Design And Implementation Of Skin Cancer Auxiliary Diagnosis And Treatment System Based On Convolutional Neural Network

Skin cancer is a common cancer.Early diagnosis and treatment can significantly reduce the mortality of skin cancer.Therefore,the automatic detection and recognition of skin cancer is of great significance in clinical practice.According to investigations,it is found that unreasonable encoder-decoder architecture design in medical image segmentation can easily lead to the problems of single feature extraction method and inadequate differentiation of feature information validity,and imbalanced data samples and insufficient feature extraction in medical image classification are important reasons for poor classification performance.To address the above problems,an research on the two major technologies for segmentation and classification of dermoscopy image has been carried out in this thesis.The main works are as follows:（1）A dermoscopy image segmentation model based on attention and Inception mechanism is proposed.In the feature extraction part,the improved model uses the Inception mechanism to extract multi-scale features,which enriches the detail information of lesion features.In the skip connection part,the improved model embeds the attention mechanism module to reallocate the weight proportion at the channel level,which strengthens the relationship between the channel-wise feature maps.The results of comparative experiments show that the attention and Inception mechanism highlight the relation between features and increase the information content of features,and that the performance of the improved model is effectively ameliorated..（2）A dermoscopy image segmentation model based on atrous spatial pyramid pooling structure and decoder with dense connection is proposed.The improved model uses the various dilated convolutions in the atrous spatial pyramid pooling structure to extract a larger range of contextual information,which provides pixel-wise feature information for image restoration.In the conventional design of decoder,the single-path restoration of image resolution leads to continuously lose information of abstract features during layer-by-layer upsampling.The improved model delivers the low-level feature output in the decoder to the deep network through skip connections,which makes up for the loss of information in the upsampling process and promotes long-distance and multilevel feature fusion.The experimental results show that the improved model can accurately segment the lesion region,fit the contours of lesion well,and promote the segmentation effect.（3）A dermoscopy image classification model based on multi-model feature fusion is improved.High-level features are the description of the most essential information of the image and also the basis for image classification.The improved model makes use of multiple models to extract diverse features and enhances the richness of features by feature fusion technology,which provides more useful features for the classifier.The final experiments verify the effectiveness of the improved model.（4）A skin cancer auxiliary diagnosis and treatment system based on convolutional neural network is designed and implemented.The system can automatically segment the lesion region of dermoscopic images and complete lesion category prediction according to the segmentation results.It can provide reliable diagnosis basis and effective lesion information for doctors.