PaperlayerView exact revision on arXiv

Source-linked AI summary

UniPET: a universal network for high-quality PET image denoising across varied dose reduction factors

Zhiwen Yang, Yang Zhou, Haowei Chen, Hui Zhang, Dan Zhao, Bingzheng Wei, Yan Xu

arXiv:2606.11131v1cs.CV

TL;DR

Universal PET denoising must handle varied dose reduction factors without losing clinically important styles and details. UniPET combines style alignment with region-aware learning, achieving state-of-the-art universal denoising while matching DRF-specific models at particular DRFs.

  • Problem

    Universal PET denoising seeks one model that handles varied dose reduction factors, but preliminary universal models struggle to recover fine styles and details across them.

  • Method

    UniPET combines a base denoising network with style alignment for cross-DRF style recovery and region-aware learning focused on stylized regions.

  • Results

    UniPET achieves state-of-the-art universal PET denoising quantitatively, perceptually, and clinically, with performance comparable to DRF-specific models.

  • Takeaways & Limitations

    UniPET adaptively recovers styles across DRFs while supporting high-quality PET denoising across varying, unknown, and out-of-distribution conditions.

  • Takeaways & Limitations

    Small lesions in synthesized images remain noticeably different from full-dose PET, with some lesion-shape distortion because they are underrepresented in the dataset.

Abstract

from arXiv · show

Most existing deep learning-based PET image denoising methods assume a fixed and known dose reduction factor (DRF) for low-dose PET images. However, these methods encounter significant performance degradation when the DRF varies beyond the assumed one in practical applications. To address the challenge posed by varied DRFs, several preliminary studies focus on the task of universal PET image denoising, aiming to train a universal model over low-dose data across DRFs. Nonetheless, these vanilla universal models often struggle with misaligned styles present in different DRF data, leading to the \textit{style elimination issue} with a significant over-smoothing effect. To deal with this issue, we innovatively introduce domain generalization to PET image denoising and propose a universal PET image denoising network (UniPET) to achieve high-quality PET image denoising across diverse DRFs. UniPET comprises two primary innovations: a style alignment network (SAN) and a region-aware learning strategy (RALS). Specifically, SAN utilizes style alignment techniques derived from domain generalization to align and recover styles across different DRFs, ensuring the model's generalizability across various DRFs while effectively preserving styles. Furthermore, to enhance style recovery, RALS distinguishes between flat and stylized regions, exclusively conducting adversarial learning on the latter, thereby more effectively guiding the model's focus towards learning stylized regions. It is demonstrated that our proposed UniPET can adaptively recover different DRF styles and achieve high-quality PET image denoising across DRFs. Comprehensive experiments show that UniPET exhibits comparable performance to individual DRF-specific models at specific DRFs and realizes state-of-the-art performance in universal PET image denoising quantitatively, perceptually, and clinically.

1. Introduction

PET denoising must recover diagnostically useful images from noisy low-dose acquisitions, but DRF-specific models generalize poorly across varying dose reductions and vanilla universal models can erase DRF-specific styles. UniPET addresses this style-elimination problem through domain-generalized style alignment and region-aware learning for high-quality denoising across diverse DRFs.

  • Motivation: Low-dose PET images suffer severe noise and reduced signal-to-noise ratio from limited photon counts, compromising quantitative and diagnostic accuracy.Denoising seeks to recover high-quality PET images while supporting reduced radiation exposure through lower tracer doses or shorter scans.
  • Problem: DRF-specific denoising models have limited generalizability because varying DRFs represent distinct domains with unique styles and induce complex domain shifts.Different DRFs exhibit distinct textures, details, and error distributions despite sharing the same underlying full-dose image.
  • Problem: Vanilla universal models share parameters across DRFs but suffer style elimination, producing generalized outputs that over-smooth DRF-specific textures and details.The issue arises from misaligned styles associated with varying noise levels across DRF data.
  • Proposed Method: UniPET combines a pre-trained base denoising network, style alignment network, and region-aware learning strategy to denoise across diverse DRFs while recovering styles and intricate details.SAN aligns DRF features into a common feature space, while the region-aware strategy prioritizes style recovery and prevents overfitting on flat regions.

2. Related works

Related work covers deep learning for PET denoising, the limitations of DRF-specific models, and preliminary universal approaches, while motivating domain generalization, style modulation, and region-aware learning for preserving clinically important image characteristics.

  • PET imaging: PET visualizes and quantifies in vivo metabolic processes from photon emissions emitted by injected radioactive tracers.Higher-metabolic-rate tissues accumulate more tracer and produce stronger detected emissions, while tracer radiation creates exposure risk.
  • PET denoising: Deep learning methods are widely used for PET denoising, but many target a specific dose reduction factor, such as DRF=4.DRF=4 signifies reduction to 25% of the original dose.
  • PET denoising: DRF-specific denoising methods suffer performance drops when practical applications use dose reduction factors beyond the assumed value.
  • Universal PET denoising: Preliminary universal PET denoising studies train a single model across diverse DRFs, including residual-learning methods that estimate noise residuals adaptively.Chan et al. (2018) reported more consistent liver-region performance across DRFs than DRF-specific methods.
  • Related learning strategies: Domain generalization learns domain-invariant features transferable across domains, while style modulation represents CNN style through feature means and standard deviations.Region-aware learning allocates more computation to selected regions; this work applies GAN training exclusively to stylized regions to emphasize style recovery.

3. Methodology

UniPET trains a single universal model for low-dose PET images across different dose reduction factors using a pre-trained base denoising network, style alignment network, and region-aware learning strategy. SAN aligns and recovers DRF-dependent styles, while RALS focuses adversarial learning on stylized regions to improve style recovery.

  • UniPET overview: UniPET comprises a pre-trained base denoising network, style alignment network, and region-aware learning strategy for universal PET denoising across DRFs.The base network produces estimates from different DRF inputs, while SAN aligns styles and RALS treats flat and stylized regions differently.
  • Base denoising network: BDN uses residual learning and multi-DRF pre-training to provide coarse-grained denoising capabilities across dose reduction factors.It estimates input residuals through convolutional and residual-block processing before synthesizing the denoised image.
  • Style alignment network: SAN addresses style elimination by embedding domain knowledge, modulating BDN features, and optimizing a style alignment loss across DRF styles.The framework dynamically aligns diverse domain styles and unifies distinct domain features within BDN.
  • Style alignment network: SAN derives style codes from domain knowledge embeddings and uses them to dynamically modulate convolutional features within BDN.Hierarchical embeddings can modulate sequential groups of BDN residual blocks, enabling flexible style manipulation.
  • Region-aware learning strategy: RALS restricts GAN training to stylized regions because flat regions dominate PET images and imbalance hampers effective learning of style information.Narrowing GAN training from the entire image to stylized regions lets the network focus on style recovery without changing GAN fundamentals.

4. Experimental setup

UniPET is evaluated on private in-distribution and out-of-distribution datasets plus the public Bern dataset, using quantitative, perceptual, clinical, and statistical assessments. Training uses UPID-Base data spanning multiple dose reduction factors, while OOD tests vary the DRF or acquisition center.

  • Datasets: Four whole-body datasets comprise UPID-Base, UPID-OOD-DRF, UPID-OOD-Center, and the public Bern dataset.UPID-Base and Bern support training, validation, and testing; the two UPID-OOD datasets evaluate out-of-distribution generalization.
  • Evaluation metrics: Evaluation covers PSNR and SSIM for intensity recovery and structural preservation, LPIPS for perceptual quality and textural-detail recovery, and comparative reader scoring.Four radiologists assess six methods on UPID-Base testing images at DRF=12, with 10 of 15 patients having lesions.
  • Evaluation metrics: Clinical evaluation measures SUV errors in blood pool, liver, and lesion ROIs, and assesses lesion detection using radiologist annotations, F1-scores, and classifier-derived ROC curves.ROC proximity to full-dose data is used as an indicator of image quality and lesion detectability.
  • Statistical testing: Paired t-tests compare methods for each metric using α = 0.05, with Bonferroni correction applied within each comparison group.The correction controls the groupwise error rate by dividing α by the number of comparisons in the group.
  • Data preparation: UniPET training uses UPID-Base data at DRFs 1, 2, 3, 6, and 12, with DRFs 2, 3, 6, and 12 repeated three times per training epoch.Images are converted to SUVs and processed in the SUV domain; training data are clipped to [0, 20] and normalized to [0, 1].

5. Experimental results

UniPET mitigates style elimination and over-smoothing, matching DRF-specific models while outperforming universal and other comparison methods across quantitative, perceptual, and clinical evaluations. SAN and RALS each improve performance, with their combination strengthening style and lesion recovery across architectures and clinical regions.

  • Universal versus DRF-specific models: UniPET outperforms BDN across DRFs, gaining 0.47 dB in PSNR, 0.005 in SSIM, and reducing LPIPS by 0.009 on average.For each individual DRF, UniPET achieves results comparable to DRF-specific models.
  • Visual and clinical evaluation: UniPET consistently preserves anatomical structures, fine textures, and small-lesion contrast across DRFs, avoiding the over-smoothing observed in BDN and other methods.It achieves the best lesion detection performance across DRFs, with F1-score > 0.9.
  • Comparison with state-of-the-art methods: On the evaluated datasets, UniPET surpasses the second-best Spach Transformer by 0.44 dB in PSNR, 0.006 in SSIM, and 0.002 in LPIPS.These improvements demonstrate strong generalizability across varying DRFs.
  • Ablation studies: SAN improves all four evaluation metrics, while RALS primarily enhances lesion-region MAE and LPIPS; together they produce the best lesion recovery.The combined method decreases lesion-region MAE by 0.024 versus the baseline and raises DRF = 12 F1-score by 0.10.
  • Cross-architecture evaluation: Integrating SAN and RALS consistently improves all four evaluation metrics across ResNet-, DenseNet-, and UNet-based denoising networks.For the ResNet-based BDN, lesion-region MAE is reduced by 0.024.
  • Region-aware learning analysis: RALS reduces MAE versus GAN by 0.002 in the blood pool, 0.005 in the liver, and 0.004 in lesions, while stylized-region guidance approaches manually annotated ROI guidance.The selected threshold δ = 0.001 generally covers the three clinical ROIs and other clinically important anatomical regions.

6. Discussion

UniPET uses domain generalization through SAN and RALS to align and recover styles while focusing learning on stylized regions, and is evaluated across four datasets and three challenging scenarios. Despite strong performance, limitations remain in small-lesion fidelity, OOD generalization, and the clinical acceptability of evaluation metrics.

  • Contributions: UniPET applies domain generalization through SAN to learn domain-transferable features and align and recover styles across different DRFs.This is intended to support model generalization and high-quality denoising.
  • Contributions: RALS guides style recovery by focusing learning on stylized regions.The strategy is introduced to facilitate style recovery.
  • Evaluation: UniPET is validated on four datasets, including private clinical data and public benchmarks, across varying-DRF denoising, OOD data, and another challenging scenario.The supplied passage names the first two scenarios but truncates the third.
  • Limitations: Small lesions in synthesized images remain visibly different from full-dose PET, with lesion-shape distortion attributed primarily to their very small size.The discrepancy is documented in Figures 7–9.
  • Limitations: UniPET still risks OOD failure, showing over-smoothing and inadequate lesion-contrast recovery on UPID-OOD-Center.This indicates that domain generalization does not eliminate risks on OOD data.
  • Future directions: FGSM slightly decreases performance on UPID-Base but improves performance on UPID-OOD-DRF and UPID-OOD-Center, suggesting protective techniques can enhance generalizability.The supplied passage reports these directional effects without quantitative values.
  • Limitations: The clinical acceptability of quantitative, perceptual, and clinical evaluation measures remains unclear relative to real-world clinical standards.Future work could investigate whether these measures align with clinical standards and whether UniPET is clinically acceptable.

7. Conclusion

The paper addresses universal PET image denoising across different dose reduction factors by introducing UniPET based on domain generalization. UniPET combines coarse-grained denoising with style alignment and recovery across DRFs.

  • Conclusion: UniPET targets recovery of low-dose PET images across different dose reduction factors.The work focuses on universal PET image denoising rather than a fixed DRF.
  • Conclusion: The method harnesses domain generalization to address universal PET image denoising.Domain generalization provides the conceptual basis for UniPET.
  • Conclusion: UniPET consists of a pre-trained base denoising network for coarse-grained denoising and a style alignment network for fine-grained style alignment and recovery across DRFs.The passage identifies BDN and SAN as components of the proposed network.
Loading 2606.11131v1…