Optimizing GPU Programming for Intensity-Based Image Registration in Surgery

Overview

This study presents a GPU-based optimization framework to accelerate intensity-based non-rigid image registration, specifically the diffeomorphic log-demons algorithm, achieving significant runtime reductions. The approach leverages performance-aware programming to overcome computational bottlenecks, enabling intraoperative registration within clinically practical timeframes.

Background

Non-rigid image registration is critical in aligning preoperative and intraoperative medical images, especially in procedures like MRI-guided cardiac electrophysiology where tissue deformation occurs. Intensity-based methods, such as the Demons algorithm, offer robustness but are computationally intensive, often exceeding acceptable intraoperative time limits. Conventional CPU processing is insufficient for the required speed, while existing GPU implementations have not met the stringent time constraints needed during surgery. Optimizing GPU utilization through performance-aware programming can potentially address these challenges.

Data Highlights

The diffeomorphic log-demons algorithm typically requires over 2 minutes on CPU systems for large 3D images. Previous GPU implementations achieved up to 20x speedup but still required more than 35 seconds for 3D image registration at 5-megapixel resolution. The proposed GPU optimization framework targets these bottlenecks to reduce runtime below the 10-second threshold necessary for intraoperative use.

Key Findings

• Intensity-based non-rigid registration algorithms are computationally demanding due to voxel-wise iterative operations like convolution and interpolation.
• Conventional CPUs are not optimized for the high-throughput computations required, leading to long processing times incompatible with surgical workflows.
• GPUs offer parallel and scalable architectures ideal for accelerating these computations but require performance-aware programming to fully exploit their capabilities.
• Performance-aware programming involves profiling bottlenecks and optimizing GPU resource utilization, memory access, and thread management.
• The diffeomorphic log-demons algorithm serves as an effective benchmark for optimization due to its widespread use and fundamental image operations.
• The optimized GPU implementation significantly reduces runtime, making real-time intraoperative image registration feasible.

Clinical Implications

Implementing optimized GPU-based intensity registration can integrate real-time image alignment into surgical workflows, providing surgeons with up-to-date anatomical information immediately after interventions such as ablation. This enhancement supports more precise and adaptive surgical guidance, potentially improving procedural outcomes and patient safety. The open-source nature of the optimized library facilitates broader adoption and adaptation across various clinical imaging applications.

Conclusion

Performance-aware GPU programming effectively addresses the computational challenges of intensity-based non-rigid image registration, enabling intraoperative use within clinically acceptable timeframes. This advancement holds promise for enhancing image-guided interventions through rapid, accurate image alignment.

References

1. Wang et al. 2024 -- Optimizing Programming for Intensity-Based Image Registration During Surgery Using Graphics Processing Units