Two-Phase Multi-Dose-Level Pet Image Reconstruction with Dose Level Awareness

Yuchen Fei, Yanmei Luo, Yan Wang, Jiaqi Cui, Yuanyuan Xu, Jiliu Zhou, Dinggang Shen

Introduction

Positron Emission Tomography (PET) is a cornerstone of modern medical imaging, providing vital metabolic information for oncology, neurology, and cardiology. However, the procedure requires the injection of radioactive tracers, posing potential health risks to patients. Adhering to the ALARA (As Low As Reasonably Achievable) principle, clinicians strive to minimize radiation exposure. This reduction leads to Low-dose PET (LPET) imaging, which is unfortunately characterized by significant Poisson noise and streak artifacts that can obscure clinical features and impede accurate diagnosis.

While deep learning has shown immense promise in denoising LPET images, most existing models are optimized for a single, fixed Dose Reduction Factor (DRF). In clinical practice, dose levels can vary significantly. Standard models trained on mixed-dose datasets often struggle to generalize because they lack the context to distinguish between varying noise intensities. This paper presents a novel two-phase framework designed to achieve high-quality reconstruction across multiple dose levels by incorporating explicit dose awareness.

Methodology

The researchers propose a coarse-to-fine reconstruction strategy that prioritizes the model's understanding of the input's noise characteristics before attempting full image restoration.

Phase 1: Dose-Aware Pre-training

The first phase focuses on feature extraction through a specialized PretrainNet based on the U-Net architecture. Unlike standard denoising models, PretrainNet is tasked with two simultaneous objectives:

  1. Dose Classification: The model must identify the specific DRF (e.g., 1/20, 1/50, or 1/100) of the input image.
  2. Self-Reconstruction: The model attempts to reconstruct the noisy input itself.

By forcing the encoder to categorize the dose level while preserving the structural integrity of the noisy input, the framework learns "dose-aware" features. These features capture the specific statistical patterns associated with different radiation levels, providing a more robust foundation for the subsequent reconstruction phase.

Phase 2: Coarse-to-Fine Reconstruction

The second phase utilizes the pre-trained weights to perform the actual image restoration through two sequential networks:

  • CPNet (Coarse Prediction Network): Initialized with the weights from the Phase 1 encoder, CPNet generates an initial estimate of the Standard-dose PET (SPET) image. This network handles the primary denoising and structural recovery.
  • RefineNet: A shallower network designed for residual learning. It takes both the coarse prediction and the original LPET image as inputs to estimate the "residual"—the fine-grained textures and high-frequency details that may have been blurred during the coarse prediction step.

The final output is the summation of the coarse prediction and the predicted residual, resulting in a sharp, high-fidelity PET image.

Experimental Setup

The framework was evaluated using the MICCAI 2022 Ultra-low Dose PET Imaging Challenge Dataset. This dataset includes scans from 206 subjects obtained via a Siemens Biograph Vision Quadra system. The data was split into 170 subjects for training and 36 for testing.

The study specifically targeted three challenging Dose Reduction Factors: DRF 20, DRF 50, and DRF 100. The model processed 2D slices with a resolution of 128 x 128 pixels extracted from the 3D PET volumes.

Results and Discussion

The proposed method demonstrated superior performance compared to several state-of-the-art (SOTA) benchmarks, including Auto-context, StackGAN, and AR-GAN.

Quantitative Performance

The model showed consistent improvements across all primary metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Normalized Mean Square Error (NMSE). Notably, at the most extreme noise level (DRF 100), the proposed model achieved a PSNR of 30.187 dB, outperforming the baseline U-Net (29.608 dB).

One significant finding was the model's stability. While some Generative Adversarial Network (GAN) approaches showed degraded performance when trained on mixed-dose data—sometimes performing worse than a simple U-Net—the dose-aware framework maintained steady gains across all DRF categories.

Qualitative Analysis

Visual error maps confirmed that the two-phase approach produced the smallest deviations from the ground-truth standard-dose images. The coarse-to-fine strategy successfully preserved sharper anatomical boundaries and more accurate tracer distribution patterns compared to single-phase models, which often produced overly smoothed results at high DRFs.

Conclusion and Future Work

This research highlights that "dose awareness" is a critical component for multi-dose PET image reconstruction. By explicitly teaching the model to recognize noise contexts through a pre-training phase, the framework achieves a level of versatility and accuracy that standard models lack.

Future iterations of this work could expand the framework from 2D slices to 3D convolutions to better leverage spatial context. Additionally, moving toward "blind" reconstruction—where the model handles a continuous range of dose levels rather than discrete categories—would further enhance its clinical utility. Finally, formal clinical reader studies will be necessary to confirm that these quantitative improvements translate into better diagnostic outcomes for radiologists.

Summary updated

BibTeX
@inproceedings{fei2024two,
  author = {Yuchen Fei and Yanmei Luo and Yan Wang and Jiaqi Cui and Yuanyuan Xu and Jiliu Zhou and Dinggang Shen},
  title = {{Two-Phase Multi-Dose-Level Pet Image Reconstruction with Dose Level Awareness}},
  booktitle = {ISBI},
  year = {2024},
  doi = {10.1109/isbi56570.2024.10635190},
}
Figure