Intravascular optical coherence tomography (IVOCT) is an advanced imaging technique that enables clear visualization of the contours and morphology of calcified coronary artery plaques, thus aiding in the diagnosis of coronary artery disease and the evaluation of percutaneous coronary intervention (PCI). However, each pullback scan generates 300?500 images. During PCI procedures, comprehensively analyzing such a large volume of images is challenging, and inconsistencies in annotations may exist between observers and the same observer at different times. Hence, a fast, accurate, and efficient approach for the automatic segmentation and evaluation of calcified plaques during surgery must be adopted. Therefore, this study proposes a convolutional neural network, named CAB-U-Net (context fusion transformer-atrous spatial pyramid pooling-bidirectional feature pyramid network), based on IVOCT images for the automatic segmentation of coronary artery calcified plaques via the integration of contextual information.

The proposed CAB-U-Net network for coronary artery calcified plaque segmentation is an improvement of the U-Net architecture. The network primarily comprises Conv2D Block, context fusion transformer (CFT), atrous spatial pyramid pooling (ASPP), and Bi-directional feature pyramid network (BiFPN) modules. The Conv2D Block comprises convolutional, batch normalization, and Sigmoid linear unit (SiLU) activation layers. It aims to enhance feature extraction, accelerate neural-network training, and improve model generalization. The CFT module accurately manages the contextual relationships within sequences via position encoding. It utilizes contextual information between input keys to guide the learning of dynamic attention matrices, thereby enhancing the feature-extraction capability. Additionally, the ASPP introduced in CAB-U-Net enlarges the receptive field through dilated convolutions to capture contextual information at different scales without increasing the network parameters and computational complexity. Furthermore, to strengthen the transmission and fusion of information between feature maps at different levels and reduce information loss, CAB-U-Net adopts a BiFPN module.

Using the same experimental setup and a dataset comprising 2181 image samples, CAB-U-Net was compared with mainstream networks, including PSPNet, DeepLabv3, U-net, SwinUnet, and TransUnet. CAB-U-Net achieves an IOU of 0.9065, a precision of 0.9332, a recall of 0.9662, and an F1-score of 0.9494, which surpass the results of TransUnet, i.e., a suboptimal network, by 0.0228, 0.0240, and 0.0128, respectively. Although the precision of CAB-U-Net is slightly lower than that of SwinUnet by 0.0012, by comprehensively considering the four abovementioned metrics, CAB-U-Net offers outstanding segmentation performance. The superiority of CAB-U-Net over U-net and TransUnet in terms of segmentation is shown in Figs. 6 and 7, respectively. Compared with U-net, our network, which is constructed based on the CFT module, shows higher IOU, precision, recall, and F1-score by 0.0718, 0.0605, 0.2834, and 0.1768, respectively. This indicates that the proposed CFT module enhances the image feature-extraction capability, thereby improving the ability of the network in capturing calcified plaque lesions. After adding the ASPP module to the CFT module, the network model shows higher IOU, precision, and F1-score by 0.0262, 0.0301, and 0.0156, respectively, whereas its recall decreases by 0.0010. This suggests that the ASPP, which utilizes parallel dilated convolutions with multiple sampling rates, extends the receptive field and captures a wider range of contextual information, thereby acquiring multiscale object information. Furthermore, adding the BiFPN module to the CFT module increases the IOU, precision, and F1-score by 0.0299, 0.0354, and 0.0176, respectively, and reduces the recall by 0.0028. This indicates that the BiFPN can learn and transmit semantic and lesion-position features at different scales better, thus effectively fusing lesion-feature information at different scales and improving the networks segmentation performance for lesions of different scales. Finally, combining the CFT, ASPP, and BiFPN, CAB-U-Net yields superior overall performance. Incorporating the ASPP and BiFPN effectively enhances the extraction and fusion of multiscale information, enables the model to learn richer and more discriminative features, and improves precision. However, an increase in the model complexity may cause overfitting, thus deteriorating the recall. Based on the comprehensive metric F1-score, incorporating the ASPP and BiFPN strengthens the improvement in the F1-score.

The CAB-U-Net proposed herein is a convolutional neural network that integrates contextual information for the automatic segmentation of coronary artery calcifications. CAB-U-Net introduces the CFT module, which accurately manages contextual relationships within each position sequence and guides the learning of dynamic attention matrices to enhance feature-extraction capability. The ASPP module is incorporated to utilize dilated convolutions to expand the receptive field and capture contextual information at different scales. Additionally, the BiFPN is adopted to enhance the transmission and fusion of features between feature maps, thereby reducing information loss and achieving more effective feature propagation. Compared with U-net, CAB-U-Net yields higher IOU, precision, recall, and F1-score by 0.1073, 0.1037, 0.2781, and 0.1972, respectively. Compared with other mainstream segmentation networks, CAB-U-Net exhibits significant advantages. Results of correlation and Bland-Altman analyses show that the area and angle of calcified plaques segmented by CAB-U-Net are consistent with those annotated by experts. Therefore, the proposed CAB-U-net is suitable for the segmentation of calcified plaques in IVOCT images, thus providing objective evidence for the clinical diagnosis of calcified plaques, assisting in the comprehensive assessment of coronary artery calcification lesions, and providing guidance for stent implantation.