Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Text controllable PET denoising

Xuehua Ye, Hongxu Yang, Adam J. Schwarz

TL;DR

This work tackles the noise and variability inherent in low-count PET imaging by introducing a text-guided denoising framework. The method leverages CLIP to embed count-level semantics into both the encoder and decoder of a U-Net denoiser, enabling controllable restoration from low-count inputs. Training data consist of low-count PET slices paired with higher-count ground truth, with count conditions randomly sampled to promote generalization, and optimization is performed with AdamW. Results show improved SSIM and PSNR over baselines and demonstrate the ability to synthesize higher-count images from low-count inputs, offering potential reductions in radiation exposure and scan time in clinical practice.

Abstract

Positron Emission Tomography (PET) imaging is a vital tool in medical diagnostics, offering detailed insights into molecular processes within the human body. However, PET images often suffer from complicated noise, which can obscure critical diagnostic information. The quality of the PET image is impacted by various factors including scanner hardware, image reconstruction, tracer properties, dose/count level, and acquisition time. In this study, we propose a novel text-guided denoising method capable of enhancing PET images across a wide range of count levels within a single model. The model utilized the features from a pretrained CLIP model with a U-Net based denoising model. Experimental results demonstrate that the proposed model leads significant improvements in both qualitative and quantitative assessments. The flexibility of the model shows the potential for helping more complicated denoising demands or reducing the acquisition time.

Text controllable PET denoising

TL;DR

This work tackles the noise and variability inherent in low-count PET imaging by introducing a text-guided denoising framework. The method leverages CLIP to embed count-level semantics into both the encoder and decoder of a U-Net denoiser, enabling controllable restoration from low-count inputs. Training data consist of low-count PET slices paired with higher-count ground truth, with count conditions randomly sampled to promote generalization, and optimization is performed with AdamW. Results show improved SSIM and PSNR over baselines and demonstrate the ability to synthesize higher-count images from low-count inputs, offering potential reductions in radiation exposure and scan time in clinical practice.

Abstract

Positron Emission Tomography (PET) imaging is a vital tool in medical diagnostics, offering detailed insights into molecular processes within the human body. However, PET images often suffer from complicated noise, which can obscure critical diagnostic information. The quality of the PET image is impacted by various factors including scanner hardware, image reconstruction, tracer properties, dose/count level, and acquisition time. In this study, we propose a novel text-guided denoising method capable of enhancing PET images across a wide range of count levels within a single model. The model utilized the features from a pretrained CLIP model with a U-Net based denoising model. Experimental results demonstrate that the proposed model leads significant improvements in both qualitative and quantitative assessments. The flexibility of the model shows the potential for helping more complicated denoising demands or reducing the acquisition time.
Paper Structure (4 sections, 4 figures)

This paper contains 4 sections, 4 figures.

Table of Contents

  1. INTRODUCTION
  2. METHODS
  3. RESULTS
  4. CONCLUSIONS

Figures (4)

  • Figure 1: Proposed text-controlled denoising model.
  • Figure 2: The visual example of synthetizing full count image from various count levels. The first row is the original count level images, the second row is the synthetic full count image from the corresponding count level with the proposed model.
  • Figure 3: The SSIM and PSNR of original various count vs. full count and synthetic full count from various count level vs. full count.
  • Figure 4: Visual and quantification comparison between the original 1/100 count image, denoised using CycleGAN, U-Net and the proposed method (all from the 1/100 count image), with the full count image as reference.