Skin Lesion Segmentation Based On Computer Vision

Lesion segmentation is the main part of medical image segmentation,which determines whether medical images can provide a reliable evidence in clinical diagnosis and treatment.It’s both key and difficult topics for using machine learning algorithms to process image lesion and complete segmentation.On the one hand,accurate segmentation of lesions can provide strong evidence for doctors’ early diagnosis,which is beneficial to treat patients with skin cancer timely and effectively reduce cancer mortality;on the other hand,the segmentation of skin lesion is an arduous task,because imaging technology has a great impact on diagnostic efficiency,and it is also a quite complex diagnosis process,there are some difficulties in skin cancer images,such as small proportion of key information,disturbance of background information,blurred edge of lesion and weak ability of information expression.Therefore,this thesis proposes two kinds of lesion segmentation methods for solving these problems.The main work is as follows:Firstly,an improved Grab Cut algorithm is constructed for skin lesion segmentation.Because the acquired skin images are often disturbed by the imaging equipment,resulting in chaotic background information and blurred lesion edge,in addition,there is noise artifact and unbalanced proportion of key information of the image,which make skin lesion difficulty to be identified and diagnosed.For this reason,this thesis improves the algorithm based on Grab Cut.Firstly,the image is preprocessed to remove the irrelevant noise such as hairs and bubbles,highlight the key information and enhance the edge of the lesion.There are two kinds of lesion masks,one is for clustering pixels with K-means++ algorithms,and enhancing the image with CLAHE method in HSV color space,extracting the lesion mask of green area.The other is to use Otsu method for the value of maximum interclass variance and extract the lesion mask adaptively.Then a threshold strategy is used to input the two masks into the Grab Cut algorithm for modeling and segmenting skin lesion.Finally carrying out morphological post-processing on the segmentation results.The proposed method is tested on skin dataset PH2 and compared the segmentation results with the original method and the widely used traditional graph cutting methods,the segmentation index Jaccard coefficient is increased by 5% on average,which verifies the superiority of the proposed algorithm.Secondly,a fully convolutional network model is proposed for skin lesion segmentation.Because there are disadvantages of high operation times and large memory occupation when the current deep learning method applied in the field of medical image segmentation,the training of the model is very time-consuming,and the inherent structure of the network also causes weak information expression ability of the image,which is difficult to improve the segmentation accuracy.In view of these situations,this thesis proposes a lightweight fully convolutional network model FCU-net.On the basis of U-net network,change the depthwise separable convolution for the standard convolution in the model,so as to realize the lightweight representation and reduce the complexity of parameter calculation.Then,in the down-sampling stage,the pooling layer is replaced by a convolution layer with step size of 2,which expands the receptive field without losing accuracy and enhances the representation ability of advanced semantic information,so as to improve the segmentation accuracy of skin lesion.The proposed lightweight model is reduced by about 1/5 in parameter quantity,and the value of FLOPS is the lowest among the similar models,the model complexity is effectively reduced.Using the public skin dataset PH2 to verify and compare with a variety of deep segmentation networks,the segmentation evaluation coefficients IOU,Dice and Accuracy are ranking first among those algorithms,which proves that the proposed model has strong segmentation performance.In order to further improve the visualization effect,this thesis also carries out an integrated method called Bagging to post-process the lesion segmentation results,the more accurate results are obtained after integrating the model FCU-net and fine-tuned Seg Net.