Efficient Parallel Implementation of a 3-D MRF-Based Deformable Image Registration Algorithm in Medical Dynamic CT Imaging

Registration techniques play a critical role in computer vision and medical image analysis, enabling the alignment of images acquired at different times or from different modalities. The increasing demand for high-resolution and real-time processing necessitates faster deformable registration algorithms; however, many accurate methods remain computationally expensive, limiting their clinical applicability. This work presents three acceleration strategies for a 3D MRF-based deformable image registration algorithm applied to dynamic CT imaging. First, the parallelizable components of the algorithm were identified and parallelized on the CPU using OpenMP. Second, a fully GPU-based implementation using CUDA was developed. Third, a heterogeneous hybrid CPU/GPU implementation was designed to optimally distribute computational tasks across both platforms. Experimental results on the Learn2Reg dataset with images of size (192 × 160 × 256), demonstrate that the hybrid approach achieves the best overall performance by strategically offloading sequential graph-based bottlenecks to the CPU. It achieves a speedup factor of ~9× compared to the Serial implementation while maintaining comparable accuracy, overcoming the severe hardware utilization limits of the fully GPU-based approach and highlighting the challenges of mapping MRF optimization to massively parallel architectures. These results indicate that hybrid computing architectures provide an effective trade-off between computational efficiency and registration accuracy. A graphical user interface was also developed to visualize the registration process and results.