Research On The Segmentation Method Of Skin Pathology Images Under Complex Conditions

According to statistics,skin cancer is a relatively difficult disease,and its fatality rate is increasing year by year.Among them,the most fatal is melanoma,which seriously threatens our health.However,as long as we can make diagnosis and clinical intervention and treatment timely through scientific and technological means at an early stage,the chance of cure will be much higher.Due to the rapid increase in the number of skin cancers,the demand for diagnosing skin wound lesions with computer is also increasing.At present,skin lesion image segmentation is one of the important methods for clinical diagnosis.However,the latest publicly available dataset of skin lesions has very limited real expert labels.In addition,the available segmentation datasets are jointly annotated by many experts,which means that accurate annotation of the skin lesion boundary is a laborious,time-consuming and expensive task.A computer-aided diagnosis system can not only assist doctors in diagnosis and improve efficiency,but also save human resources to a certain extent.Accurate skin segmentation algorithms can achieve high-sensitivity positioning in dermoscopic images and quickly find the location of the lesion,which plays a vital role in the subsequent diagnosis of different types of skin diseases.Therefore,this paper focuses on the three problems that the lesion area in the current dermoscopic image is highly similar to the healthy skin area,the edge of the skin lesion area is blurry,and the artifact information interference in the dermoscopic image..It mainly includes the following three aspects:First of all,in order to effectively improve the segmentation accuracy of skin lesions,this paper proposes a segmentation model that can filter out noise such as hair.By filtering the interference information in the feature extraction process,it effectively solves the problem such as noise transmission caused by the jump connection operation of the U shape network.The feature purification block proposed in this paper can predict a variety of low-level features during the model training process,and match the predicted features in a weighted manner to obtain more effective feature information.At the same time,according to the principle of similarity,the weight is dynamically adjusted by feedback to weaken the proportion of interference information and enhance the proportion of effective information,so as to match the best effective features.Experimental results show that the model achieves a good segmentation effect on the benchmark ISIC2017 challenge dataset,and achieves the purpose of filtering interference information such as hair and equipment artifacts while learning features.Secondly,considering that the proportion of the area occupied by medical equipment artifacts in the dermoscopic image is relatively large,as the number of layers in the medical image segmentation network deepens,there will be a problem of mixing effective feature information and invalid feature information.Therefore,this paper combines the positional attention mechanism and the channel attention mechanism to capture the relationship between distant pixels to distinguish between artifact characteristics and lesion characteristics.We integrate it into the general encodingdecoding segmentation network,named AF-Unet.The network emphasizes similar local features through the above two mechanisms,and can obtain a global view when dealing with skin artifacts.Regardless of the distance,AF-Unet selectively emphasizes the existence of interdependent channels based on the similarity between artifact features and skin lesion features.Add the output of these two attention modules to further improve the representation of feature information.No matter how far away,similar pixels will be related to each other.The attention mechanism integrates the relevant feature information between all the channel graphs in the channel domain,and selectively emphasizes the existence of interdependent channels.Adding the output of these two attention modules can further improve the representation of feature information,which will be of the great help to the accurate segmentation effect.Finally,for the low-contrast data images that exist in ISIC2017,this paper designs a low-contrast segmentation module.We use nested residual connection operations to replace the jump connection operation of the traditional U-shaped network,which makes it easier to deal with inconspicuous feature information.This can effectively improve the expressive ability of the network without deepening the number of network layers without introducing too many network parameters.With the improvement of presentation ability,the network can further fit such difficult lesions,so as to deal with them more fully and carefully.At the same time,we introduce focal loss to reduce the weight of easy-toclassify sample pixels,so that the model focuses on learning the difficult-to-classify sample pixel features.The new model formed is named the LCSA network,which can improve the images that are difficult to segment under low-contrast conditions in the ISIC2017 dataset,thereby further improving the overall segmentation accuracy.