Research On Dermoscopic Images Segmentation Algorithm Based On Deep Convolutional Network

In recent years,the prevalence of skin diseases has been on the rise.Among them,skin cancer has the characteristics of easy metastasis and strong aggressiveness,and has become one of the main causes of cancer deaths.The clinical treatment of skin lesions mainly relies on timely detection of the condition and delineation of the boundary of the lesion to accurately locate the cancerous area.Advanced dermoscopy technology can easily obtain high-definition skin images of patients,but the diagnosis process still needs to segment the lesions in the dermoscopy images.Manually marking the lesion area of the dermoscopy image not only requires the doctor’s professional ability,but also the division process is very cumbersome,consumes a lot of time and energy of the doctor,and is prone to mislabeling.In order to improve the efficiency and quality of doctors in diagnosing skin diseases,this article aims to study an effective segmentation method to assist doctors in diagnosing skin diseases.Traditional image segmentation methods are poor in robustness and generalization.The effect of segmentation is largely dependent on the algorithms designed and extracted by experts,and the implementation of these algorithms is quite complicated.Nowadays,image segmentation methods based on deep learning emerge in endlessly,and have been widely used in the segmentation task of medical imaging lesions,and as a key component of auxiliary diagnosis and treatment.The main work of this article is divided into the following three points:(1)This paper proposes a new deep U-shaped network(Deep-Unet)based on the encoding-decoding structure to automatically and accurately segment the lesion area of the dermoscopic image.The encoder part of the U-Net network is replaced with a Res Net network structure of different depths,and the U-shaped network is adjusted to obtain the Res-U-Net network.Experiments have verified that within a certain range,the segmentation accuracy of the Res-U-Net network increases with the increase in the number of backbone network layers.On the basis of weighing the complexity of time and space,finally decided to use resnet34 as the backbone network for feature extraction.In order to make the final output layer contain richer feature information,this paper designs a jump connection to transfer the features output by the decoders of different levels of the Res-U-Net network to the output layer,thereby determining the basic model of the Deep-Unet network.(2)This paper designs three schemes for optimizing the basic model of Deep-Unet network:(a)Explore the introduction of residual learning in the decoder part to ensure that the network better returns the gradient and avoids the problem of gradient disappearance,so that the network can learn more stably;(b)Explore the use of CBAM hybrid attention mechanism in different positions of the basic model to improve the network’s capture of key features as much as possible,suppress the interference of irrelevant information,and improve the learning performance of the model;(c)Explore in the atrous spatial pyramid pooling operation is introduced at different locations of the basic model to aggregate the characteristics of the upper layer network to obtain richer context information and pass it to the next layer network to improve the segmentation accuracy of the model.(3)This paper uses the ISIC-2017 dermatoscope image data set to verify the proposed network with a series of experiments.The experimental results show that the proposed network model achieves a competitive segmentation performance,and the Jaccard coefficient and Dice coefficient index reach 0.7858 and 0.8645,respectively.At the same time,compared with similar research algorithms,the Jaccard coefficient and Dice coefficient are increased by1.3% and 0.6% respectively,which further confirms the superiority of the method in this paper.