Research On Dermoscopic Image Segmentation And Classification Method Based On Deep Convolutional Network

At present,skin cancer has become the cancer with the highest incidence in the world.For skin cancer patients,only early detection and removal of the diseased area can prevent the spread of cancer cells,thereby avoiding life-threatening.The identification of skin lesion areas and the diagnosis of benign and malignant,early and late stages play an extremely important role in the survival of patients with skin diseases.However,in clinical practice,due to the complexity of the lesion and the significant increase in the number of dermoscopic images,manual inspection may be subjective,time-consuming,and unreproducible.Therefore,research on the segmentation of lesion areas and classification of skin lesions in dermoscopic images has become a hot issue in the field of computer-aided diagnosis.At present,some methods such as threshold-based methods,active contour-based methods,K-nearest neighbor algorithm,and support vector machines have been used to segment and classify dermatoscope images,but these methods can only learn the shallow features of the data.These algorithms cannot find many deep-level features hidden in medical images and rely heavily on manual extraction of features,resulting in poor dermoscopic image segmentation and classification performance.In response to the above problems,this paper combines the deep convolutional network model to study the dermoscopic image segmentation and classification methods.The main research contents of the thesis are as follows:(1)This paper proposes an attention-guided dermoscopy image segmentation network CSAG&DCCNet(channel & Spatial Attention-guided network with Densely Connected Convolutional)with densely connected convolution.The network uses U-Net as the basic network structure,and builds an image pyramid input on the left side of the coding layer to achieve a multi-scale layer of multi-level receiving field fusion;on the right side of the decoder,it introduces a side output layer,average all side output maps as the final prediction map.Simultaneously use the multi-label loss function to update the parameters to train the model;In the last step of the CSAG&DCCNet encoding path,densely connected convolutions are used to replace ordinary convolutional layers,and through the idea of "collective knowledge",the gap between low-level features is bridged and features are effectively aggregated.Densely connected convolution can effectively solve the problem that the spatial resolution of features is continuously reduced during the downsampling process,and the spatial position information is continuously lost.We verified the effectiveness of this method through ablation experiments,and discussed that when the number of densely connected convolutional blocks is 3,the best segmentation effect can be achieved;(2)Design a novel attention-oriented filter module Channel Spatial Fast Attention-guided Filter(CSFAG),and embed it in the skip connection of CSAG&DCCNet segmentation network.This filter can recover spatial information by filtering low-resolution feature maps and high-resolution feature maps,and combine structural information at various resolution levels to filter out noise from the background and reduce boundary blur problems caused by upsampling.On the ISIC-2017 dermoscopic image standard data set,through a large number of ablation experiments,we have verified that the CSFAG module is superior to other mainstream attention modules and can be combined with multiple basic segmentation networks,which can effectively improve the segmentation performance of the model.The segmentation performance of CSAG&DCCNet was compared with other latest algorithms,and the six indicators of accuracy,sensitivity,specificity,Dice coefficient,Jaccard coefficient,and Matthew correlation coefficient were all achieved very competitive result.Especially in the segmentation of Melanoma and Seborrheic Keratosis,the latest segmentation performance has been achieved.CSAG&DCCNet trained in ISIC-2017 was tested on another publicly available data set,PH2,to verify the robustness and cross-dataset performance of our method.(3)This paper proposes an image erasing method based on the maximum activation point guidance.During training,the maximum pixel value of the maximum activation feature map will be calculated to find the most discernible part of the image.Centering on the point with the highest weight,the occlusion area is generated,forcing the model to learn more image context information when making a decision.In order to evaluate the performance of the method,the proposed method is combined with Res Net,Wide Residual Networks(WRN),and Res Ne Xt networks,and experiments are carried out on the Cifar and Fashion-MNIST data sets,and very competitive experimental results have been obtained.(4)Quantitatively evaluated the performance of the width residual network(WRN-28-10)combined with the maximum activation point guided erasure method in ISIC2017 dermoscopic image classification,and verified that this method can also achieve more advanced dermoscopic image classification performance.