Loading...

Attenuation and Scatter Correction in Whole Body Positron Emission Tomography Images using Neural Networks

Adeli, Zahra | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57903 (46)
  4. University: Sharif University of Technology
  5. Department: Energy Engineering
  6. Advisor(s): Hosseini, Abolfazl; Sheikhzadeh, Peyman
  7. Abstract:
  8. Objective: Accurate quantification in PET imaging is essential for proper evaluation and requires attenuation and scatter correction. Typically, this correction is performed using CT; however, the propagation of CT image errors to PET, the additional radiation dose in PET/CT, especially in pediatric cases and repeated scans, as well as challenges associated with standalone PET and PET/MRI, highlight the need for alternative approaches. This study investigates the capability of deep learning in jointly correcting attenuation and scatter for PET images with different radiotracers. Methods: Various neural network models, including ResNet, U-Net, and swin UNETR, were trained for direct correction of PET images. The performance of the models was evaluated using metrics such as Relative Error (RE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM), with the CT-corrected image serving as the reference. Both 2D and 3D training performances were assessed for one of the networks, U-Net, and 3D training was selected due to its superior performance. The training data included images from the radiotracers 68Ga-PSMA and 18F-FDG, and the models were trained on these radiotracers separately as well as simultaneously to compare the generalizability of U-Net and swin UNETR. Finally, transfer learning with the swin UNETR model pre-trained on 68Ga-PSMA was employed for correcting 64Cu PET images. Results: The U-Net model demonstrated robust performance in processing images of the radiotracer 18F-FDG, and this advantage persisted in simultaneous training on both radiotracers. However, for 68Ga-PSMA images, due to the requirement for learning more complex features and handling a broader value range, the U-Net model performed weaker compared to swin UNETR. This weakness was further pronounced in simultaneous training. In contrast, the swin UNETR model, leveraging attention mechanisms and transformers, maintained its stability in simultaneous learning and provided higher accuracy for images with complex distributions, such as 68Ga-PSMA. These findings suggest that swin UNETR, with its high flexibility, is a suitable choice for improving diagnostic accuracy in complex clinical applications. Conclusion: This study demonstrates that deep learning can effectively correct PET images without the need for CT. The swin UNETR model showed promising results in simultaneous training on two radiotracers, highlighting its generalizability in clinical applications with different radiotracers. This approach is particularly beneficial for sensitive populations, such as pediatric patients and those requiring frequent imaging, as it eliminates the CT imaging dose
  9. Keywords:
  10. Positron Emission Tomography (PET) ; Attenuation Correction ; Scatter Correction ; Neural Network ; Deep Learning ; Quantification

 Digital Object List

 Bookmark

No TOC