Scholarlink

Purpose <p>Positron‑emission tomography (PET) is indispensable for metabolic tumour imaging, yet clinical pressure to shorten scan times or lower injected activity often produces noisy images that obscure lesions and bias quantitative parameters such as standardized uptake value (SUV). Recently, transformer‑based architectures have shown superior capacity to capture long‑range 3‑D context; however, their application to PET denoising and integration into freely available clinical software remain limited. In this study we aimed to develop and evaluate a transformer-based denoising model for F-18 FDG PET images using SwinUNETR architecture, integrating tumor-sensitive loss functions. Additionally, a secondary objective was to implement an open-source module for the 3D Slicer platform to enable clinical deployment of the trained models.</p> Methods <p>This retrospective study included 90 patients who underwent whole-body F-18 FDG PET/CT. Noisy PET images were reconstructed from 30-second list-mode data, while full-count reference images were reconstructed from 90-second data. A SwinUNETR-based residual learning model was developed in three configurations: standard Charbonnier loss (std_char), and two tumor-weighted loss models (x19w1.5 and x14w3) that penalized SUV underestimation more severely in tumor regions. Models were trained on 64 × 64 × 64 voxel patches and evaluated using MSE, MAE, inverted SSIM (I-SSIM), PSNR, and Bland–Altman plots of tumor SUVmax and SUVmean.</p> Results <p>All models demonstrated a significant reduction in image noise when compared to the noisy input data. Among them, the std_char model provided the greatest enhancement in global image quality metrics, achieving a MSE reduction of 28.97%, a MAE reduction of 18.61%, a 37.83% decrease in I-SSIM, and an increase in PSNR by 1.50 dB. In contrast, the tumor-weighted models x19w1.5 and x14w3 exhibited superior performance in preserving tumor quantification accuracy. The x19w1.5 model had a tumor SUVmax bias of -0.28&#xa0;g/ and an SUVmean bias of -0.03&#xa0;g/mL, while the x14w3 model achieved an SUVmax bias of -0.11&#xa0;g/mL and an SUVmean bias of + 0.04&#xa0;g/mL.</p> Conclusion <p>Transformer-based denoising with tumor-aware loss functions significantly improves PET image quality while preserving tumor SUV values. The proposed models, particularly x14w3, provide a clinically viable balance of denoising performance and quantitative accuracy. The open-source 3D Slicer module enables accessible clinical integration and future research use.</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

Denoising of PET with SwinUNETR neural networks: impact of tumor oriented loss function, denoising module for 3D slicer

  • 期刊论文

  • 发表日期：

    2026年2月8日

查看全文
  • Burak Demir,
  • Merve Atalay,
  • Hatice Kubra Yurtcu,
  • Fikret Ertek

摘要

Purpose

Positron‑emission tomography (PET) is indispensable for metabolic tumour imaging, yet clinical pressure to shorten scan times or lower injected activity often produces noisy images that obscure lesions and bias quantitative parameters such as standardized uptake value (SUV). Recently, transformer‑based architectures have shown superior capacity to capture long‑range 3‑D context; however, their application to PET denoising and integration into freely available clinical software remain limited. In this study we aimed to develop and evaluate a transformer-based denoising model for F-18 FDG PET images using SwinUNETR architecture, integrating tumor-sensitive loss functions. Additionally, a secondary objective was to implement an open-source module for the 3D Slicer platform to enable clinical deployment of the trained models.

Methods

This retrospective study included 90 patients who underwent whole-body F-18 FDG PET/CT. Noisy PET images were reconstructed from 30-second list-mode data, while full-count reference images were reconstructed from 90-second data. A SwinUNETR-based residual learning model was developed in three configurations: standard Charbonnier loss (std_char), and two tumor-weighted loss models (x19w1.5 and x14w3) that penalized SUV underestimation more severely in tumor regions. Models were trained on 64 × 64 × 64 voxel patches and evaluated using MSE, MAE, inverted SSIM (I-SSIM), PSNR, and Bland–Altman plots of tumor SUVmax and SUVmean.

Results

All models demonstrated a significant reduction in image noise when compared to the noisy input data. Among them, the std_char model provided the greatest enhancement in global image quality metrics, achieving a MSE reduction of 28.97%, a MAE reduction of 18.61%, a 37.83% decrease in I-SSIM, and an increase in PSNR by 1.50 dB. In contrast, the tumor-weighted models x19w1.5 and x14w3 exhibited superior performance in preserving tumor quantification accuracy. The x19w1.5 model had a tumor SUVmax bias of -0.28 g/ and an SUVmean bias of -0.03 g/mL, while the x14w3 model achieved an SUVmax bias of -0.11 g/mL and an SUVmean bias of + 0.04 g/mL.

Conclusion

Transformer-based denoising with tumor-aware loss functions significantly improves PET image quality while preserving tumor SUV values. The proposed models, particularly x14w3, provide a clinically viable balance of denoising performance and quantitative accuracy. The open-source 3D Slicer module enables accessible clinical integration and future research use.