Memory-enhanced and multi-domain learning-based deep unrolling network for medical image reconstruction.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Reconstructing high-quality medical images from corrupted data is challenging.
• Deep unrolling (DUN) methods offer a blend of interpretability and deep learning power.
• Existing DUN methods struggle with information flow and global feature capture across stages.

Purpose of the Study:

• To develop an interpretable, high-fidelity medical image reconstruction method using a novel DUN network.
• To address limitations in information flow and global feature extraction in current DUN approaches.

Main Methods:

• Proposed a memory-enhanced (ME) and multi-domain learning-based DUN network.
• Integrated a ME module for adaptive historical output integration.
• Introduced a cross-stage spatial-domain learning transformer (CS-SLFormer) for local and non-local feature extraction.
• Implemented a frequency-domain consistency learning module for fine detail recovery.

Main Results:

• The proposed method demonstrated superior performance across positron emission tomography (PET), MRI, and CT.
• Achieved state-of-the-art quantitative metrics and visual quality in reconstructions.
• Specifically, attained a PSNR of 37.835 dB and SSIM of 0.970 in low-dose PET reconstruction.

Conclusions:

• The developed ME DUN framework significantly enhances medical image reconstruction quality.
• This work expands the application of model-driven deep learning in medical imaging.
• The proposed method shows great potential for high-fidelity image reconstruction in clinical settings.