Source-linked AI summary
Implementation and evaluation of various demons deformable image registration algorithms on GPU
Xuejun Gu, Hubert Pan, Yun Liang, Richard Castillo, Deshan Yang, Dongju Choi, Edward Castillo, Amitava Majumdar, Thomas Guerrero, Steve B. Jiang
TL;DR
Online ART is constrained by the need for rapid image segmentation, for which DIR can transfer contours but remains computationally slow. This paper implements six demons DIR variants with CUDA on GPUs and evaluates them using pulmonary 4DCT landmark data. GPU computation takes about 7–11 seconds with average 3D errors of 1.5–1.8 mm, while the study finds the original passive-force methods most suitable overall for online ART.
Problem
Online ART needs rapid image segmentation to transfer tumor and organ contours to daily images, but existing DIR computation is too slow.
Method
The study implements six demons DIR variants on GPUs with CUDA and evaluates them on five pulmonary 4DCT datasets containing over 1,100 expert-determined landmark pairs each.
Results
GPU DIR required about 7–11 seconds and produced average 3D errors of 1.5–1.8 mm; passive-force methods offered the strongest overall combination of accuracy and efficiency.
Takeaways & Limitations
The original passive-force demons algorithms seem more suitable for online ART applications based on the available pulmonary data.
Takeaways & Limitations
Evaluation used only pulmonary 4DCT images; other anatomical regions and CT-to-CBCT registration require further investigation.
Abstract
from arXiv · showhide
Online adaptive radiation therapy (ART) promises the ability to deliver an optimal treatment in response to daily patient anatomic variation. A major technical barrier for the clinical implementation of online ART is the requirement of rapid image segmentation. Deformable image registration (DIR) has been used as an automated segmentation method to transfer tumor/organ contours from the planning image to daily images. However, the current computational time of DIR is insufficient for online ART. In this work, this issue is addressed by using computer graphics processing units (GPUs). A grey-scale based DIR algorithm called demons and five of its variants were implemented on GPUs using the Compute Unified Device Architecture (CUDA) programming environment. The spatial accuracy of these algorithms was evaluated over five sets of pulmonary 4DCT images with an average size of 256x256x100 and more than 1,100 expert-determined landmark point pairs each. For all the testing scenarios presented in this paper, the GPU-based DIR computation required around 7 to 11 seconds to yield an average 3D error ranging from 1.5 to 1.8 mm. It is interesting to find out that the original passive force demons algorithms outperform subsequently proposed variants based on the combination of accuracy, efficiency, and ease of implementation.
1. Introduction
The paper implements six demons deformable image registration variants on GPUs to assess their accuracy and efficiency for online adaptive radiation therapy. It evaluates these methods systematically on pulmonary 4DCT data and examines multi-scale processing, stopping criteria, and CUDA implementation choices.
- Motivation: Online ART requires rapid image segmentation to transfer contours from planning images to daily images despite changing patient geometry.
- Approach: Six demons variants are implemented on GPU with CUDA, including passive, evolved passive, active, double, adjusted double, and inverse consistent methods.The variants differ primarily in gradient or force determination.
- Demons variants: The passive force method uses a static-image gradient that is computed once, while evolved passive force incorporates both static- and moving-image information.Gaussian smoothing is used to suppress noise and preserve geometric continuity.
- Demons variants: The active force method uses the gradient of the iteratively updated moving image, while double force combines passive and active forces.
- Implementation: A multi-scale scheme down-samples images to reduce displacement magnitude relative to voxel size and can reduce computation time.The study adds one down-sampling level; further down-sampling did not improve accuracy or efficiency.
- Stopping criteria: The study stops DIR when the displacement-field change over ten iterations is at most ε, with ε = 0.1 × 10^-4.The authors found this measure corresponded more closely to spatial accuracy than correlation coefficient.
3. Experimental Results
The study evaluated GPU-based demons registration on five thoracic 4DCT cases using expert-defined landmarks and compared accuracy, stopping behavior, and computational time across variants.
- Clinical data set: Five clinical thoracic 4DCT cases provided moving and static images for DIR evaluation, with images cropped around the rib cage and sampled to 256×256 voxels.Maximum-inhale and maximum-exhale phases served as moving and static images, respectively.
- Clinical data set: 1,166–1,561 expert-identified landmark pairs per case were used to assess 3D registration accuracy.Mean landmark displacement ranged from 4.01 to 9.42 mm per case.
- Stopping criteria study: Higher correlation generally corresponded to lower 3D spatial error, but correlation alone was insufficient as an accuracy-based stopping criterion.The study compared correlation coefficient and relative norm against landmark-evaluated 3D error.
- Stopping criteria study: Correlation-based performance varied with α, while spatial error reached its best accuracy near α = 0.5 for the reported case.Beyond α = 0.5, correlation continued to decline while spatial-error improvement became minimal.
- Computational performance: GPU computational times across demons variants ranged from 6.5 to 11.5 seconds, with passive-force methods among the fastest.The reported average completion time for the fastest passive-force algorithms was 7.0 seconds.
4. Discussion and Conclusions
GPU implementations of six demons variants achieved execution times and spatial accuracy considered suitable for online ART, with passive-force methods favored overall. Adjusted double force had the lowest average error but required selecting an optimal α value, while the study’s evidence was limited to pulmonary 4DCT images.
- 7.0 seconds was the average execution time for passive force and evolved passive force methods, compared with 8.2–8.5 seconds for three other methods and 11.1 seconds for inverse consistency.All tested GPU implementations were considered likely suitable for online ART based solely on speed.
- 1.51–1.57 mm were the lowest average displacement errors, achieved by adjusted double force, evolved passive force, active force, and passive force methods.All four methods were considered likely suitable DIR candidates for online ART.
- Adjusted double force achieved the lowest average error, but its performance depended on determining an optimal α value that may not generalize beyond pulmonary cases.The study reports that α=2.0 produced maximum spatial accuracy in the pulmonary cases, while further investigation is needed for other anatomical regions.
- Passive and evolved passive force variants were preferred because they lacked an α requirement, had a slight accuracy advantage, and ranked among the fastest algorithms.Double force and inverse consistency were considered less suitable largely because of lower accuracy.
- 100x was the GPU speedup versus default-compiled C, decreasing to about 40x when the C code used the -O3 optimization flag.The study focused on absolute GPU running time rather than relative speedup.
- Pulmonary 4DCT images were the exclusive evaluation data, leaving other anatomical regions and CT-to-CBCT registration for future investigation.The authors identify additional image and landmark datasets as necessary to confirm the findings elsewhere.