用于超声图像分割的自适应混合注意力算法
An Adaptive Hybrid Attention Algorithm for Ultrasonic Image Segmentation

乳腺超声图像通常包含丰富的结构和纹理，并且这些结构具有不同的形状、大小和对比度，所以准确分割出乳腺超声图像中的病灶区域是一项具有挑战性的任务。在本研究中，我们提出了一种创新的方法，旨在提高乳腺超声图像分割任务的性能。我们的方法结合了混合注意力编码器和自适应深度监督技术，可以有效地处理乳腺超声图像的复杂结构和多尺度信息。首先，我们提出的混合注意力模块引入了空间和通道注意力机制，用于更好地捕获图像的空间信息和通道特征。其次，我们采用了深度监督技术使解码器输出不同尺度大小的分割图，每个分割结果与真实标签计算损失。最后，我们引入了混合损失函数，将多尺度分割结果结合在一起，使模型自动权衡不同尺度信息的重要性。大量实验表明，我们的网络在定量分析和视觉效果方面都提高了乳腺超声图像分割的性能。
Breast ultrasound images often contain a wealth of structures and textures, and these structures have different shapes, sizes, and contrasts; so accurately segmenting focal areas in breast ultra-sound images is a challenging task. In this study, we propose an innovative approach aimed at improving the performance of the breast ultrasound image segmentation task. Our approach combines hybrid attention encoders and adaptive deep supervision techniques to efficiently process the complex structure and multi-scale information of breast ultrasound images. Firstly, our proposed hybrid attention module introduces spatial and channel attention mechanisms to better capture spatial information and channel features of images. Secondly, we use the deep supervision technique to make the decoder output the segmentation graph of different scales, and calculate the loss of each segmentation result with the real label. Finally, we introduce a mixed loss function to combine the multi-scale segmentation results so that the model can automatically weigh the importance of information at different scales. A large number of experiments have shown that our network improves the performance of breast ultrasound image segmentation in both quantitative analysis and visual effects.