Physics-/Model-Based and Data-Driven Methods for Low-Dose Computed Tomography: A survey.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning (DL) has significantly advanced tomographic imaging, particularly low-dose computed tomography (LDCT), since 2016.
• However, pure DL methods for LDCT denoising and reconstruction face challenges like the 'black box' nature and instabilities, hindering clinical application.

Purpose of the Study:

• To systematically review physics/model-based data-driven methods for LDCT.
• To summarize loss functions and training strategies for these hybrid approaches.
• To evaluate the performance and discuss future directions in hybrid LDCT methods.

Main Methods:

• Review of hybrid deep learning models integrating imaging physics and models for LDCT.
• Analysis of various loss functions and training strategies employed in these hybrid networks.
• Performance evaluation of different physics-informed DL approaches for LDCT.

Main Results:

• Hybrid methods combining physics/model-based and data-driven elements show potential to overcome limitations of pure DL in LDCT.
• Integration of physical principles enhances stability and interpretability of DL models for LDCT.
• Systematic review provides a comprehensive overview of current trends and challenges.

Conclusions:

• Hybrid physics/model-based data-driven methods represent a promising direction for advancing low-dose CT imaging.
• Addressing instabilities and interpretability issues is crucial for the clinical translation of DL in LDCT.
• Further research is needed to optimize these hybrid approaches and explore future directions.