Research On Biomedical Image Segmentation And Classification Based On Convolutional Neural Network

Biomedical images can show the presence of lesions or abnormalities in different organs or cells,so it is an important way for medical and biological researchers to diagnose diseases and understand biological processes.However,currently,the annotation of biomedical image data is still done by experts.Due to the complexity of images and the subjectivity of experts,there are some differences between the annotations made by different experts.The sheer volume of data also poses challenges for manual annotation.Thus these difficulties are giving rise to the need to automatically and accurately identify and interpret biomedical images.In particular,the rapid development of deep learning techniques in recent years has greatly improved the efficiency and accuracy of biomedical image processing.This thesis focuses on two tasks in the field of biomedical image processing,retinal vessel segmentation and protein subcellular localization.For retinal vessel segmentation,changes in the blood vessels of the fundus can reflect the occurrence of ocular diseases,this allows us to explore other physiological diseases that cause fundus lesions.However,existing computational methods lack efficient and accurate segmentation of thin vessels,so it is important to construct a reliable and quantitative automatic segmentation method to improve diagnostic efficiency.Therefore,this work proposes a multi-channel deep neural network for retinal vessel segmentation.First,U-net is applied to the original vessels and thin(and thick)vessels to perform multi-objective optimization.Then,the three prediction probability maps are fused to form the final binary segmentation map using the special fusion mechanism designed in this thesis.For protein subcellular localization,by detecting abnormalities in the subcellular location of a protein,the occurrence of certain diseases can be inferred and new drug targets can be explored.However,among the existing image-based protein subcellular localization methods,traditional techniques lack efficiency and accuracy,and the potential of deep learning methods has not been fully exploited.Therefore,in this thesis,a multi-scale multi-model deep neural network via ensemble strategy is proposed.First,a deep convolutional neural network is used to extract multi-scale features,and the global average pooling layers are used to map the features extracted at different stages into feature vectors,then multi-scale feature vectors are concatenated together in different ways to form a multi-model structure for image classification,finally the multi-model classification results are integrated using the ensemble strategy to obtain the final results.In addition,the method incorporates a squeeze-and-excitation module in the network to emphasize more informative features.In conclusion,this thesis proposes two novel and efficient automatic segmentation and classification methods based on biomedical images.For image segmentation,the method in this thesis can effectively improve the segmentation quality of difficult-tosegment parts of the image,thus improving the overall segmentation results,which can be applied to the automatic analysis of biomedical images;for image classification,the method proposed in this thesis provides a framework for image processing and classification,which can significantly improve the accuracy of image classification and lay the foundation for the research of other biological processes.