Author + information
Quantitative measurement based on precise segmentation in intravascular ultrasound (IVUS) is the most critical process for identifying plaque and guiding PCI. However, manual segmentation is time-consuming and labor-intensive. Moreover, results may vary among observers of diverse experience and skills. AutoIVUS, a deep learning–based algorithm, was developed to segment borders of external elastic membranes (EEM) and lumen automatically for quantification.
As a reference standard, borders of EEM and lumen of 4,097 images from 29 IVUS datasets (Boston Scientific, Natick, Massachusetts) were delineated by an expert observer. Segmentation results were used for quantitative analysis by a predefined calculation framework. This study evaluated the accuracy of segmentation and agreement of quantification with AutoIVUS in corresponding images in comparison with the reference standard.
For all images, mean Hausdorff distance of EEM and lumen were 0.215 ± 0.114 mm and 0.237 ± 0.124 mm, respectively. And mean Dice coefficient of EEM and lumen were 0.959 ± 0.021 and 0.927 ± 0.045, respectively. Results of different types of lesions are shown in the Table. There was strong correlation and good agreement in EEM area (r = 0.96), lumen area (r = 0.97), plaque burden (r = 0.93), and minimum lumen diameter (r = 0.96) between AutoIVUS and expert observer. The mean time of segmentation per frame was substantially shorter with AutoIVUS (0.086 s vs. 25.2 ± 15.4 s; p < 0.001).
|Lesion Type||Hausdorff Distance (mm)||Dice Coefficient|
|Calcified||EEM||0.362 ± 0.135||0.930 ± 0.025|
|Lumen||0.240 ± 0.119||0.936 ± 0.026|
|Noncalcified||EEM||0.216 ± 0.065||0.960 ± 0.018|
|Lumen||0.287 ± 0.150||0.886 ± 0.071|
|Bifurcation||EEM||0.243 ± 0.043||0.952 ± 0.143|
|Lumen||0.264 ± 0.078||0.912 ± 0.027|
|Stent||EEM||0.130 ± 0.023||0.976 ± 0.004|
|Lumen||0.130 ± 0.035||0.964 ± 0.011|
The deep learning–based algorithm for automated segmentation of IVUS showed excellent agreement with an expert observer across most types of lesions, allowing real-time analysis in the catheterization laboratory.
IMAGING: Cath Lab of the Future