Convolutional networks for the segmentation of intravascular ultrasound images: Evaluation on a multicenter dataset. (March 2022)

• Record Type: Journal Article
• Authors: Du, Haiyan; Ling, Li; Yu, Wei; Wu, Peng; Yang, Yuan; Chu, Miao; Yang, Junqing; Yang, Wei; Tu, Shengxian
• Is Part Of: Computer methods and programs in biomedicine. Volume 215(2022)
• Publication Date: 2022-03
• Subjects: IVUS Intravascular ultrasound -- EEL External elastic lamina -- FCN Fully convolutional network -- DSC Dicesimilarity coefficient -- JI Jaccard index -- HD Hausdorffdistance -- CNN Convolutional neural network; Intravascular ultrasound -- Coronary artery disease -- Image segmentation -- Deep learning -- Convolutional neural network; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.106599
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: Construct a large IVUS image dataset including the images acquired with different vessel characteristics and with ultrasonic probes of different frequencies. Evaluation of several state-of-the-art deep learning methods on a large and multiple-center dataset. Training the models with the IVUS images of two frequencies can improve the segmentation performance. Abstract: Background and objective: The delineation of the lumen contour and external elastic lamina (EEL) in intravascular ultrasound (IVUS) images is crucial for the quantitative analysis of coronary atherosclerotic plaques. However, the presence of ultrasonic shadows and anatomical structures (such as bifurcations and calcified plaque) complicates the automatic delineation of the lumen contour and EEL. The purpose of this paper is to evaluate the IVUS segmentation performances of different convolutional networks and the impact factors on a large-scale multiple-center dataset. Methods: A total of 6516 cross-sectional images from 175 IVUS pullbacks acquired in different centers by different IVUS imaging catheters were screened from a corelab to evaluate the segmentation methods. The IVUS images included bifurcation, side branch ostia, and various image artifacts to reflect the general image characteristics in routine clinical acquisition. We compared three generic fully convolutional networks (FCNs) and two FCNs specifically designed for the segmentation of IVUS images and explored the factors impacting the segmentation performance, including the training images and the input of consecutive images to the models. The performance of the FCNs was evaluated by using the Dice similarity coefficient (DSC), the Jaccard index (JI), the Hausdorff distance (HD), linear regression and Bland-Altman analysis. Results: The 4-cascaded RefineNet and DeepLabv3+ outperformed U-net and IVUS-net in the segmentation of the lumen contour and EEL on IVUS images. DeepLabv3+ had the best segmentation performance, with DSCs of 0.927 and 0.944, JIs of 0.911 and 0.933, and HDs of 0.336 mm and 0.367 mm for delineation of the lumen and EEL, respectively. Excellent agreement between DeepLabv3+ and the manual delineation was found in the quantification of the coronary plaque area ( r = 0.98). Conclusions: The convolutional network architecture is effective in the automatic segmentation of IVUS images. It might contribute to the clinical application of quantitative IVUS analysis in real-world as well as the efficient assessment of coronary atherosclerosis.