Novel U-net based deep neural networks for transmission tomography

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning and computer tomography (CT) fusion improves image quality and reduces artifacts in reconstructed images.
• Beam hardening and electrical noise are common artifacts in tomographic imaging.

Purpose of the Study:

• Present novel neural network architectures for tomographic reconstruction.
• Reduce the effects of beam hardening and electrical noise in CT images.

Main Methods:

• Developed two novel neural network architectures with integrated image reconstruction.
• Trained networks considering the mathematical model of projections for enhanced data and image quality.
• Evaluated models on simulated datasets with beam hardening and electrical noise using numerical error measurements.

Main Results:

• The TomoNet2 architecture demonstrated superior performance compared to other methods.
• TomoNet2 significantly improved Structural Similarity Index (SSI) scores on two datasets.
• Achieved high Peak-Signal-to-Noise-Ratio (PSNR) percentages, outperforming alternative techniques.

Conclusions:

• Integrating reconstruction steps within deep neural networks, particularly with skip connections, enhances reconstruction quality.
• The proposed method shows potential for broad application in various tomographic imaging tasks.