Loading...

Enhancement Of the Quality In PET Brain Images Using AutoGANcoder Algorithms

Razavi Satvati, Reyhaneh Sadat | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58033 (46)
  4. University: Sharif University of Technology
  5. Department: Energy Engineering
  6. Advisor(s): vosoughi, Naser; Ghafarian, Pardis; Attari, Ali Akbar
  7. Abstract:
  8. Positron Emission Tomography (PET) is a cutting-edge medical imaging technique that enables the examination of functional activities in different regions of the body, particularly the brain. This technique involves injecting a radiopharmaceutical into the body, which emits gamma rays that are subsequently detected by the PET scanner to generate images. However, a significant challenge in PET imaging is the degradation of image quality due to noise and photon scattering, which can lead to diagnostic errors, especially in sensitive applications such as detecting brain diseases. Various image reconstruction techniques have been developed in the past to mitigate attenuation and scattering effects and enhance PET image quality. These approaches include classical techniques based on traditional algorithms, such as iterative regularization, repetitive filtering, and optimization of directional vectors. While these methods have shown partial success in noise reduction and image quality enhancement, they suffer from computational complexity, requiring extensive processing time and high computational resources. Additionally, classical methods often struggle to preserve fine details and subtle features in complex environments like the brain. With advancements in deep learning technology, neural network-based methods have recently emerged as alternatives to traditional approaches. These techniques include convolutional neural networks (CNNs) for spatial feature extraction, recurrent neural networks (RNNs) for sequential data processing, and other deep learning models such as deep belief networks (DBNs) for noise correction and PET image reconstruction. Compared to traditional methods, these models demonstrate superior performance in reconstructing high-detail images and mitigating errors caused by photon scattering. However, challenges such as network fine-tuning and dependence on high-quality datasets remain critical areas of research. This study aims to enhance the quality of brain PET images of epilepsy patients by addressing attenuation and scattering effects using a novel deep learning-based approach. The proposed model integrates Generative Adversarial Networks (GANs) with Autoencoders to reduce noise, correct attenuation and scattering, and improve image resolution. This hybrid approach not only combines the advantages of both methods but also achieves significant performance improvements due to the interaction between the generative capabilities of GANs and the compression properties of Autoencoders. Experimental results indicate that the proposed hybrid model, named AutoGANcoder, significantly outperforms standalone U-Net and Autoencoder models in key evaluation metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Fréchet Inception Distance (FID), Perceptual Similarity Metric (LPIPS), and Mean Squared Error (MSE). Furthermore, this approach effectively preserves the structural details and subtle features of brain images while reducing noise and scattering artifacts. Therefore, the proposed model can serve as an efficient tool for improving PET image quality and correcting attenuation and scattering effects.
  9. Keywords:
  10. Generative Adversarial Networks ; Image Quality ; Positron Emission Tomography (PET) ; Attenuation Correction ; Deep Learning ; Functional Brain Imaging ; Image Enhancement ; Scatter Correction

 Digital Object List

 Bookmark

No TOC