基于MCTUnet的双模态超声图像的颈动脉斑块分割

Carotid Plaque Segmentation in Dual-Modal Ultrasound Images Based on MCTUnet

摘要: 超声检查(US)和对比增强超声(CEUS)是分析病变的空间和时间特征，以及诊断或预测疾病的重要成像工具。然而，超声图像通常存在边界模糊和噪声干扰强等特点，因此，逐帧评估斑块并描绘病变是一项繁琐且耗时的任务，这对利用深度学习技术分析超声视频提出了挑战。尽管如此，现有的超声和对比增强超声图像分割方法中，能够有效融合这两种不同类型图像的特征信息的仍较为稀缺，且这些方法在全局上下文信息的提取能力上需优化。为此，本文提出了一种基于改进Transformer的混合卷积自注意力(MCT)U形结构双分支网络模型，并在跳跃连接处引入卷积注意力模块。MCT的设计旨在结合卷积和自注意力的优势，不仅增强了全局上下文信息的捕捉能力，同时也保留了卷积方法的良好归纳偏置。实验结果表明，所设计的网络在颈动脉数据集上的表现优于临床医生的诊断结果。

Abstract: Ultrasound (US) and contrast-enhanced ultrasound (CEUS) are essential imaging tools for analyzing the spatial and temporal characteristics of lesions, as well as for diagnosing or predicting diseases. However, ultrasound images are often characterized by blurred boundaries and significant noise interference, making the task of frame-by-frame plaque assessment and lesion depiction labor-intensive and time-consuming. This presents a challenge for using deep learning techniques to analyze ultrasound videos. Despite advancements, there are still few segmentation methods capable of effectively integrating feature information from both types of images, and existing approaches require further optimization in their ability to capture global contextual information. To address this, we propose a U-shaped dual-branch network model based on a hybrid Convolutional Self-Attention (MCT) framework with an improved Transformer, incorporating convolutional attention modules in the skip connections. The MCT framework is designed to leverage the strengths of both convolution and self-attention, enhancing the ability to capture global contextual information while retaining the beneficial inductive biases of convolution. Experimental results demonstrate that our designed network outperforms clinical experts on the carotid artery dataset.