Research On Deep Convolutional Networks Applied To Skin Lesion Segmentation In Dermoscopic Image

Melanoma is a malignant skin tumor caused by an abnormal growth of melanocytes in the uppermost epidermis of the skin.It has a low incidence but a high fatality rate.Early diagnosis of melanoma is crucial to the treatment of patients.With the help of a noninvasive,effective,and convenient dermoscopy,clinical dermatologists have greatly improved diagnosis accuracy of melanoma.Due to the influence of clinical diagnosis experience,working environment,labor intensity and other factors,the diagnosis results show that the same doctor at different times or different doctors at the same time has a large deviation,which leads to a high rate of missed diagnosis and misdiagnosis.However,computer-aided diagnosis of melanoma can effectively relieve the work intensity of doctors,reduce the differences between or within diagnostics,and promote the level of diagnosis and treatment.Skin lesion region segmentation,as a key and basic step in computer aided diagnosis system,can remove the irrelevant background noise and improve the accuracy of skin cancer classification.Therefore,it is of great clinical significance to study the accurate segmentation algorithm of skin lesions.In this work,a Multi-scale Context-guided Network(MSCGnet)is proposed for the challenging task of skin lesion segmentation in dermoscopic images.In MSCGnet,the context information is added to the multi-scale feature layer of the U-shaped backbone network to enhance the ability of skin lesion detection and thus improve the segmentation results.In the encoding path,multi-scale context information is used to guide the early feature learning,so that it can learn more discriminative features.In the decoding path,the context-based attention structure is designed to filter the multi-scale information before it is integrated,so that the confidence of the added information is higher.In addition,we applied an iterative mechanism to gradually improve the segmentation results,and MSCGnet was upgraded to iMSCGnet.The training of iMSCGnet adopts the deep hierarchical supervision mechanism to make the network learn more discriminative features.The experimental results of iMSCGnet on ISBI 2016,ISBI 2017,ISIC 2018 and PH2 datasets confirm that iMSCGnet is robust and accurate in detecting skin lesions in dermoscopic images.As the cascading model cannot mine the category information of difficult pixels in the skin lesion region well,a Progressive Content Guided Recurrent U-net(PGRUnet)is proposed in this paper to segment the skin lesion in the dermoscopic image.Unlike traditional cascading methods,PCG-RUnet specifically achieves pixel-level segmentation from easy to hard.Because it is easier to learn a uniform distribution than to learn a long-tail distribution directly,PCG-RUnet makes progressive segmentation with the guidance of different content curriculum.Considering that pre-defined curriculum is time consuming,laborious and challenging,this paper adopts U-Net++to automatically learn curriculum from raw medical data.In PCG-RUnet,the recurrent U-Net framework implements a progressive segmentation from easy to hard.Experimental results of PCG-RUnet on ISBI 2017,ISIC 2018 and PH2 databases demonstrate that the proposed method is robust and accurate in detecting skin lesions in dermoscopic images,and achieves higher segmentation accuracy than the cascaded model.