Related Experiment Video
Updated: Jun 28, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
ReconU-Net: a direct PET image reconstruction using U-Net architecture with back projection-induced skip connection
1Central Research Laboratory, Hamamatsu Photonics K. K., 5000 Hirakuchi, Hamana-ku, Hamamatsu 434-8601, Japan.
Physics in Medicine and Biology
|April 19, 2024
Summary
A new deep learning model, ReconU-Net, enhances direct positron emission tomography (PET) image reconstruction by integrating physical models. This novel approach improves image quality and reconstruction accuracy, even with limited training data.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
Background:
- Direct image reconstruction in Positron Emission Tomography (PET) is crucial for diagnostic accuracy.
- Existing deep learning models like U-Net and DeepPET have limitations in direct PET image reconstruction.
Purpose of the Study:
- Introduce ReconU-Net, a novel U-Net-shaped architecture for deep learning-based direct PET image reconstruction.
- Compare ReconU-Net with U-Net and DeepPET to visualize direct PET image reconstruction behavior.
Main Methods:
- Developed ReconU-Net by integrating the back projection physical model into the U-Net skip connection.
- Trained and tested ReconU-Net using Monte Carlo simulation data (Brainweb phantom) and real phantom data (Hoffman brain phantom).
Main Results:
- ReconU-Net achieved superior peak signal-to-noise ratio and contrast recovery compared to U-Net and DeepPET.
- Demonstrated effective transfer of multi-resolution features, including high-resolution information, via skip connections.
- Successfully reconstructed a real Hoffman brain phantom with limited training data.
Conclusions:
- ReconU-Net improves direct PET image reconstruction fidelity by combining data-driven learning with the physics model.
- The architecture is effective even with small training datasets, enhancing the synergy between imaging physics and AI.
