基于Transformer和双注意力机制的颈动脉超声造影图像斑块分割方法

Segmentation of Carotid Plaque in Contrast-Enhanced Ultrasound Image Based on Transformer and Dual Attention Mechanism

摘要: 脑卒中是一项重大的公共卫生挑战，同时也是全球导致死亡人数最多的疾病之一。颈动脉粥样硬化斑块与脑卒中等缺血性疾病密切相关。颈动脉斑块的早发现和早治疗对预防未来缺血性脑卒中疾病的发生具有重要意义。超声造影(CEUS)已经成为常见的诊断颈动脉斑块的成像方式，因此从CEUS图像中准确分割动脉粥样硬化斑块对于预防和治疗缺血性脑卒中至关重要。然而，由于斑块边界模糊和图像噪声强烈等原因，CEUS图像颈动脉斑块自动分割面临巨大挑战。因此，如何提高颈动脉斑块分割性能仍然是迫切需要解决的问题。本文提出了一种创新的医学图像分割框架，称为DATU-Net，该框架将Swin Transformer模块和双注意力机制集成到U形架构中，以实现CEUS图像颈动脉斑块自动分割。DATU-Net采用基于Swin Transformer模块构建的编码器，可以有效地建模远程依赖关系和多尺度上下文信息。为了获得更丰富的特征表示，我们在编码器–解码器的跳跃连接中引入了双级注意力(Dual-Level Attention)模块，以增强图像特定的位置特征和通道特征，从而有效提高了斑块分割性能。此外，我们在解码器中引入了Swin Transformer模块，用于进一步探索上采样过程中的全局上下文信息，同时逐步恢复特征图。我们利用实际的临床数据集对提出的框架性能进行了评估。广泛的实验仿真结果显示，本文提出的方法在Dice系数(0.8548)、交并比(0.7632)、精确度(0.8746)和召回率(0.8863)等方面始终优于其他分割网络。这些实验证明了DATU-Net的有效性，为颈动脉超声造影图像斑块自动分割问题提供了一种可行的解决方案。

Abstract: Stroke is a major public health challenge and one of the leading causes of death around the world. Carotid atherosclerotic plaque closely correlates with ischemic diseases such as stroke. Early detection and treatment of carotid plaques are important for preventing future ischemic stroke diseases. Contrast-enhanced ultrasound (CEUS) has emerged as a prevalent imaging modality for the diagnosis of carotid plaques, so accurate segmentation of atherosclerotic plaques from CEUS images is crucial for the prevention and treatment of ischemic stroke. However, automatic segmentation of carotid plaques from CEUS images is tremendously challenging due to blurred plaque boundaries and strong noise in images. Therefore, how to improve the performance of carotid plaque segmentation remains an urgent problem. In this paper, we propose an innovative medical image segmentation framework, called DATU-Net, which aims to integrate the Swin Transformer block and the dual-attention mechanism into a U-shaped architecture for automatic carotid plaque segmentation of CEUS images. DATU-Net employs an encoder built upon the Swin Transformer block, proficient in effectively modelling long-range dependencies and multi-scale contextual information. In order to obtain richer feature representations, we introduce a Dual-Level Attention module in the encoder-decoder skip connection to enhance the image-specific positional and channel features, which effectively improves the performance of plaque segmentation. In addition, we introduce the Swin Transformer block in the decoder for further exploring the global contextual information during the up-sampling process while progressively recovering feature maps. We evaluate the performance of the proposed framework using real clinical datasets. Extensive experimental simulation results consistently show that the method proposed in this paper outperforms other segmentation networks in terms of Dice coefficient (0.8548), intersection over union (0.7632), precision (0.8746) and recall (0.8863). These experiments demonstrate the effectiveness of DATU-Net and provide a viable solution to the problem of automatic plaque segmentation in carotid CEUS images.