Study On Skin Lesion Segmentation Based On Deep Learning Methods

Skin melanoma is one of the malignant cancerous symptoms that threaten people’s health.Early diagnosis and treatment can not only reduce disease suffering,but can even heal it and bring about better quality of health for the population.With the help of medical technology such as dermoscopy,segmentation of skin lesions can provide clinicians with reliable pathological information to support the treatment.Deep learning methods have been developed to a considerable level for the segmentation task of dermoscopic images.Due to a variety of issues,however,there are still significant challenges in accurately segmenting skin lesions to meet clinical diagnostic needs.These issues are mainly as follows: Firstly,due to the variable size,irregular morphology,irregular distribution and hair interference of skin lesions,loss of features and erroneous predictions often arise when segmentation;the shortage of datasets also imposes limitations on the performance of deep learning algorithm.Secondly,the main of the skin lesion is often blurred,with complex boundary features and variable colour differences between the lesion area and the skin.This poses a great challenge to lesion boundary prediction,and large segmentation errors.Thirdly,prevailing deep learning methods usually have complicated algorithm structures and huge number of parameters,and thus have high hardware dependency and long training time,which are often difficult to realize for clinical applications.Therefore,this paper designs and constructs related algorithms to improve the accuracy and efficiency of lesion detection.The details are as follows:(1)A Dual-branch Feature Extraction Network(DFE-Net)is proposed for capturing lesions information effectively.The network effectively extracts skin lesions by introducing two types of modules.One is the Transformer-based GAAM(Gated Axial-Attention Feature Extraction Module)feature extraction module,which is mainly used to capture the global spatial feature information of the lesion.Another is the convolutional attention feature extraction module E2CAM(Enhanced Efficient Channel Attention Feature Extraction Module),which is mainly used to extract local feature information.The network is designed with skip-connections from the E2 CAMs to the up-sampling stage to compensate for the loss information.Experiments on datasets ISIC-2018,ISIC-2016 and PH2 demonstrate that DFENet outperforms previous deep learning methods with good predictive segmentation performance.(2)A Multi-attention mechanism and Multi-feature interaction Network(Ma Mfi-Net)is proposed for the inaccuracy of lesion image segmentation caused by blurred lesions,complicated boundaries and diverse color differences.The feature encoding part of the network consists of the Gated Axial-Attention Feature Extraction Module(GAAM)and the Hybrid Attention Feature Extraction.The HAM contains a weight map generated by High Efficiency Channel Attention to improve the feature extraction accuracy;the GAAM extracts the input features twice and generates spatial weights to be passed into the HAM module to improve the feature extraction accuracy again from a global basis.In the feature decoding stage,the weight information generated from the high-level features interacts with the skip connections from the encoding stage for feature interaction,enabling the retention of the significant content of the low-level features.To reduce errors in general and improve segmentation accuracy.Extensive experiments on datasets ISIC-2016,ISIC-2017 and ISIC-2018 shows that the algorithm can obtain segmentation results closer to real lesions with minimal error.(3)An Efficient Multi-attention Convolutional Neural Network for Rapid Skin Lesion Segmentation(Rema-Net)is proposed for dealing with the issues of large number of parameters,complex structure and difficulty in meeting clinical needs of deep learning algorithms.The feature extraction module C-GMP(Convolution and global max-pooling layers)of Rema-Net is simplified to a convolutional layer and a max-pooling layer;while the feature extraction stage contains the weight information generated by C-ECA(Convolution and efficient channel attention resampling module).To preserve more information and enhance the ability to capture small lesions,the RAF(Reverse Attention Feature Interaction Module)is introduced to achieve high and low-level feature interaction.Experiments on datasets ISIC-2016,ISIC-2017,PH2,ISIC-2018 and HAM10000,all show that compared to U-Net,Rema-Net reduces parameters by nearly 40% and increases the training speed by 7%while ensuring accuracy in predicting lesions.