Memory-Efficient Training for Fully Unrolled Deep Learned PET Image Reconstruction with Iteration-Dependent Targets.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Machine Learning

Background:

• The forward-backward splitting expectation-maximisation network (FBSEM-Net) unfolds the maximum a posteriori expectation-maximisation algorithm for image reconstruction.
• Existing methods often require significant computational resources, limiting the training of complex, unrolled network architectures, especially in 3D.

Purpose of the Study:

• To introduce an enhanced FBSEM-Net with a novel memory-efficient training method.
• To enable the training of fully unrolled 3D FBSEM-Net implementations.
• To improve network performance through iteration-dependent regularization, targets, and losses.

Main Methods:

• Developed a new version of FBSEM-Net incorporating iteration-dependent networks, targets, and losses.
• Implemented a sequential training strategy to enhance memory efficiency.
• Trained and evaluated the modified FBSEM-Net on 2D and 3D simulated data.

Main Results:

• Iteration-dependent targets and losses improved optimization consistency and generalization capabilities.
• Iteration-dependent regularization slightly reduced reconstruction error compared to fixed regularization.
• Sequential training reduced memory usage by up to 98%, enabling 3D FBSEM-Net training, with no significant performance impact from truncated backpropagation.

Conclusions:

• The proposed FBSEM-Net with iteration-dependent components and sequential training offers improved performance and generalization.
• Sequential training effectively addresses memory limitations, making deep unrolled network training feasible.
• The enhanced FBSEM-Net is a viable solution for complex image reconstruction tasks, particularly in 3D.