针对 U-Net 分割算法无法提取多尺度特征、易受到伪影和噪声干扰而导致在肺部 X 射线图像中肺实质分割不精确的问题，提出一种基于选择性自校正卷积的U-Net改进算法。改进后的U-Net算法将普通卷积模块替换为选择性自校正卷积模块，该模块采用多分支结构提取多尺度特征信息，使用 Sigmoid 函数和 Softmax 函数对多尺度特征信息进行选择性校正，使校正后的特征信息聚焦于肺实质区域，输出特征更加具有针对性。实验表明，该方法对骰子系数、交并比、F1评分结果以及对肺实质分割结果都有一定程度的提升。
Lung Parenchyma Segmentation of Lung X-ray Images Based on Selective Self-Calibration Convolution U-Net
Aiming at the problem that U-Net segmentation algorithm cannot extract multi-scale features and is susceptible to artifacts and noise,which leads to imprecise segmentation of lung parenchymal in lung X-ray images,an improved U-Net algorithm based on Selective Self-Calibration convolution is proposed in this article.The improved U-Net algorithm replaces the common convolutional module with the Selective Self-Calibration convolution module,which adopts a multi-branch structure to extract multi-scale feature information,and uses Sigmoid function and Softmax function to selectively correct the multi-scale feature information,so that the corrected feature information focuses on the lung parenchyma region and the output features are more targeted.Experiments showed that this method brought some improvement on dice coefficient,intersection over union,F1 score and improved the segmentation accuracy of the lung parenchymal.
国家科技期刊平台
NSTL
万方数据
