Quasi-supervised MR-CT image conversion based on unpaired data

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiotherapy Planning

Background:

• Simultaneous acquisition of MRI and CT images is vital for radiotherapy planning but is costly, time-consuming, and involves ionizing radiation.
• Generating CT images from MRI can mitigate these issues, but requires accurate cross-modality image synthesis.

Purpose of the Study:

• To develop a novel quasi-supervised learning framework for generating CT images from radiation-free MR images.
• To improve the accuracy and fidelity of MR-to-CT image translation for radiotherapy applications.

Main Methods:

• A quasi-supervised framework was developed to analyze unpaired MR and CT images.
• Normalized Mutual Information (NMI) and the Hungarian algorithm were used to establish optimal MR-CT pairings.
• These pairs and NMI scores served as priors to guide the MR-to-CT image translation model.

Main Results:

• The proposed method significantly outperformed existing unsupervised methods in image quality and feature consistency.
• Generated CT images demonstrated higher accuracy and fidelity, preserving anatomical details and structural integrity.
• The framework successfully converted unpaired MR and CT images into structurally consistent pseudo-pairs.

Conclusions:

• The quasi-supervised framework enhances MR-to-CT conversion accuracy and reliability.
• This approach reduces dependence on expensive and scarce paired medical imaging datasets.
• It offers a practical and scalable solution for medical imaging applications lacking paired annotations.