Automatic lumen border detection in IVUS images using dictionary learning and kernel sparse representation. (April 2021)

• Record Type: Journal Article
• Authors: Tong, Jijun; Li, Kai; Lin, Wenting; Shudong, Xia; Anwar, Ali; Jiang, Lurong
• Is Part Of: Biomedical signal processing and control. Volume 66(2021)
• Publication Date: 2021-04
• Subjects: Intravascular ultrasound -- Lumen contours -- Dictionary learning -- Kernel method -- Sparse coding; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102489
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Fully automatic method for the detection of the lumen contours in IVUS images. Sparse coding and kernel dictionary learning are used to classify pixels. Assisting the diagnosis of coronary heart disease and effective interventional therapy. Simple preprocessing helps to reduce the impact of artifacts and calculation cost. Abstract: Intravascular ultrasound (IVUS) images display the cross-sectional information of the lumen, from which the diameter and length of the lumen are accurately measured to calculate the volume. More importantly, they provide the tissue information of the plaque, thus assisting the diagnosis of coronary heart disease and effective interventional therapy. In this study, a fully automatic method is presented for the detection of the lumen contours in IVUS images of the coronary artery. First, texture feature vectors are extracted from the original images with a patch size of 3 × 3. The sparse coding and kernel dictionary learning are used to employ the features to construct two dictionaries for positive and negative tissues, respectively. Then a series of preprocessing helps to reduce the impact of artifacts, calculation cost and get approximate Region of interest (ROI). A kernel-cluster algorithm based on linear discrimination method is developed to classify the pixels in the ROI. In the end, morphological operations are used to improve the detection quality. Publicly available evaluation indicators are applied to the proposed algorithm with 326 test images of different structures. Mean value of the total results (JACC: 0.87, HD: 0.35, PAD: 0.09) outperforms the other automatic methods of the participants in the challenge. Besides, compared with the recent methods used on the same dataset, the proposed method shows good performance and high accuracy. Furthermore, kernel method and preprocessing steps are effective in acquiring better detection results by reducing the influences of artifacts.