Research On Diagnostic Method Of Skin Cancer Based On Semi-Supervised Learning

The mortality rate of skin cancer is among the top five in the world,while it can be significantly reduced by the early diagnosis and timely treatment.Therefore,to explore how to improve the accuracy,objectivity and efficiency of skin cancer screening is of great value in critical practice.According to the investigations,most of the skin cancer diagnostic methods are based on supervised learning and demand huge amounts of available labeled data.Its diagnostic precision is also limited by imbalanced data distributions and inadequate feature differentiation.Therefore,this thesis proposes a semi-supervised framework for medical image classification to address the above issues.Comprehensive experiments are conducted on four publicly accessible dataset,i.e.,PADUEFC-20,ISIC 2018,ISIC 2019 and ISIC-Archive.The main innovations and contributions are summarized as follows:(1)A newly model training paradigm,curriculum learning(CL),is proposed to be integrated into semi-supervised learning framework,training from the easier samples to the more advanced ones.The curriculum learning draws on the experience of meaningful learning order in human curricula which is from easy to difficult.This thesis proposes two ideas to implement that,i.e.,the clustering algorithm based on the distribution density and the sample confidence value.The former splits the whole training set into several subsets(easy-normal-hard)based on the complexity ranking of data,and the later arranges the training order based on the confidence level.Simulations show that training the semi-supervised learning model based on the curriculum learning can improve its unreliable performance during the initial phase,accelerate the training process and enhance the robustness of the model.(2)This thesis discusses and proposes several multi-modal information fusion methods to fuse pathological image features and clinical features.Different types of skin cancer often accompanied by different clinical manifestations.For example,pigmented skin diseases hardly occur wound ulceration,bleeding and pain,and the distribution of incidence rate varies significantly at different ages.Simulations show that using additional clinical information can tackle the issue of inadequate differentiation of image features,and thus to achieve an improvement for skin cancer diagnosis.(3)This thesis proposes a novel feature enhancement module to explore and fuse multi-scale features.The deep learning models explore more local feature such as texture and contour at the shallow layers,and rich semantic information at the deep layers.This module is applied to fuse both features to address the issues of inter-class similarity and intra-class variation,focusing more on the fine-grained information of deep layers.Simulations show that it can further enhance the feature fitting and achieve the performance improvement of the model.(4)The supervision mode of semi-supervised learning is improved.Traditional semisupervised learning almost only considers the consistency of the single sample.This thesis further proposes the consistency paradigm of sample relationship,and thus to train the model under the joint supervision of individual consistency and sample relationship consistency,to explore richer information from unlabeled data.