Research Of Skin Lesion Segmentation Method Based On Collaborative Learning And Attention Mechanism

The incidence rate of skin diseases worldwide is increasing yearly,significantly threatening human health and life.If skin diseases(such as melanoma)can be diagnosed and treated as quickly as we can,the survival rate of patients can be improved.However,due to limitations in timeliness and convenience,manual diagnosis cannot meet the needs of patients.Dermoscopy can assist doctors in treatment by clearly presenting lesion characteristics.During the diagnosis and treatment of diseases,whether it is disease analysis or cutting surgery,accurate segmentation of the skin lesion areas is required.Accurate segmentation of skin lesions using computer-aided diagnostic methods is essential for assisting doctors in diagnosis and treatment.Computer-aided image segmentation methods for skin lesions are divided into traditional and deep learning-based skin lesion segmentation methods.Traditional image segmentation methods rely on artificially designed features,which are inefficient and less robust.Deep learning-based image segmentation methods have achieved automatic feature extraction and significant progress.However,in the case of the small number of labeled samples and the complexity characteristic of skin lesion images,achieving accurate segmentation of lesion regions remains challenging.This thesis deeply studies the skin lesions segmentation method based on collaborative learning and attention mechanism for the above situation.The main research contents are summarized as follows:Firstly,for the problem of small labeled samples in skin lesions images,a segmentation and classification model for skin lesions based on collaborative learning is proposed.In order to improve segmentation accuracy under limited labeled samples,this thesis uses unlabeled data to generate high-quality pseudo labels to expand training samples and uses image-level labeled data to generate class activation maps to provide location information for the segmentation network.In order to alleviate the interference of noise in dermatoscopic images on disease diagnosis,this thesis uses lesion masks generated by the student segmentation network to provide shape prior to the classification network,thereby improving the discrimination ability of the classification network.Then,for the problem of low segmentation accuracy caused by the complexity of skin lesion images,a skin lesion segmentation model based on the attention mechanism is proposed.The model learns global and local information about skin lesion images by integrating multiple attention mechanisms.In order to fully extract the feature information of skin lesions,this thesis employs Transformer to establish global context relationships to alleviate inductive biases caused by limited receptive fields.In order to make the model adaptively focus on essential features,this thesis integrates spatial channel attention into the model,promoting the model to learn task-related features in both spatial and channel dimensions.In order to make the model focus on the fuzzy boundaries of skin lesions,this thesis integrates residual attention into the model,utilizing boundary prior knowledge to capture more local detail information.Finally,many ablation and comparative experiments were conducted on the proposed method.In order to verify the effectiveness of the skin lesions image segmentation and classification model based on collaborative learning,this thesis visualizes the segmentation results generated at each stage and uses ablation experiments to demonstrate the effectiveness of reliable pseudo-labels and class activation maps for improving segmentation accuracy,demonstrate the effectiveness of masks for improving classification performance.In order to verify the effectiveness of the skin lesion segmentation model based on the attention mechanism,this thesis visually compares the segmentation results generated by this model with other methods and proves the effectiveness of the combination of various attention modules through ablation experiments.