Skin Lesion Image Segmentation Based On Deep Learning

In recent years,with the deep integration of artificial intelligence and various industries,especially the rise of the deep learning research boom,the relationship between deep learning and clinical skin disease diagnosis has become closer.For clinical skin lesion images,effective segmentation of skin lesions can improve the accuracy of the diagnosis of skin diseases,and provide a powerful means for dermatologists to examine pigmented skin lesions.However,due to the low contrast between the lesion area and the surrounding skin in the collected images,the blurring of the lesion boundary,the large difference between melanoma types,and the presence of artifacts,the segmentation of the skin lesion area is challenging.Because of that,this article proposes three efficient and accurate methods for segmentation of skin lesions.The specific content is as follows:Firstly,in view of the fact that the existing network is difficult to accurately segment the complex shape of the lesion area,a U-shaped skin lesion image segmentation algorithm that automatically adapts to the target shape is proposed.First,perform grayscale,normalization,and adaptive histogram equalization processing on the original lesion image in sequence to improve the contrast between the foreground and the background.Then input the pre-processed pictures into the U-shaped network for training.The network fuses the modulated deformable convolution into the U-Net encoder and decoder to automatically adapt to the proportion and shape of the lesion target,making it complicated the lesion structure can be accurately detected.Finally,the segmentation results are obtained through the Soft Max classifier.The experimental results on the ISBI2016 skin lesion image data set show that the algorithm can accurately segment the skin lesion area,and the overall performance is better than existing algorithms.Secondly,an improved UNet++ model is proposed to solve problems such as misidentification and low segmentation accuracy of the standard U-shaped network in the segmentation of skin lesion images.In addition to maintaining the advantages of UNet++multi-scale feature fusion and low image feature loss,the Tversky-Focal loss function is proposed for UNet++ network’s misrecognition of unbalanced images of data categories,which reduces the negative impact of imbalanced data categories.Avoid training difficulties due to drastic changes in gradients;In view of the poor segmentation effect of the network on small targets,so the soft attention gate is introduced.In the lesion image of small targets with large backgrounds,large attention coefficients are given to small targets to improve the segmentation accuracy of the model.The experimental results on the ISBI2017 and ISBI2016 skin lesion image datasets show that all evaluation indicators are better than U-Net and UNet++.Thirdly,aiming at the problem of hair occlusion in the skin lesion area and the single fusion of existing network features,a multi-scale dense network skin lesion image segmentation algorithm is proposed.Firstly,the original skin lesion image is preprocessed with morphological closing operation and non-sharpening filter in sequence,and a refined image without skin hair and vascular artifacts is obtained.Then input the preprocessed image into the segmentation network,which is based on the encoding-decoding architecture and uses two multi-scale feature fusion methods of parallel multi-branch structure and pyramid pooling model to achieve feature extraction under different receptive fields.At the same time,Dense Net is merged into the encoder to realize the multiplexing of image features.In addition,the proposed loss function combining target loss and content loss further improves the accuracy of image segmentation.Finally,the segmentation results are obtained through the Soft Max classifier and the relevant evaluation indicators are calculated.The experimental results on the ISBI2016 skin lesion image data set show that the overall performance of the algorithm is better than the existing algorithms.