关键词: 关键词: 颈动脉斑块分割, 半监督, 双编码器, 多尺度融合, 级联通道空间注意力

Abstract: Abstract: Segmenting carotid atherosclerotic plaques from ultrasound images is valuable for the diagnosis and early prevention of ischemic stroke. However, ultrasound images often suffer from strong noise, low quality, and blurry boundaries, making reliable annotation time-consuming and labor-intensive. To address these issues, this paper proposes a semi-supervised segmentation method called Uncertainty Rectified Pyramid Consistency (URPC) that utilizes unlabeled data to train the model. The URPC method calculates the loss by minimizing the difference between the predictions at different scales and their average. Then, to tackle the issue of frequently blurry plaque regions, a dual-encoder structure based on edge detection is introduced to assist the feature encoder in capturing edge information from ultrasound plaque images. Moreover, as plaque shapes in ultrasound images are often irregular, a multi-scale fusion module (MSFM) is designed to adaptively fuse multi-scale features and improve the extraction of irregular-shaped plaques. A cascaded channel spatial attention (CCSA) module is also incorporated to better focus on the plaque regions. Finally, the proposed method is evaluated on a dataset of ultrasound carotid plaque images. The experimental results demonstrate that the proposed method achieves a Dice score of 81% on the private dataset, which is 1.8 percentage points higher than the suboptimal results obtained by the Cycle-Consistent Prototype Consistency Learning (CPCL) method. Thus, the effectiveness of the proposed method is validated.

Key words: Keywords: carotid plaques segmentation, semi-supervised, dual encoder, multi-scale fusion, cascaded channel spatial attention