Melanoma Detection Techniques In Dermoscopy Images Based On Deep Learning

Malignant melanoma is a highly lethal form of skin cancer with the characteristics of poor prognosis and rapid metastasis,and its incidence is increasing yearly.Early diagnosis and clinical intervention are the most effective ways to reduce mortality.However,the number of dermatologists with melanoma detection ability is difficult to match the rate of increase in incidence.It is subjective to diagnose the skin tumor by naked eyes and the reproducibility of diagnostic results is not ideal.Automated melanoma detection in dermoscopy images is now an effective solution for these problems,but the complexities of dermoscopy images,such as the irregular or fuzzy lesion borders,the intraclass variations and the similarity between classes,present great technical challenges to the melanoma detection.In recent years,deep learning represented by the convolutional neural network(CNN)can automatically learn hierarchical features from the data without complicated manual extraction,and it shows advanced effects in the field of natural image and medical image analysis.Therefore,this thesis carried out systematic analysis and research to the application of deep learning in the melanoma detection techniques of dermoscopy images.The main research work and achievements of this thesis are as follows:1.This thesis proposes a melanoma lesion segmentation model based on FCN called mUnet and optimizes it with a loss function based on dice coefficient.Accordingly,it proposes a set of effective training schemes.The jaccard index of the optimized m-Unet model is 7.1% more efficient than the classic FCN model based on VGG16 and provides powerful supports for subsequent image recognition.2.This thesis applies deep residual network(DRN)to melanoma recognition and proposes a set of training strategies based on transfer learning.In the melanoma recognition task,DRN can achieve better results than shallower networks,and the training strategies can ensure effective training under limited training data and take full advantage of pre-trained knowledge as well as the properties of melanoma itself.3.In this thesis a color constancy algorithm called Shades of Gray is used to solve the problem of multi-source images.And then,this work seamlessly integrates the lesion segmentation,the color constancy and the DRN model to form a three-stage framework.This framework can achieve a 3.4% auc improvement in the recognition task compared to the baseline of a common single model.At last,the thesis proposes a model ensemble framework based on the weighted average of auc,which increases by 7.7% relative to baseline on auc criteria,and it has some advantages compared to the current mainstream algorithm models.At last,the methods proposed by this thesis are extensively evaluated on ISBI2017 dataset with multiple evaluation criteria.Experimental results fully verify the validity of the proposed methods and models.Therefore,the results of this thesis have a good reference value for the further study of melanoma detection in dermoscopy images.