Multi-energy CT material decomposition using graph model improved CNN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Accurate tissue composition estimation in spectral CT relies on material decomposition, crucial for disease diagnosis.
• Traditional Convolutional Neural Networks (CNNs) struggle to capture non-local features essential for precise material decomposition.
• Existing methods face limitations in extracting comprehensive image features for multi-material decomposition (MMD).

Purpose of the Study:

• To develop a novel Multi-Material Decomposition (MMD) method using a graph-based U-net architecture (GECCU-net) to enhance spectral CT material image quality.
• To improve the extraction of both local and non-local features for more accurate tissue composition analysis.
• To reduce noise and artifacts in spectral CT images, thereby improving diagnostic accuracy.

Main Methods:

• Proposed a novel GECCU-net incorporating a multi-scale encoder and local and non-local feature aggregation (LNFA) blocks.
• Utilized graph edge-conditioned convolution on non-Euclidean spaces to effectively extract non-local features.
• Implemented a hybrid loss function to handle multi-scale inputs and prevent result over-smoothing.

Main Results:

• GECCU-net generated material images with reduced noise and artifacts compared to baseline CNN models.
• The method successfully retained more detailed tissue information.
• Achieved high Structural SIMilarity (SSIM) values (0.9976 for abdomen, 0.9990 for chest water maps) and low RMSE (0.1218 g/cm³ for abdomen, 0.4903 g/cm³ for chest).

Conclusions:

• The proposed GECCU-net method significantly enhances Multi-Material Decomposition (MMD) performance in spectral CT imaging.
• The technique offers improved image quality, crucial for accurate disease diagnosis.
• GECCU-net demonstrates potential for widespread application in advanced medical imaging analysis.