Research On Skin Mirror Image Segmentation Based On Deep Learning

Skin disease is one of the most common diseases in life,and its incidence has been increasing in recent years.Due to the strong predictability of skin diseases in the early stage,early intervention can effectively improve the success rate of cure.Before computer-aided dermoscopic lesion segmentation,the clinical processing method of dermoscopic images was to manually segment the lesion area by experienced surgeons,and the segmentation results were highly subjective and time-consuming,so the objective and rapid computer-aided Diagnosis is becoming more and more important.With the development of deep learning,machine vision-related technologies are changing rapidly,and image segmentation algorithms have also made great progress in the application of clinical dermatology diagnosis.Compared with previous traditional image processing and traditional machine learning methods,deep learning algorithms have stronger technical advantages in terms of final effect and generalization.However,in the process of dermoscopic image acquisition,there will be problems such as blurred edges of lesions,complex and changeable boundaries,and hair interference,which makes the segmentation of skin lesions extremely challenging.In this context,to better meet the clinical needs,this paper proposes two segmentation algorithms for skin lesions to address the above problems.The specific contents are as follows:First,because of the problems of blurred boundaries and large differences in the area of lesions in clinically collected dermoscopic images,a segmentation algorithm for double U-shaped skin diseases based on multi-scale feature fusion is proposed.The algorithm consists of two parts: a coarse U-shaped network and a subdivided U-shaped network.First of all,the U-shaped network coding part uses the pre-training model to extract the multi-scale features of the relevant features.In the decoding stage,the improved attention residual block is used to effectively map and fuse the low-level and high-level information to obtain preliminary prediction results.The preliminarily generated prediction results are then aggregated with the original image and input into the multi-channel feature extraction encoder for secondary feature distillation.The subdivision U-shaped network decoder fuses the coding part of the coarse U-shaped network and the coding part of the subdivided U-shaped network at the same time to ensure that the network can aggregate more effective information.Finally,the Focal Tversky loss function is used to further improve the segmentation effect.Experiments show that the algorithm can effectively segment dermoscopic lesions.Secondly,a skin lesion segmentation algorithm based on context feature extraction is proposed to solve the problems of low contrast in the lesion area and hair interference in the process of skin lesion image segmentation.First,the input image is preprocessed to obtain a refined image with enhanced contrast.Then,the preprocessed image is input into the network,and the multi-receptive field residual coding module of the coding part of the network fully extracts the features while taking into account the speed.At the same time,the feature fusion module at the bottom consists of an asymmetric fusion non-local module and an asymmetric pyramid non-local module,which are used to fuse the contextual features of the image.The decoding part is composed of multiple micro U-shaped network modules,which ensures that the low-level features and high-level mapping features are effectively fused and deeply re-extracted.The algorithm is simulated in ISIB 2016 and PH 2,and its experimental results are superior to various current skin lesion segmentation algorithms.