High-fidelity 3D geometric reconstruction of pelvic organs from MRI: a hybrid deep learning and iterative optimization approach.
Authors
Affiliations (4)
Affiliations (4)
- Institute of Medical Technology, Peking University Health Science Center, 38 Xueyuan Rd, Haidian District, Beijing, China, Beijing, Beijing, 100191, China.
- Department of Obstetrics and Gynecology, Peking University People's Hospital, No.11 Xizhimen South Street,Xicheng District, Beijing, 100044, China.
- Institute of Medical Technology, Peking University Health Science Center, 38 Xueyuan Rd, Haidian District, Beijing, China, Beijing, 100191, China.
- Department of Obstetrics and Gynecology, Peking University People's Hospital, No.11 Xizhimen South Street, Xicheng District, Beijing, China, Beijing, 100044, China.
Abstract
The generation of high-fidelity, high-quality three-dimensional (3D) pelvic organ geometries from magnetic resonance imaging (MRI)-derived segmentation masks remains challenging. This study aimed to develop and evaluate a hybrid framework for reconstructing the bladder, uterus, and rectum, and to assess its geometric fidelity, mesh-quality proxy performance, and reconstruction efficiency.
Approach. We developed a hybrid deformable shape modeling framework that integrates deep learning prediction with iterative optimization for the reconstruction of the bladder, uterus, and rectum. The framework consists of three core components: (1) a geometry-aware multi-level deep learning architecture that preserves topological consistency of pelvic organs; (2) a two-stage amortized optimization training strategy that balances global shape capture and local surface refinement; and (3) a holistic synergy mechanism-where iterative optimization provides supervision for deep learning during the training phase, and during inference, deep learning rapidly predicts the global organ morphology, followed by iterative optimization to refine local surfaces and mesh quality.
Main results. Before iterative refinement, the proposed model outperformed representative deep learning-based reconstruction methods in geometric fidelity, achieving CD/DSC values of 5.5 × 10⁻³/0.788 for the bladder, 3.7 × 10⁻³/0.839 for the rectum, and 4.2 × 10⁻³/0.900 for the uterus. Deep learning-based initialization enabled the complete hybrid framework to converge within approximately 100 iterations, with a reconstruction time of approximately 35 s per organ on a CPU and 1.5 s on a GPU. At the patient level, the mean minSICN and minSIGE values of the 10 worst volumetric elements were 0.0807 and 0.0844, respectively, exceeding those achieved by traditional geometric post-processing methods.
Significance. The proposed framework connects image-domain segmentation with structural-domain applications by efficiently generating high-fidelity, high-quality pelvic organ geometries with improved mesh-quality proxy performance. These geometries may support statistical shape assessment and may provide candidate surface representations for subsequent volumetric meshing and solver-specific biomechanical validation.