3-D Reconstruction in Canonical Co-Ordinate Space From Arbitrarily Oriented 2-D Images

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Optimization-based 2D/3D image registration methods require high-quality initialization, limiting their performance in 3D reconstruction and motion compensation.
• Significant subject motion in clinical scenarios, like fetal in-utero imaging, complicates 3D image and volume reconstruction.

Purpose of the Study:

• To present a learning-based image registration method for predicting 3D rigid transformations of 2D image slices relative to a learned canonical atlas.
• To eliminate the need for spatial transform initialization in 2D/3D image registration.

Main Methods:

• Utilized a convolutional neural network (CNN) architecture to learn a regression function mapping 2D image slices to a 3D canonical atlas space.
• Employed only image slice intensity information for registration and canonical alignment.

Main Results:

• Achieved an average spatial prediction error of 7 mm on simulated magnetic resonance imaging (MRI) fetal brain data with synthetic motion.
• Produced qualitatively improved 3D reconstructions for fetuses with significant motion (approx. 20 weeks gestational age).

Conclusions:

• The proposed learning-based registration method offers a general, computationally efficient solution for 2D/3D registration initialization.
• The method is suitable for real-time applications and enhances 3D image reconstruction and motion compensation pipelines.