Research On Digital Pathological Image Enhancement And Accurate Labeling

Histopathological analysis of stained specimens is very important for cancer diagnosis,treatment and clinical prognosis.However,the current pathologic diagnosis is faced with the lack of professional pathologists and the diagnosis is susceptible to subjective factors.Artificial intelligence（AI）can effectively solve the problems existing in the current pathological diagnosis.The combination of AI and digital pathology has been a trend in this field.It has become a new target to use AI to automatically identify and quantify tumor regions in pathological images,guiding pathologists to make reliable diagnosis.Currently,supervised learning is one of the most effective solutions while AI analysis tasks for pathological images usually require pixel-level annotation.All data need to be manually annotated by professionals,which leads to some problems,such as laborious,tedious and error-prone.Since AI based pathological diagnosis has been limited by annotated data sources,this paper studies the method of automatically generating high-quality and accurately labeled pathological image datasets based on hyperspectral imaging technology.In terms of image quality,low-quality images will not only cause misdiagnosis by pathologists,but also affect the recognition accuracy of various deep learning models.It is necessary to provide high-quality images as much as possible while providing annotations.This paper proposes Deep RFT（Deep Residual Fourier Transform）network for image deblurring.Experimental results show that Deep RFT performs better than the existing networks on both public datasets and pathological dataset collected by the self-developed pathological imaging system.It provides strong algorithmic support for the establishment of high-quality dataset and the diagnosis of low-quality pathological images.In terms of label generating,this paper obtains the automatic and accurate annotations of tumor samples from different tissues through utilizing the labeling ability of cytokeratin（CAM5.2）for tumor regions and the difference between CAM5.2stained area and hematoxylin-eosin（H＆E）stained area in spectral dimension.Deep learning is used to convert CAM5.2 and H＆E double-stained tumor images into standard H＆E virtual stained images,which are more common and economical at present.In terms of automatic labeling for tumor regions,this paper proposes to leverage gradient boosting decision tree（GBDT）and Grab Cut algorithm to achieve automatic and accurate segmentation annotations of tumor samples based on both spectral and spatial information of CAM5.2 and H＆E double-stained microscopic hyperspectral tumor images.Simultaneously,the CAM5.2 and H＆E double-stained microscopic hyperspectral pseudo-color images are converted to standard H＆E virtual stained images by Cycle GAN.Finally,the establishment of high-quality H＆E virtual stained pathological image dataset with fine annotation is realized.This paper provides a new idea for the establishment of pathological image datasets with automatic and accurate annotations,which is expected to solve the problem of lacking high-quality datasets for AI pathological image analysis.