Compressive X-ray tomosynthesis using model-driven deep learning.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Compressive X-ray tomosynthesis reconstructs 3D objects from limited 2D projections using coding masks.
• Current methods require substantial computational resources and fail to optimize encoding and reconstruction synergistically.
• Existing approaches face challenges in computational efficiency and reconstruction accuracy.

Purpose of the Study:

• To develop a model-driven deep learning (MDL) approach for efficient and accurate tomosynthesis reconstruction.
• To create a unified framework for joint optimization of coding masks and neural network parameters.
• To enhance the degrees of freedom in optimization for improved tomosynthesis imaging.

Main Methods:

• A model-driven deep learning (MDL) framework was developed for compressive X-ray tomosynthesis.
• The framework jointly optimizes coding masks and neural network parameters.
• This unified approach aims to improve computational efficiency and reconstruction accuracy.

Main Results:

• Computational efficiency for coding mask optimization and image reconstruction improved by over an order of magnitude.
• The model-driven deep learning approach significantly enhanced the performance of tomosynthesis reconstruction.
• Joint optimization of encoding and reconstruction stages led to superior results.

Conclusions:

• The proposed model-driven deep learning approach offers a significant advancement in compressive X-ray tomosynthesis.
• This method substantially improves computational efficiency and reconstruction accuracy compared to existing techniques.
• The unified framework provides a more globally optimal solution for 3D object reconstruction.