Research On Algorithm For Dermoscopy Image Segmentation On Generative Adversarial Networks

Melanoma is a kind of skin cancer with a high mortality rate.In recent years,the number of cases has increased year by year.Melanoma,if detected in its early stages of growth,is highly curable,dermoscope image plays an important role in stage of early diagnosis of melanoma.Computer-aided diagnosis can provide objective diagnosis of melanoma.As a key step in the dermalogical image diagnosis process,the segmentation of the lesion area has a direct impact on the subsequent diagnosis.There are many difficulties in medical image processing,such as much noise,uneven grayscale,shape changes of organs and difficult extraction of features.Traditional segmentation method is not effective.Image segmentation based on deep learning can solve the above difficulties and improve the accuracy of image segmentation by using its powerful feature extraction ability and automatic diagnostic ability.In this paper,a segmentation algorithm based on generative adversarial networks for lesion segmentation is proposed,which aims at the problems of high complexity,long computation time and poor segmentation of difficult images.First,a segmentation network is generated based on the Unet network,a certain Gaussian noise is added and using Batch Normalization to improve the stability of the training and avoid the occurrence of overfitting.Secondly,abandon pooling layer and use a full-convolutional layer connection to generate a adversarial network,and achieves a pool-like effect through the strides of the convolution layer,so that the network can learn its own spatial sampling.In order to improve the accuracy and convergence speed,a new loss function is designed for adversarial network and gan network.The algorithm proposed in this paper has a high accuracy in the segmentation of difficult images,such as low-contrast images,and it can achieve automatic end-to-end segmentation.In training process,the four dimensional vector data of original image and tag image can be input into the antagonism network,and the segmentation ability of the network can be improved.Through the test of ISBI2016 skin lesion competition data,the proposed method can correctly segment the damage area,compared with the SegNet,Unet which are commonly used in medical image segmentation and the results of some existing dermoscopic image segmentation methods.The algorithm is obviously superior to these methods in terms of segmented evaluation criteria or the effect of segmentation.