Research Of Skin Lesion Images Segmentation Based On Class Attention And Recurrent Convolution

Melanoma caused by malignant lesions of skin melanocytes has a high mortality rate.Early diagnosis and treatment of melanoma are essential to effectively reduce the mortality of the disease.At present,the main diagnostic method of skin pigmented lesions is to show clearly pathological features of the skin lesion by using dermatoscopy,and professional dermatologists make a judgment based on rich clinical experience.This method mainly depends on the clinical experience of doctors,has a certain degree of subjectivity,and diagnosis for a large number of patients is a tedious and timeconsuming process.Through the construction of a computer-aided diagnosis system for skin lesions to assist doctors in diagnosis,doctors can be freed from the repeated timeconsuming screening and diagnosis work,and focus on the treatment process of diagnosed patients,and strive for precious treatment opportunities for patients.As an important part of the computer-aided diagnosis system,the accurate segmentation of skin lesions is still an unsolved challenge due to the complex and changeable characteristics of the lesion area and the interference of many external factors.In recent years,the deep learning method has been used for medical image segmentation,and its results have been significantly improved in various evaluation indicators,showing the possibility of solving the problem of skin lesions image segmentation.In this paper,based on deep learning,the class attention feature extraction module of skin lesions was constructed to extract the attention class feature map of skin lesions from the coarse segmentation results.And an iterative optimization strategy for skin lesion image segmentation was developed,and the attention class feature map of skin lesions was fused with multi-scale feature information and integrated into the network through recurrent supervision.In this way,the segmentation results of skin lesions can be gradually refined and accurate segmentation results can be obtained.The main work of this paper is as follows:（1）A compact skin lesion image segmentation network,named RCA-Net,was proposed to solve the problems of small dataset and skin lesion with blurred boundary.By constructing the feature extraction module of class attention,the attention class feature in coarse segmentation results was extracted and integrated into the network through multi-scale recurrent supervision.In the formulated iterative optimization strategy,the model obtained accurate segmentation accuracy and high generalization ability.（2）Due to the small proportion of lesions in skin lesion images,the paper added dice loss on the basis of cross entropy loss function to solve the imbalance of positive and negative samples in skin lesion segmentation.In addition,the level set auxiliary loss function was used to introduce the supervision of the lesion boundary in the skin lesions segmentation task to further improve the segmentation performance of skin lesions at the boundary.（3）In order to verify the segmentation performance and generalization ability of the model,RCA-Net was trained and tested on isic-2017 dataset,and the model generalization ability was tested on PH2 dataset.The experiment shows that compared with the existing optimal model for skin lesion segmentation,the RCA-Net has achieved a lead in multiple indicators while ensuring the model is lightweight.The paper aims to implement a deep learning algorithm that can accurately segment skin lesions in complex scenes,and reduce the number of model parameters while ensuring the accuracy of the model,so as to improve the practicality of the model in a computer-aided diagnosis system.