Unpaired Deep Cross-Modality Synthesis with Fast Training.

Area of Science:

• Medical imaging
• Computer vision
• Artificial intelligence

Background:

• Cross-modality synthesis converts images between different imaging types, valuable for research and clinical use.
• Existing methods often require perfectly aligned paired data, which is difficult to obtain.
• Misalignment in paired data can negatively impact training and synthesized image quality.

Purpose of the Study:

• To develop a novel method for cross-modality image synthesis using unpaired data.
• To address the challenge of acquiring perfectly aligned multi-modal image datasets.
• To improve the quality and reliability of synthesized medical images.

Main Methods:

• Utilized generative adversarial networks (GANs) for image synthesis.
• Implemented a cyclic training approach for efficient model training.
• Introduced a novel structural dissimilarity loss function to preserve anatomical details.

Main Results:

• Successfully performed cross-modality synthesis using only unpaired data.
• Demonstrated effective synthesis for brain MR-to-CT, prostate MR-to-CT, and brain 3T-to-7T tasks.
• Achieved good synthesis performance, validating the proposed method's capabilities.

Conclusions:

• The proposed method enables robust cross-modality image synthesis without requiring paired data.
• The novel loss function enhances the preservation of anatomical structures in synthesized images.
• This approach offers a viable solution for medical image synthesis challenges where paired data is scarce.