Research On Skin Lesion Image Segmentation Based On Deep Learning And Feature Fusion

In recent years,the number of people suffering from melanoma has been on the increasing trend year by year.Melanoma poses a great threat to people’s health.Early diagnosis and treatment of melanoma is very helpful to prolong the life of patients.Segmentation of skin lesions is a critical step before classification can be performed,and good segmentation results will help improve the correct classification of skin diseases and help clinicians make better diagnoses.Manually labeling and processing the collected skin lesions images requires a high level of physician expertise and experience,and is tedious,time-consuming,and error-prone.As a result,the development of algorithms for automatic segmentation of skin lesion images is very necessary.At the same time,skin lesion images have characteristics such as large size variation,irregular shape and boundary,complex background information and blurred boundary,which make the segmentation of skin lesions a very challenging problem so far.Algorithms based on deep convolutional neural networks have shown outstanding effects in the field of computer vision,and moreover,they have become the mainstream algorithms for image segmentation tasks.Many researchers have applied deep learning methods to medical image analysis problems and achieved many excellent research results.On the skin lesion image segmentation task,although the existing segmentation algorithms have achieved good segmentation results,they still have some deficiencies:（1）the feature extraction capability of these models is limited and they do not provide rich feature maps for the segmentation task;（2）the feature fusion methods they used are inefficient and do not select more important feature maps,which also affects the segmentation results of skin lesion images.To address these problems,this paper designs a skin lesion segmentation model with encoder-decoder structure.The main work of this paper can be summarized as follows:（1）Surveyed the current status of domestic and international research on the problem of semantic segmentation and skin lesion image segmentation,and introduced the basic theory and classical models of convolutional neural networks.（2）A skin lesion image segmentation model based on encoder-decoder structure is designed.To enhance the feature learning ability of the model,a modified Efficient Net is used as the backbone network;to use each feature map generated in the encoding stage efficiently,a gated fully feature fusion module is used to supplement the feature information for the decoder;to improve the decoding ability of the decoder,a dense feature fusion method is used to deliver feature maps;the final model showed a better segmentation effect on the skin lesion images,which can be used to build a computer-aided diagnosis system to improve the efficiency of melanoma diagnosis and treatment.（3）In this paper,a series of experiments are conducted on ISIC-2017 dataset and PH2 dataset to verify the effectiveness of the proposed model.The model in this paper achieves 87.55% and 93.01% of Dice Coefficients and 79.06% and 87.52% of Jaccard Coefficients on ISIC2017 and PH2 datasets,respectively,due to the efficient feature fusion approach and stronger decoder it uses,which is better validated in the ablation experiment subsection.The comparison with classical segmentation models FCN-8s,Attention U-Net and some previous related works show the effectiveness of the proposed method.