基于注意力Unet的多尺度胎儿图像分割方法

Multi-scale Fetal Image Segmentation Method Based on Attention Unet

针对使用传统机器学习方法分割胎儿图像中头部和股骨的精度较低且效果差,提出了一种新型的注意力Unet架构。在注意力Unet中加入了通道注意力机制平均最大模块(AMB),将原有的卷积层模块替换为不同卷积块组合的InceptionV2+模块,并在网络深层处加入了不同尺寸的空洞卷积模块。同时,研究了Dice损失函数和Focal损失函数相结合替换二元交叉熵对图像分割效果的影响。实验结果表明,所提方法对胎儿头部和股骨图像的分割效果良好,在准确率、Dice系数、交并比(IOU)、豪斯多夫距离(HD)评价指标方面优于如今主流的医学图像分割方法。

For the reason that the traditional machine learning method has low accuracy and poor effect in the segmentation of head and femur from fetal images, a novel Attention Unet architecture is proposed. The channel attention mechanism average maximum block(AMB) is added to Attention Unet. InceptionV2+ block with combination of different convolution blocks is used to replace the original convolution layer block, and the dilation convolution blocks of different sizes are added to the deep part of the network. At the same time,the effect of Dice loss function and Focal loss function combination instead of binary cross entropy on image segmentation is studied. The experimental results show that the proposed method has a good effect in segmentation of head and femur from fetal images, and is superior to the current mainstream medical image segmentation methods in terms of accuracy, Dice coefficient, intersection-over-union(IOU) and Hausdorff distance(HD) evaluation indexes.