Path and bone-contour regularized unpaired MRI-to-CT translation

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Accurate MRI-to-CT translation integrates complementary imaging data without extra sessions.
• Unpaired MRI-to-CT translation is crucial due to challenges in acquiring paired scans.
• Existing methods struggle with translating bone structures, limiting radiation therapy applications.

Purpose of the Study:

• To develop a robust unpaired MRI-to-CT translation method.
• To enhance the accurate translation of bone structures for improved treatment planning.
• To overcome limitations of current methods in medical image synthesis.

Main Methods:

• Proposed a path- and bone-contour regularized approach for unpaired MRI-to-CT translation.
• Utilized neural ordinary differential equations for MRI-to-CT mapping in a shared latent space.
• Introduced a trainable network for bone contour generation and focused on bone regions.

Main Results:

• The proposed method outperformed existing unpaired MRI-to-CT translation approaches across three datasets.
• Achieved lower overall error rates in MRI-to-CT translation.
• Demonstrated superior performance in bone segmentation, preserving bone structure fidelity.

Conclusions:

• The novel approach significantly improves unpaired MRI-to-CT translation, particularly for bone structures.
• This advancement is vital for accurate radiation therapy planning.
• The method offers a more reliable solution for medical image synthesis from unpaired data.