Scholarlink

<p>Positron Emission Tomography (PET) diagnostic precision is often compromised by low spatial resolution. Deep learning restoration models tend to sacrifice quantitative accuracy for visual sharpness, and most are trained on a single fixed degradation profile, limiting generalization across scanners. This paper presents a metabolically faithful 3D restoration framework pairing a volumetric extension of SwinFIR with two innovations: (1) a composite metabolic-aware loss enforcing structural, distributional, and frequency-domain agreement with the ground truth, and (2) a stochastic degradation augmentation strategy that randomizes point spread function parameters, voxel sampling, and counting noise during training, exposing the model to a distribution of simulated scanner-like degradations rather than a single fixed simulation. Evaluated on NeuroEXPLORER data, the proposed method outperforms baselines with a Structural Similarity Index Measure (SSIM) of 0.843, Peak Signal to Noise Ratio (PSNR) of 27.08 dB, and Normalized Root Mean Squared Error (NRMSE) of 0.117, while maintaining metabolic fidelity (Concordance Correlation Coefficient (CCC) 0.948, Wasserstein distance 0.018). Ablation experiments confirm that stochastic degradation augmentation improves robustness over fixed-profile training. The framework recovers anatomical detail with only small but measurable regional SUVR biases (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\le \)</EquationSource> </InlineEquation>3.3%) in selected cortical regions, which are symmetric across diagnostic groups.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Metabolically Faithful 3D PET Restoration via Volumetric Swin Transformers

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Ovidijus Grigas,
  • Rytis Maskeliūnas

摘要

Positron Emission Tomography (PET) diagnostic precision is often compromised by low spatial resolution. Deep learning restoration models tend to sacrifice quantitative accuracy for visual sharpness, and most are trained on a single fixed degradation profile, limiting generalization across scanners. This paper presents a metabolically faithful 3D restoration framework pairing a volumetric extension of SwinFIR with two innovations: (1) a composite metabolic-aware loss enforcing structural, distributional, and frequency-domain agreement with the ground truth, and (2) a stochastic degradation augmentation strategy that randomizes point spread function parameters, voxel sampling, and counting noise during training, exposing the model to a distribution of simulated scanner-like degradations rather than a single fixed simulation. Evaluated on NeuroEXPLORER data, the proposed method outperforms baselines with a Structural Similarity Index Measure (SSIM) of 0.843, Peak Signal to Noise Ratio (PSNR) of 27.08 dB, and Normalized Root Mean Squared Error (NRMSE) of 0.117, while maintaining metabolic fidelity (Concordance Correlation Coefficient (CCC) 0.948, Wasserstein distance 0.018). Ablation experiments confirm that stochastic degradation augmentation improves robustness over fixed-profile training. The framework recovers anatomical detail with only small but measurable regional SUVR biases ( \(\le \) 3.3%) in selected cortical regions, which are symmetric across diagnostic groups.