Pathological Image Automatic Labeling Based On Biological Tissue Polarization Characteristics And Machine Learning

Pathomorphological examination is a routine method of cancer diagnosis.In this method,the suspected cancerous tissue is made into stained sections then is examined by professional pathologists with microscope.The detection process is usually time-consuming,laborious,kind of subjective,and lack of automatic annotation for the cancerous area of pathological images.With the increasing demand of the clinic,the pressure of pathologists is also increasing.Polarization characteristics of biological tissues are ignored in the traditional pathological examination,the polarization information differences of different tissues directly reflect the differences of their microstructure,it can provide more comprehensive and organized optical information for the detection and diagnosis of cancer and other diseases.This paper combines tissue polarization characteristics and deep learning method,utilizes tissue polarization imaging system to obtain polarization information images of pathological sections.Meanwhile,a coding-decoding deep convolutional neural network is designed,which learns the polarization differences between cancerous and non-cancerous tissues in Mueller matrix images by the supervised learning method to achieve automatic labeling of pathological images.The experimental results show that compared with intensity images,the contrast of cancerous and non-cancerous tissue regions in polarization images is significantly improved;Compared with the traditional segmentation method,our method can complete the segmentation and recognition of cancerous tissue at the same time,the accuracy of the average segmentation and recognition rate for liver cancer and thyroid cancer has reached 78%,this method provides a new solution for cancer detection in biological tissues.