Local metric learning in 2D/3D deformable registration with application in the abdomen

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Image Analysis

Background:

• Respiratory motion and soft tissue deformations in image-guided radiotherapy (IGRT) can compromise localization accuracy.
• Accurate soft tissue localization is critical for effective disease site targeting during radiotherapy, especially in the abdomen.
• Existing methods like REALMS (registration efficiency and accuracy through learning a metric on shape) use global metrics for 3D deformation parameter interpolation.

Purpose of the Study:

• To refine the REALMS method for improved 2D/3D deformable registration accuracy in abdominal IGRT.
• To address challenges in abdominal IGRT, including low contrast and non-respiratory motion.
• To enhance real-time soft tissue localization for abdominal disease sites.

Main Methods:

• Application of a refined REALMS method using a localized Riemannian metric trained on specific deformation spaces.
• Introduction of a rigid translation vector to account for non-respiratory deformations.
• Development of a region-of-interest around fiducial markers for more reliable abdominal registration.

Main Results:

• The localized REALMS approach demonstrated superior 2D/3D deformable registration accuracy compared to the global approach.
• Testing on both synthetic and actual abdominal datasets confirmed the enhanced performance.
• The refined method effectively handles challenges like low contrast and non-respiratory motion in abdominal IGRT.

Conclusions:

• Localized training of the Riemannian metric significantly improves interpolation accuracy within specific deformation regions.
• The proposed algorithm enhances soft tissue localization accuracy for abdominal IGRT.
• This refined method offers a more reliable solution for real-time image-guided radiotherapy in challenging abdominal cases.