Research On Deep Learning-Based Intelligent Segmentation And Recognition For Skin Lesion

Skin lesion is one of the most common diseases,and if malignant skin lesions such as melanoma are diagnosed early,it can significantly prolong the lifespan of patients.However,due to the diverse manifestations and complex symptoms of skin lesions,traditional diagnostic methods have problems such as high misdiagnosis rates and low efficiency,which cannot meet the growing diagnostic needs.With the development of artificial intelligence technology,deep learning-based intelligent segmentation and recognition schemes for skin lesion have made significant progress.However,current schemes still face some challenges,such as small inter-class variations and large intra-class variations in skin lesions,imbalanced distribution of lesion types,and patient privacy leakage.To solve the above problems,this thesis studies the deep learning-based intelligent segmentation and recognition scheme for skin lesion,and the main contributions include the following three aspects:1)Residual inception and bidirectional Convolutional Gated Recurrent Unit(Conv GRU)empowered intelligent segmentation for skin lesion: Due to the shape,color,and texture differences of skin lesions,and the unclear boundary,it is difficult for traditional deep learning methods to accurately segment them.Therefore,this thesis proposes a residual inception and bidirectional Conv GRU empowered intelligent segmentation model for skin lesion.Specifically,a cloud-edge collaboration intelligent segmentation service network model for skin lesion is designed.By this network model,users can obtain quick and accurate segmentation services.Furthermore,a novel intelligent segmentation model for skin lesion is developed.By integrating residual inception and bidirectional Conv GRU,this model can fuse multi-scale features and make full use of the relationship between low-level features and semantic features.It improves the ability of the model to extract features and capture global context information,and leads to better segmentation performance.Finally,the experimental results indicate that the proposed intelligent segmentation model performs excellently.Compared with several recently proposed U-Net extension models,its accuracy and Jaccard coefficient are higher,reaching 0.934 and 0.831 respectively.2)Transfer learning-based intelligent recognition for skin lesion: Limited training samples,imbalance in disease class distribution,and similarities between different skin lesions make it difficult to accurately classify skin lesions.Therefore,this thesis proposes a transfer learning-based intelligent recognition for skin lesion.First,a cloud-edge computing model is constructed.Then,by combining transfer learning,model fine-tuning,data augmentation,and model ensemble technology,an intelligent recognition mechanism for skin lesion is proposed,which can solve the problem of low classification accuracy caused by insufficient training samples.In addition,a class-balanced crossentropy loss function is designed to improve the classification accuracy when the class distribution of the training dataset is imbalanced,and further improves the classification accuracy of melanoma by introducing a weighting factor.The experimental results show that the proposed intelligent recognition mechanism for skin lesion achieves better classification performance than other existing mechanisms,especially for melanoma classification,which shows a significant improvement.The balanced accuracy and melanoma sensitivity reach 0.873 and 0.837,respectively.3)Federated semi-supervised learning-based intelligent recognition for skin lesion: To address problems such as limited data in a single medical institution,insufficient labeled samples,and privacy leakage in centralized learning,a federated semi-supervised learning-based intelligent recognition mechanism skin lesion is proposed.Specifically,a federated learning-based cloud-edge collaborative intelligent recognition model for skin lesion is designed,which collaboratively trains data from various medical institutions while protecting the privacy of users.This model can provide users with accurate and convenient diagnostic services.Then,a semi-supervised loss function for heterogeneous data is designed to effectively control the difference between local models and global models.In addition,by combining multiple random sampling and accuracy-based weighting methods,the contribution of each local model is clarified,and all uneven local models are aggregated into a global consensus model to further reduce the impact of data heterogeneity.Finally,experimental results show that the proposed mechanism has better performance and scalability than several recently proposed mechanisms.