基于多尺度编码-解码网络的皮肤病变图像分割

Segmentation of skin lesion image based on multi-scale encoder-decoder network

针对皮肤病变图像分割在医疗诊断中的作用,提出一种基于多尺度编码-解码网络的皮肤病变图像分割算法。该算法继承了SegNet网络结构的训练速度快、训练模型存储小等特点,采用多尺度输入的方式增强了网络对皮肤病变图像的充分学习。此外,在编码网络中的pool2层输出一个二进制双线性插值的中间预测特征图到解码层的最后一层卷积块进行级联输入提高最终的分割精度。实验结果表明,采用多尺度编码-解码网络对皮肤病变图像分割具有极好的效果,在其他医学图像分割方面也能进行广泛应用。

An image segmentation algorithm based on multi-scale encoder-decoder network is proposed for the segmentation of skin lesion image in medical diagnosis.The proposed algorithm inherits the characteristics of SegNet network structure,such as fast training speed and small training model storage.And the multi-scale input method enhances the ability of network to comprehensively learn the skin lesion image.In addition,the output of a layer of binary bilinear interpolated intermediate prediction features to the final layer of convolution blocks of the decoder layer in the pool2 layer of the encoder network is cascaded to increase the final segmentation accuracy.The experimental results show that using multi-scale encoder-decoder network can achieve an excellent segmentation of skin lesion image,and that the proposed network can also be widely used in other medical image segmentations.