A region of interest focused Triple UNet architecture for skin lesion segmentation
Guoqing Liu, Yu Guo, Qiyu Jin, Guoqing Chen, Barintag Saheya, Caiying Wu
Abstract
Abstract Skin lesion segmentation is a crucial step for skin lesion analysis and subsequent treatment. However, it is still a challenging task due to the irregular and fuzzy lesion borders, and the diversity of skin lesions. In this article, we propose Triple‐UNet, an organic combination of three UNet architectures with suitable modules, to automatically segment skin lesions. To enhance the target object region of the image, we design a region of interest enhancement module (ROIE) that uses the predicted score map of the first UNet. The enhanced image and the features learned by the first UNet help the second UNet obtain a better score map. Finally, the results are fine‐tuned by the third UNet. We evaluate our algorithm on a publicly available dataset of skin lesion segmentation. Experiments have shown that TripleUNet achieves an accuracy of 92.5% on the ISIC‐2018 skin lesion segmentation benchmark, with Dice and mIoU of 0.909 and 0.836, respectively, which outperforms the state‐of‐the‐art algorithms.