CT Reconstruction With PDF: Parameter-Dependent Framework for Data From Multiple Geometries and Dose Levels.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning computed tomography (CT) reconstruction typically requires fixed scanning parameters, increasing training costs and data demands.
• Training separate models for diverse clinical applications with varying geometries and dose levels is inefficient.

Purpose of the Study:

• To develop a unified deep learning framework for CT reconstruction adaptable to multiple scanning geometries and dose levels.
• To reduce the computational and data requirements for training CT reconstruction networks.

Main Methods:

• Proposed a parameter-dependent framework (PDF) that incorporates parameterized geometry and dose levels.
• Utilized two multilayer perceptrons (MLPs) to modulate feature maps within the CT reconstruction network.
• Conditioned network outputs on varying geometries and dose levels simultaneously during training.

Main Results:

• The PDF achieved competitive performance compared to networks trained on specific or mixed datasets.
• Demonstrated efficient training across multiple geometries and dose levels with a single model.
• Significantly reduced the extra training costs associated with diverse scanning conditions.

Conclusions:

• The parameter-dependent framework offers an efficient solution for deep learning-based CT reconstruction.
• This approach alleviates the burden of extensive training data and computational resources for varied clinical scenarios.
• The PDF enables flexible and cost-effective CT image reconstruction adaptable to different scanning parameters.