Abstract:

In order to solve the problem of inaccurate boundary segmentation of skin lesion images, an improved skin lesion segmentation algorithm based on dense lesion convolution network (DenseNet-BC) is proposed. Firstly, the connection way between the traditional algorithm layers is changed. Through dense connections, all layers can directly access the gradient from the original input signal to the loss function, so as to maximize the image feature information. Secondly, in order to reduce the number of parameters and the calculation amount of the network, the small convolution kernel is used in the bottleneck layer and the transition layer to halve the number of channels of the input feature map. The performance of this method is compared with the algorithms of VGG-16, Inception-v3 and ResNet-50 on the ISIC 2018 task 1 skin lesion segmentation data set. The experimental results show that the DenseNet-BC algorithm has the segmentation accuracy of 0.975 and the Threshold Jaccard of 0.835. The segmentation accuracy is significantly improved compared with other algorithms, and it is an effective method for skin lesion segmentation.

Key words: dermoscopic image, skin lesion segmentation, deep learning, dense convolutional network