Design And Implementation Of Pathological Image Annotation And Auxiliary Diagnosis System Based On Weak Supervised Nuclear Segmentation

With the progress of computer intelligent science and technology and medical imaging science and technology,the digital pathology technology is booming,and the acquisition of pathological sections is more convenient,but also changes the traditional way of reading films.In recent years,the computing performance of GPU has been significantly improved,and deep learning technology has also developed rapidly.This also makes segmentation algorithms combined with deep learning technology gradually become possible and have been increasingly applied in practical scenarios such as medical diagnosis,mobile robots and unmanned driving.Deep learning is used to process full-field digital slices to obtain target regions of interest.Since a large proportion of lesions occur at the cellular level,nuclear recognition is often essential in many cases.Nuclear segmentation is a basic task in histopathological image analysis.This segmentation task requires considerable effort to manually generate accurate pixel-level annotations for fully supervised training.In order to reduce such tedious manual operations,this thesis based on the weakly supervised segmentation framework of local point labeling,that is,only a small part of the nuclear positions in each image are marked,trained the model,and conducted research on domain adaptation to enhance the segmentation ability in different nuclear scenarios.Finally,a script was developed based on the existing open source annotation tool ASAP,in which the trained weakly supervised nuclear segmentation model was called,and the function was expanded according to the actual needs of traditional annotators,so as to design and develop pathological image automatic annotation software.