-TomoGNN- Linking the Dots: Pixel-Detectors Associations for Improved PET Direct Image Reconstruction.
TomoGNN introduces a state-of-the-art approach for Positron Emission Tomography (PET) image reconstruction using Graph Neural Networks (GNNs). It addresses limitations of traditional and deep learning-based methods by providing high-fidelity reconstructions with minimal noise even with sparse data.
PET imaging faces challenges due to its inherently noisy and ill-posed nature. Conventional model-based iterative reconstruction (MBIR) methods are effective but computationally intensive. Deep learning approaches often suffer from over-smoothing and lack generalizability.
TomoGNN formulates the PET inverse problem within a graph-theoretical framework, modeling the relationship between lines of response (LORs) and pixels using GNNs. This leads to accurate and efficient reconstructions that preserve anatomical structures without over-smoothing.
The PET reconstruction problem is approached through a Poisson noise model, emphasizing the need for accurately inferring the underlying activity distribution from measured projection data (sinogram), considering system resolution, attenuation, and random and scatter events.
TomoGNN's architecture consists of three main components:
- Sinogram-to-Sinogram Denoising (CNN): Pre-processes the sinogram to reduce noise and improve data quality.
- Sinogram-to-Image Mapping (GCN): A novel single-layer GCN effectively projects sinogram data into the image domain.
- Image-to-Image Refinement (CNN): Further refines the mapped image to enhance local features and details.
The model was trained on a dataset comprising PET images with artificially introduced lesions, using Mean Squared Error (MSE) and Gradient Difference Loss (GDL) to preserve high-frequency image features.
- Efficient and Accurate Reconstruction: TomoGNN outperforms traditional MBIR and other deep learning methods, providing detailed images with controlled noise.
- Minimal Model Complexity: Despite superior performance, TomoGNN requires fewer trainable parameters, making it efficient in memory and computational resources.
- Robustness to Varied Data Conditions: Capable of handling sinograms of any dimension and adaptable to different noise levels and counts, demonstrating exceptional generalization capabilities.
TomoGNN was trained using the Adam optimizer over 300 epochs, with specific attention to data augmentation techniques for model robustness. Training utilized a sophisticated loss function balancing reconstruction fidelity with the preservation of image textures and edges.