Author + information
Image interpretation remains an obstacle for increased use of intravascular ultrasound (IVUS) in the catheterization laboratory. Accurate measurement of lumen area allows for appropriate stent selection, whereas detailed knowledge of plaque phenotype can inform the clinician on post–percutaneous coronary intervention optimization strategy. This work aims to develop and validate a deep learning–based algorithm for IVUS image segmentation and phenotyping.
IVUS pullbacks (n = 305) were acquired with Philips Eagle Eye Platinum 20 MHz (n = 227), Philips Revolution 45 MHz (n = 44), or Boston Scientific OptiCross 40 MHz (n = 34) catheters. This dataset was split into training (n = 270) and validation (n = 35) sets. A convolutional neural network was used to segment and classify the phenotype of each IVUS image. Training images are fed into the network and the internal and external elastic lamina borders are determined from the segmentation component. The phenotyping component classifies each IVUS image (20 MHz only) as normal, fibrotic, fibroatheroma, calcified, or stented.
Very good agreement (concordance correlation coefficient [CCC] = 0.9) was seen between algorithm-computed (8.25 ± 4.25 mm2) and expert-measured (7.87 ± 3.89 mm2) lumen areas. Good agreement was also seen between both plaque area (CCC = 0.85) and plaque burden (CCC = 0.88) for the algorithm (7.11 ± 4.16 mm2; 45.21 ± 15.16%) and expert analyst (6.78 ± 3.88 mm2; 45 ± 15.96%). Normal, stented, and calcified images were classified with high accuracy (Figure).
The authors have developed and validated a deep learning–based platform for automatic segmentation and phenotyping of IVUS pullbacks.
IMAGING: Imaging: Intravascular