Researchers Propose Generative Adversarial Networks Based Framework for Skin Lesion Segmentation

Date:30-06-2020   |   【Print】 【close

Skin lesion is a primary and essential step of melanoma recognition and diagnosis. 

But different melanoma appearances and low contrast between the lesion region and normal areas makes it hard for precise segment lesion region recognition. 

According to the study published in Medical Image Analysis, researchers from the Shenzhen Institutes of Advanced Technology (SIAT) of the Chinese Academy of Sciences and their collaborators from Shenzhen University proposed a novel and effective automatic segmentation method by using Generative Adversarial Networks with dual discriminators. 

Researchers conducted extensive experiments to evaluate proposed method based on the international skin imaging collaboration (ISIC) skin lesion segmentation datasets released in 2016, 2017 and 2018. 

To estimate the performance of the proposed skip connection and dense convolution U-Net (UNet-SCDC) and dual discrimination mechanism, researchers carried out the ablation study through 5 evaluation criterions, Accuracy(Acc), Jaccard index(JI), Dice coefficient(DC), Sensitivity(Sen) and Specificity(Spec).  

More remarkably, they compared the proposed dual adversarial GAN (DAGAN) with some state-of-the-art segmentation methods, include the top three methods on ISIC 2017, the top five methods on ISIC 2016 and 2018. The experimental results exhibited that DAGAN achieved the best performance on the ISIC 2017 skin lesion segmentation challenge datasets so far. 

"Extensive experimental results demonstrate that our model is more generalized and achieves superior segmentation performance, which shows a bright prospect on skin lesion segmentation, we believe it can be extended to other medical image segmentation tasks in the future," said Dr. WANG Shuqiang, corresponding author of the study.  

Figure. Segmentation performance comparison of different state-of-the-art networks on ISIC 2017 dataset. (Image by WANG Shuqiang) 

Media Contact: 
ZHANG Xiaomin 
Email: xm.zhang@siat.ac.cn