Research Of Segmentation Methods For Pigmented Skin Lesions Based On Random Forest

The mortality rate of malignant melanoma is ranked first in all skin diseases.The incidence of the disease has been growing rapidly over the past few decades.Since the late skin melanoma is still uncured at present,early detection is an important step in reducing mortality.Dermoscopy is usually used for the diagnosis of melanoma,which can capture the detailed features of the area of the skin lesion.However,the use of dermoscopy for diagnosis is inseparable from the clinical experience of dermatologists and there is a great deal of subjectivity.Therefore,in order to improve the accuracy of melanoma diagnosis,it has become extremely necessary to research a method for detecting melanoma automatically or semi-automatically.Segmentation of pigmented lesions is a key part of automatic diagnosis and is very important for the early diagnosis and treatment of melanoma.The main research content of this paper is the segmentation of pigmented skin lesions based on random forest algorithm.Around this main content,the following work has mainly been done:(1)Segmentation of pigmented skin lesions based on random forest and wavelet textures: The dataset is divided into the train set and test set two parts.At first,the statistical region fusion algorithm is used to perform rough segmentation after preprocessing for the dermoscopy images.Then the color features are extracted in different regions,the texture features are extracted using Gabor wavelet.Finally,all of features extracted form feature vectors were inputted into random forests to train random forests.The test set is used to test the trained random forest modelIn addition,in order to further test the performance of the model,a 5-fold cross validation was used to estimate the model's generalization error.(2)Pigmented skin lesion segmentation based on random forest and full convolutional network(FCN): First,a random forest algorithm is used to segment the data set to obtain a segmentation result;then,a full convolutional network is used to segment the data set to obtain another segmentation result;Finally,the average value of two results per pixel whichever fused to get the final segmentation result.In addition,the image is preprocessed and the segmentation results are post-processed: some hair,black borders,etc.in the image are removed.After segmentation,there are uneven edges,holes and isolated islands in segmented images.Post-processing operations are required,including hole filling,island removal and edge smoothing.Finally,performance indexes such as sensitivity,specificity and accuracy are calculated to evaluate the performance of the model.The experimental results show that the combination of random forest and wavelet-based texture can accurately segment the lesion area,and the sensitivity is above 90%,the specificity and accuracy is as high as 95%.The combination of random forests and FCN can increase the individual indicators compared with using random forest segmentation alone,and in particular can increase the sensitivity by about 20%.