MetaInv-Net: Meta Inversion Network for Sparse View CT Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• X-ray Computed Tomography (CT) is crucial for clinical diagnosis and image-guided interventions.
• Current CT image reconstruction methods often rely on complex, data-adaptive deep learning models.
• Existing strategies incorporate numerous data-adaptive components, increasing model complexity.

Purpose of the Study:

• To develop a more efficient and effective deep learning model for CT image reconstruction.
• To investigate a novel approach by learning only essential components of iterative algorithms.
• To introduce the MetaInv-Net, a meta-learning model for CT image reconstruction.

Main Methods:

• Proposed a deep learning model, MetaInv-Net, based on unrolling an iterative algorithm.
• Focused on learning the initializer for the conjugate gradient (CG) algorithm within the backbone model.
• Kept traditional components like image priors and hyperparameters unchanged.
• Utilized a hypernetwork for CG module initialization, classifying the model as meta-learning.

Main Results:

• MetaInv-Net demonstrated superior CT image reconstruction performance compared to state-of-the-art deep models.
• The model requires significantly fewer trainable parameters.
• Achieved excellent results in both simulated and real-world data experiments.
• Showcased generalization capabilities across different scanning settings, noise levels, and datasets.

Conclusions:

• The MetaInv-Net offers an efficient and high-performing solution for CT image reconstruction.
• Learning only specific, intuition-driven parts of iterative algorithms is sufficient for high performance.
• The model's meta-learning nature contributes to its effectiveness and generalizability.