Journal
IEEE TRANSACTIONS ON ULTRASONICS FERROELECTRICS AND FREQUENCY CONTROL
Volume 68, Issue 5, Pages 1762-1772Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TUFFC.2021.3052486
Keywords
Imaging; Calcium; Kernel; Image segmentation; Gray-scale; Convolution; Annotations; Atherosclerosis; convolutional neural network (CNN); deep learning; intravascular ultrasound (IVUS) imaging; plaque
Categories
Funding
- Show Chwan Memorial Hospital, Changhua, Taiwan [NCKUSCMH10805]
- National Health Research Institutes in Taiwan [NHRI-EX107-10712EI]
- Ministry of Science and Technology of Taiwan [MOST 107-2221-E-006-024-MY3]
- Medical Device Innovation Center (MDIC)
- National Cheng Kung University (NCKU) from the Featured Areas Research Center Program
Ask authors/readers for more resources
An end-to-end deep-learning convolutional neural network was developed to automatically detect various structures in IVUS images, and the performance was evaluated for different loss functions.
Atherosclerosis is the major cause of cardiovascular diseases (CVDs). Intravascular ultrasound (IVUS) is a common imaging modality for diagnosing CVDs. However, an efficient analyzer for IVUS image segmentation is required for assisting cardiologists. In this study, an end-to-end deep-learning convolutional neural network was developed for automatically detecting media-adventitia borders, luminal regions, and calcified plaque in IVUS images. A total of 713 grayscale IVUS images from 18 patients were used as training data for the proposed deep-learning model. The model is constructed using the three modified U-Nets and combined with the concept of cascaded networks to prevent errors in the detection of calcification owing to the interference of pixels outside the plaque regions. Three loss functions (Dice, Tversky, and focal loss) with various characteristics were tested to determine the best setting for the proposed model. The efficacy of the deep-learning model was evaluated by analyzing precision-recall curve. The average precision (AP), Dice score coefficient, precision, sensitivity, and specificity of the predicted and ground truth results were then compared. All training processes were validated using leave-one-subject-out cross-validation. The experimental results showed that the proposed deep-learning model exhibits high performance in segmenting the media-adventitia layers and luminal regions for all loss functions, with all tested metrics being higher than 0.90. For locating calcified tissues, the best result was obtained when the focal loss function was applied to the proposed model, with an AP of 0.73; however, the prediction efficacy was affected by the proportion of calcified tissues within the plaque region when the focal loss function was employed. Compared with commercial software, the proposed method exhibited high accuracy in segmenting IVUS images in some special cases, such as when shadow artifacts or side vessels surrounded the target vessel.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available