Classification And Semantic Segmentation Of Pathological Images Based On Deep Learning

Medical image processing is a complex and interdisciplinary field that combines medical imaging,mathematics,physics,and computer science.The extraction and processing of medical image features,as well as the analysis and decision-making processes,have a significant impact on the performance of medical image processing models.Traditionally,medical image feature processing has relied on feature engineering,which is a manual process that requires prior knowledge and can be time-consuming.However,current advancements in medical image processing are shifting towards feature learning,which uses deep learning methods,such as artificial neural networks,to analyze images end-to-end.In this thesis,the focus is on the classification of pathology images and small-scale medical image segmentation.Segmentation is used as a more refined classification task to further analyze medical images and provide assistance to healthcare professionals.To improve the accuracy and efficiency of pathology images detection,this thesis proposes a multidimensional convolutional lightweight network based on Shuffle Net V2 called MCLNet.The thesis considers the association between pixels of the same category in an image to be similar to the association between words in the same sentence.To mine this association,the thesis uses one-dimensional convolution to extract the association between image elements,which enriches the extracted information and complements the deficiency of two-dimensional convolution in global feature extraction.To address the problems of small medical datasets and coarse-grained modular integration approaches,this thesis proposes KDTUNet,a knowledge refinement model based on hierarchical fusion.The thesis proposes three innovative strategies to solve these problems: a lightweight network design applicable to small datasets,a knowledge distillation network to improve network performance and interactive feature extraction,and inter-layer fusion of Transformer and UNet.The proposed KDTUNet is compared with the current mainstream and commonly used methods,and the best results are illustrated through experiments.In summary,this thesis focuses on two parts of medical image processing: the classification of gastric cancer pathology images and small-scale medical image segmentation.The proposed algorithms in this thesis have all achieved SOTA advanced levels.The multidimensional fusion approach and knowledge distillation strategy presented in this thesis contribute to the advancement of medical image processing,and have the potential to assist healthcare professionals in the detection and diagnosis of diseases.