A two-step deep learning method for 3DCT-2DUS kidney registration during breathing.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Accurate kidney registration between 3D CT and 2D ultrasound is challenging due to breathing motion.
• Existing methods may struggle with the semantic gap between different imaging modalities and resolutions.

Purpose of the Study:

• To develop and validate KidneyRegNet, a novel deep registration pipeline for 3D CT and 2D ultrasound kidney scans.
• To address the difficulties in 3D CT-2D US kidney registration during free breathing using advanced network architectures and training strategies.

Main Methods:

• Proposed KidneyRegNet, a pipeline featuring a handcrafted texture feature network and a 3D-2D CNN registration network.
• Employed a feature-image-motion (FIM) loss within an encoder-decoder structure for hierarchical regression.
• Utilized unsupervised one-cycle transfer learning for adaptation to patient-specific data after pretraining.

Main Results:

• Achieved a mean contour distance (MCD) of 0.94 mm for CT-US kidney registration and 1.15 mm for CT-CT registration.
• Demonstrated robust performance across varying transformation magnitudes, with MCDs of 0.82-1.10 mm for CT-US and 1.02-1.28 mm for CT-CT.
• Validated on diverse datasets including 132 US sequences, 39 multi-phase CT, 210 single-phase CT, and 25 CT-US pairs.

Conclusions:

• KidneyRegNet effectively addresses the complexities of 3D CT-2D US kidney registration in free-breathing conditions.
• The novel network structures and transfer learning strategies enhance registration accuracy and applicability in clinical settings.
• This pipeline offers a promising solution for improved non-rigid registration in medical imaging.