2D layers for efficient feature processing (reducing parameters by 50% compared to full 3D models), and concludes with 3D deconvolutional layers to restore fine structural details.
  2. Task-Specific Initialization: To overcome GAN instability, the authors employ transfer learning. The generator is first pre-trained as a standard CNN using Mean Squared Error (MSE) loss. These weights are then transferred to initialize the WGAN, providing a "warm start" that prevents divergence.
  3. Refined Loss Function: The model optimizes a combination of Wasserstein distance (with a gradient penalty) and MSE loss. This ensures the output is both visually realistic and anatomically accurate.
The hybrid 2D/3D generator architecture of PT-WGAN.

Datasets and Experimental Setup

The study utilized clinical torso PET scans from nine patients acquired on a Neusoft NeuWise ToF PET/CT scanner. Seven patients were used for training (169,000 volume patches), and two were reserved for testing. Low-dose images were simulated by undersampling raw data to 20% of the original count, with normal-dose scans serving as the ground truth.

Key Results

PT-WGAN was compared against state-of-the-art methods like RED-CNN and CPCE-3D, demonstrating significant improvements across multiple metrics:

  • Quantitative Superiority: The model achieved a peak PSNR of 55.048 and an RFSIM of 0.776, outperforming baseline models.
  • Training Efficiency: Task-specific initialization allowed the network to reach peak performance in just 10 epochs, compared to over 40 epochs for models trained from scratch. This approach also avoided the divergence issues seen in 40% of standard WGAN trials.
  • Qualitative Fidelity: Visual analysis confirmed that PT-WGAN preserves metabolic textures in organs like the spleen and maintains sharp edges in vertebral structures, avoiding the over-smoothing common in pure MSE-based networks.
Visual comparison of denoising results across different models.

Conclusion

PT-WGAN provides a robust and efficient solution for low-dose PET image denoising. By bridging the gap between 2D and 3D feature spaces and utilizing objective-based transfer learning, the framework offers a stable method for generating high-quality medical images. Future work will focus on validating the model with larger clinical cohorts and exploring unsupervised learning techniques to eliminate the need for paired high-dose training data.

Reference: Yu Gong, et al. "Parameter-Transferred Wasserstein Generative Adversarial Network (PT-WGAN) for Low-Dose PET Image Denoising." arXiv:1910.06749v3 [eess.IV].

Summary updated

Abstract

Due to the widespread use of positron emission tomography (PET) in clinical practice, the potential risk of PET-associated radiation dose to patients needs to be minimized. However, with the reduction in the radiation dose, the resultant images may suffer from noise and artifacts that compromise diagnostic performance. In this paper, we propose a parameter-transferred Wasserstein generative adversarial network (PT-WGAN) for low-dose PET image denoising. The contributions of this paper are twofold: i) a PT-WGAN framework is designed to denoise low-dose PET images without compromising structural details, and ii) a task-specific initialization based on transfer learning is developed to train PT-WGAN using trainable parameters transferred from a pretrained model, which significantly improves the training efficiency of PT-WGAN. The experimental results on clinical data show that the proposed network can suppress image noise more effectively while preserving better image fidelity than recently published state-of-the-art methods. We make our code available at https://github.com/90n9-yu/PT-WGAN.
BibTeX
@misc{gong2019parameter,
  author = {Yu Gong and Yueyang Teng and Hongming Shan and Taohui Xiao and Ming Li and Guodong Liang and Ge Wang and Shanshan Wang},
  title = {{Parameter Constrained Transfer Learning for Low Dose PET Image Denoising}},
  year = {2019},
  eprint = {1910.06749},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.06749},
}
Figure