CT synthesis with deep learning for MR-only radiotherapy planning: a review

Area of Science:

• Medical physics
• Radiotherapy
• Artificial intelligence

Background:

• MR-only radiotherapy planning offers time and safety benefits over traditional CT-based methods.
• Synthetic CT generation is crucial for accurate dose calculation in MR-only workflows.
• Deep learning models show promise in translating MR images to synthetic CT.

Purpose of the Study:

• To provide a comprehensive overview of deep learning approaches for MR-to-CT synthesis.
• To categorize and compare different deep learning models used for this task.
• To evaluate the potential and limitations of current methods for clinical application.

Main Methods:

• Review of deep learning techniques including Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), Transformer models, and Diffusion models.
• Analysis of image-to-image translation strategies for MR-to-CT synthesis.
• Comparative evaluation of model performance based on structure preservation and detail synthesis.

Main Results:

• Deep learning models effectively retain domain-invariant structures while synthesizing domain-specific details for CT images from MR images.
• Different model classes (CNNs, GANs, Transformers, Diffusion) offer distinct advantages and challenges for MR-to-CT synthesis.
• The performance of these models is influenced by architectural choices and training strategies.

Conclusions:

• Deep learning-based MR-to-CT synthesis is a viable approach for advancing MR-only radiotherapy planning.
• Further research into model optimization and validation is needed for widespread clinical adoption.
• Comparative analysis highlights areas for improvement in accuracy and robustness of synthetic CT generation.