Structure-preserved meta-learning uniting network for improving low-dose CT quality.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Deep neural networks (DNNs) show promise for low-dose computed tomography (LDCT) but struggle with simulated data limitations.
• Simulated datasets often deviate from real clinical scenarios, causing overfitting, instability, and poor robustness in DNN models.
• Existing methods lack effective strategies for handling unlabeled LDCT data in real-world applications.

Purpose of the Study:

• To develop a structure-preserved meta-learning uniting network (SMU-Net) for noise artifact suppression and detail preservation in unlabeled LDCT imaging.
• To address the limitations of simulated datasets by leveraging real-world unlabeled LDCT data.
• To enhance the robustness and performance of LDCT reconstruction methods.

Main Methods:

• Introduced SMU-Net, comprising a teacher and a student network, utilizing a meta-learning strategy.
• The teacher network, trained on simulated data, generates pseudo-labels for the student network.
• The student network is trained on unlabeled real LDCT data, incorporating a Co-teaching strategy for improved robustness.

Main Results:

• SMU-Net demonstrated superior performance in reducing noise-induced artifacts and preserving structural details across multiple datasets.
• Quantitative analysis showed SMU-Net achieved the highest peak signal-to-noise ratio (PSNR) and structural similarity index measurement (SSIM).
• The method resulted in the lowest root-mean-square error (RMSE) and natural image quality evaluator (NIQE) scores, indicating enhanced image quality.

Conclusions:

• The proposed meta-learning strategy effectively utilizes unlabeled CT images to improve reconstruction performance in LDCT.
• SMU-Net offers a robust solution for generating high-quality LDCT images in real clinical scenarios.
• This approach overcomes the limitations of traditional DNN-based methods trained solely on simulated data.