Multitask Weakly Supervised Generative Network for MR-US Registration.

Area of Science:

• Medical imaging
• Computer-assisted surgery
• Machine learning

Background:

• Accurate registration of pre-operative imaging (MRI, CT) to intra-operative ultrasound is vital for surgical and biopsy guidance.
• Current deep learning methods for image registration require extensive, costly ultrasound domain supervision.
• This limitation hinders the development and accessibility of advanced registration techniques.

Purpose of the Study:

• To develop a novel multitask generative framework for deformable image registration.
• To enable training of registration models using only weak supervision from the pre-operative imaging domain.
• To translate magnetic resonance (MR) images to the ultrasound domain while preserving anatomical structures.

Main Methods:

• A multitask generative framework was designed for deformable registration.
• The framework translates 3D MR images to transrectal ultrasound (TRUS) images for prostate biopsies.
• An in-house dataset of 600 patients was utilized for training, validation, and testing.

Main Results:

• Achieved a target registration error of 3.58 mm on expert-selected landmarks.
• Obtained a Dice score of 89.2% for prostate mask segmentation.
• Reported a 95th percentile Hausdorff distance of 1.81 mm on prostate masks.
• Demonstrated successful translation of MR to ultrasound images with preserved structural details via an ultrasound-specific two-path design.

Conclusions:

• The proposed generative framework effectively translates MR images to the ultrasound domain.
• The method requires only weak supervision from the pre-operative domain, reducing reliance on expensive ultrasound data.
• This approach facilitates the training of deep learning-based registration methods for improved surgical guidance.