Research On Dermatological Medical Image Segmentation Algorithm Based On Semi-supervised Learning

In recent decades,the incidence of skin cancer-related diseases has gradually increased,causing serious harm to the health and life of modern people.For high-risk diseases like skin cancer,early identification and intervention is an important means of treatment.With the continuous deepening of artificial intelligence technology in the field of image recognition,computer programs can effectively analyze lesion changes in skin imaging,assist doctors in diagnosis and treatment.However,commonly used fully supervised deep learning algorithms require a large amount of labeled data for training,which greatly increases the cost and time consumption of model training.The idea of semi-supervised learning proposes to use lowdifficulty unlabeled data to further enhance the segmentation performance of the model.With the assistance of low-level labeled data,the model obtains additional feature information from high-level unlabeled data.Therefore,semi-supervised learning algorithms reduce training costs and can obtain reliable prediction results,making it a promising research direction.Based on the idea of semi-supervised learning,this article proposes two different semantic segmentation models for skin imaging segmentation tasks.The main research contents are as follows:In the task of segmenting skin microscopy images,there are several challenges,such as significant differences in color and shape size in tumor regions,fuzzy tumor edges,and misleading from certain occlusions.To address these difficulties,this article proposes a Confidence-Aware Cross Supervised Model(CCSM)for accurate segmentation of tumor regions in skin microscopy images.The network consists of two parallel and independently initialized segmentation networks.The segmentation networks are based on the DenseUNet model,with a confidence calculation module added to the decoder.The output of one segmentation branch is filtered using the confidence map of the same network to remove low confidence areas and used as pseudo-labels to supervise the other branch.To effectively capture the semantic information of the input image,a simple and effective attention module is also proposed to capture the features of the target region.A series of experiments on the ISIC2017 and ISIC2018 dataset demonstrate the reliability and superiority of this network model.In the task of semantic segmentation of skin lesion images using dermoscopy,both boundary information and overall information of the lesion area have a positive effect on improving the segmentation performance of the model.Most existing models only focus on the extraction and utilization of regional information,and do not make good use of edge information to enhance model performance.To address this issue,this paper designs a Boundary Feature-guided Semi-supervised Self-ensembling Model(BGSM)to achieve automatic and accurate segmentation of skin lesion images.Before training,the network incorporates two data augmentation modules.The color transformation module uses the mutual transformation between the RGB and LAB color spaces to obtain brightness information from unlabeled data and construct new labeled data.The rotation transformation module uses the rotation inconsistency of convolution to apply corresponding rotation operations to the input and output.Both data augmentation modules can effectively enhance consistency loss.The network is built based on ResNet and includes a corresponding decoder at each level.An edge extraction module is built between the encoder and decoder,using the Sobel operator to obtain edge information.In addition,a multi-scale feature fusion module is designed to fuse the outputs of the upper and lower layers of the encoder.Through this progressive feature fusion,the network can obtain richer shallow and deep information.This paper verifies the accuracy and superiority of the proposed network model on the ISIC2017 dataset and the ISIC2018 dataset.