Deep implicit optimization enables robust learnable features for deformable image registration

2024-06-11

Medical Image Analysis

AI Summary

Existing deep learning-based image registration methods, though efficient and amenable to weak supervision, neglect the invariance and robustness of optimization procedures, leading to performance degradation under domain shifts. To address this, we propose an implicit deep optimization framework that, for the first time, embeds implicit neural representations into a differentiable unrolled optimization process—enabling end-to-end coupling of feature learning and deformation field optimization within a differentiable implicit function. Leveraging gradient-truncated backpropagation and an unsupervised deformation consistency constraint, our method significantly enhances registration robustness across cross-modality and low-contrast scenarios. On benchmarks including BraTS and LPBA40, it achieves a 3.2% improvement in Dice coefficient and a 21% reduction in mean surface distance, while demonstrating strong resilience to noise and imaging artifacts.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search
• Machine Learning: Optimization

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Problem

• Bridges gap between statistical learning and optimization in image registration
• Improves performance under domain shift like anisotropy and intensity variations
• Enables switching transformation representations at test time without retraining

Innovation

• Incorporates optimization as a deep network layer
• Uses multi-scale feature images for registration
• Enables arbitrary transformation switching at test time