Author: Michael Baine, Yang Lei, Yu Lei, Ruirui Liu, Tian Liu, Jing Wang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, University of Nebraska Medical Center 🌍
Purpose: Accurate 3D deformable registration of MRI and ultrasound (US) is essential for real-time image guidance during high-dose-rate (HDR) prostate brachytherapy. However, MRI-US registration of the prostate is difficult due to several inherent challenges associated with the differences between the two imaging modalities and the pelvic anatomy. This study proposes a weakly supervised spatial implicit neural representation (SINR) learning method to overcome these barriers by leveraging surface information for robust 3D MRI-US deformable registration.
Methods: The proposed framework integrates sparse surface supervision from segmented MRI and US contours instead of dense voxel-wise intensity matching. A SINR was developed to model deformations as continuous, differentiable functions over space. Patient-specific surface priors guide the deformation field, parameterized as a stationary velocity vector field to ensure biologically plausible, temporally consistent deformations. The method was validated on 20 cases from a public Prostate-MRI-US-Biopsy database and 10 institutional HDR brachytherapy cases. Alignment was assessed between the organ contours in US and deformed MRI, using Dice similarity coefficient (DSC), mean surface distance (MSD), and 95th percentile Hausdorff distance (HD95).
Results: The proposed method demonstrated strong performance in deformable image registration. For the public database, the prostate achieved a DSC of 0.93±0.05, MSD of 0.87±0.10 mm, and HD95 of 1.58±0.37 mm. In the institutional database, prostate CTV achieved a DSC of 0.88±0.09, MSD of 1.21±0.38 mm, and HD95 of 2.09±1.48 mm. Lower performance for bladder and rectum was attributed to ultrasound’s limited field of view. Visual results confirmed accurate alignment with minimal discrepancies.
Conclusion: This study introduces a novel weakly supervised SINR-based method for 3D MRI-US deformable registration, addressing the challenges of multimodal differences. By leveraging sparse surface supervision and spatial priors, this method achieves accurate, robust, and computationally efficient registration. The framework enhances real-time image guidance during HDR prostate brachytherapy, improving treatment precision and clinical outcomes.