Research On Complex Skin Lesion Image Segmentation Method Based On Adversarial Learning

Skin cancer,especially melanoma,which has a very high fatality rate,poses a great threat to the life and health of patients.In clinical practice,the diagnosis of skin cancer mainly relies on the recognition of skin lesion images.However,images of complex skin lesions present the features of various shapes,blurred boundaries,and more hair and blood vessel interference,which affect the efficiency and accuracy of doctors’ diagnosis and treatment.To address these issues,this thesis focuses on the high-quality extraction of skin lesion image features and the efficient utilization of lesion image datasets in complex skin lesion segmentation.The main contents include:(1)A skin lesion segmentation network that combining multi-scale attention and boundary enhancement(BEMA U-Net)is proposed.The network includes a spatial multi-scale attention module for extracting global features and a boundary enhancement module for enhancing the boundary features of lesion regions.These two modules are added to the U-Net backbone network,which can effectively suppress the interference of background noise in the lesion image and enhance the boundary details of the lesion.In addition,a hybrid loss function combining Dice Loss and Boundary Loss is also designed.By dynamically adjusting the weight of the hybrid loss function during the training process,the network can perform multiple supervisions on the extraction of the lesion image’s overall and boundary detail features.(2)In order to further improve the segmentation accuracy of BEMA U-Net,and to address the problem of fewer labeled samples in the skin lesion image dataset,a dual-discrimination skin lesion segmentation generative adversarial network based on feature channel selection(FSD-GAN)is proposed.BEMA U-Net segmentation network is used as the generator,and an adaptive feature channel selection module is added to suppress the interference information of irrelevant channels in the discriminator network.The spatial discrimination strategy in the dual discriminator focuses on the spatial characteristics of the generated results,and the boundary discrimination strategy focuses on the boundary details features of the generated results.The dual discriminators work simultaneously to complement the features that are difficult to capture,thus achieving better segmentation performance.(3)A skin lesion segmentation system based on adversarial learning was designed and implemented,which includes functions such as input of skin lesion images,data preprocessing,segmentation model,and selection of model weights.The system provides an easy-to-use interface for automatic skin lesion segmentation,which can be used as an assistant tool for clinical diagnosis and research purposes.This thesis aims to achieve high-quality extraction of complex skin lesion features,efficient utilization of lesion image datasets and accurate segmentation results by designing deep learning based algorithms.Experimental results on ISIC2017 and ISIC2018 datasets verify the effectiveness of the algorithm proposed in this thesis.The obtained segmented images have continuous boundary and clear outlines.In addition,the skin lesion image segmentation system based on the proposed algorithm in this thesis can be applied to practical clinical diagnosis and treatment.