Skin lesion image segmentation by using backchannel filling CNN and level sets

Pigmented skin lesion image segmentation is a key step in the computer-aided diagnosis of melanoma. This image segmentation task is also difficult due to blurred skin lesion boundaries, uneven color, and diverse tex-tures. To improve the accuracy of skin lesion image segmentation, this paper proposes an image segmentation method that combines convolutional neural networks (CNN) and level sets. First, a backchannel filling (BCF) method is proposed. Through the first forward propagation of the CNN, the target position estimated by the CNN is backfilled into the position prior channel. Then, end-to-end deep learning models based on BCF multibranch encoding-decoding CNN (BCF-CNN) and level sets are built. The initial level set and driving energy are predicted through the second forward propagation of the BCF-CNN and then input into the level set segmentation model to complete the evolution. Additionally, a learning algorithm that combines a joint BCF-CNN and a level set is designed to train the segmentation model. The experimental results of the proposed algorithm are compared with those of other state-of-the-art algorithms on both the International Skin Imaging Collaboration (ISIC) 2017 and 2016 skin lesion challenge datasets, demonstrating that the proposed algorithm outperforms the others in terms of accuracy with a Jaccard index of approximately 77.0% and 85.1% on the ISIC 2017 and 2016 skin lesion challenge datasets, respectively.