Low dose computed tomography reconstruction with momentum-based frequency adjustment network.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Model-Based Data-Driven (MBDD) approaches integrate Model-Based Iterative Reconstruction (MBIR) with Deep Learning (DL) for Low-Dose Computed Tomography (LDCT) reconstruction.
• Existing MBDD methods face challenges with DL efficacy, high computational costs, numerous iterations, and ensuring pixel accuracy.
• Loss functions are critical for achieving the required pixel accuracy in DL-based CT reconstruction.

Purpose of the Study:

• To enhance the performance and efficiency of MBDD methods for LDCT reconstruction.
• To address limitations in computational cost and iteration count within current MBDD frameworks.
• To improve the pixel accuracy and overall quality of reconstructed CT images.

Main Methods:

• Introduction of a Frequency Adjustment Network (FAN) to optimize high and low-frequency components during inference.
• Development of a Momentum-based Frequency Adjustment Network (MFAN) utilizing momentum terms for accelerated convergence.
• Proposal of a Focal Detail Loss (FDL) function to preserve fine details during DL model training.

Main Results:

• Validation on AAPM-Mayo and real-world piglet datasets confirmed superior performance of the proposed contributions.
• MFAN achieved convergence in just 10 iterations, demonstrating a significant speed improvement over existing methods.
• Ablation studies underscored the individual and combined effectiveness of FAN, MFAN, and FDL.

Conclusions:

• A novel MBDD-based LDCT reconstruction method is presented, combining MFAN and FDL.
• The proposed approach substantially reduces the required iterations for convergence.
• Superior reconstruction results, both visually and numerically, were achieved.