Background: Pulmonary embolism (PE) remains a leading cause of cardiovascular mortality with a low early diagnosis rate, highlighting the urgent need for more efficient detection methods. Non-contrast-enhanced computed tomography (NCCT) offers a promising alternative to computed tomography pulmonary angiography (CTPA) but requires accurate image registration for artificial intelligence (AI) model training. This study evaluates and compares three commonly used image registration algorithms (Elastix, ANTs, and Demons) to determine their efficacy in aligning pulmonary arteries on CTPA and NCCT images for potential AI applications in PE detection.
Methods: This retrospective study included 324 PE patients diagnosed via CTPA across three hospitals. Pulmonary arteries and their branches were segmented on NCCT and CTPA images using Totalsegmentator with manual radiologist review. Three registration algorithms (Elastix, ANTs, Demons) were compared against rigid registration. Accuracy was assessed using Dice coefficient, Jaccard similarity [Intersection over Union (IoU)], 95th percentile Hausdorff distance (HD95), and clinical physician scores (1-5 scale). Pairwise comparisons were performed using Wilcoxon signed-rank tests with rank biserial correlation for effect size calculation.
Results: Elastix demonstrated significantly superior performance compared to ANTs and Demons across all metrics. Elastix achieved Dice coefficient of 0.819±0.030 vs. 0.732±0.032 for ANTs and 0.692±0.062 for Demons (both P<0.001, effect size r=1.0). IoU values were 0.695±0.043 for Elastix, 0.578±0.040 for ANTs, and 0.532±0.071 for Demons (both P<0.001, r=1.0). HD95 was significantly lower for Elastix (3.42±0.57 mm) compared to ANTs (7.89±1.29 mm) and Demons (6.49±1.68 mm) (both P<0.001, r=-0.998 to -1.0). Median clinical physician scores for Elastix were 4 [interquartile range (IQR): 4, 5], indicating "clinically acceptable with minimal or no modifications needed", vs. 3 (IQR: 3, 4) for ANTs and 4 (IQR: 3, 4) for Demons.
Conclusions: Among the evaluated registration methods, Elastix showed the highest accuracy and clinical applicability for transferring thrombus landmarks from enhanced CTPA to NCCT images (Dice =0.819, IoU =0.695, HD95 =3.42 mm, all P<0.001 vs. comparators). This study provides a foundation for developing AI models to detect PE in NCCT images, although further validation of thrombus ground standards on NCCT is required.
Keywords: Pulmonary embolism (PE); algorithm evaluation; artificial intelligence (AI); image registration; unenhanced chest computed tomography (unenhanced chest CT).
© AME Publishing Company.