Recognition Analysis And Application Research Of Digital Pathology Images Based On Deep Learning

Pathology is a basic medical discipline that studies the etiology,patho-genesis,pathological changes,outcomes,and mutation outcomes of diseases.Through the study of pathological mechanisms,we can understand the mor-phological structure and functional changes of different organisms at differ-ent stages.Pathology has been regarded as an axial discipline between basic medicine and clinical medicine,which can provide more accurate and available information for the prevention,diagnosis and treatment of diseases.Digital pathology refers to the integration of computer and communication network-related technologies into the field of pathology.It is a technology that organ-ically combines modern digital systems with traditional optical magnification devices.The higher resolution of digital pathology,the appropriate observa-tion and analysis of stable imaging,the easy storage and management,and the convenient browsing and transmission provide high convenience for remote consultation and pathology teaching.Drawing on the cutting-edge experience and advanced results of deep learning in computer vision,we aim to apply deep learning techniques to the field of digital pathological image recognition and analysis,while referring to some ideas of transfer learning strategies in deep neural network related algorithms.It is intended to solve the problems of time-consuming and labor-intensive,inaccurate diagnosis,and inconsistent diagnostic indicators in the current digital pathological diagnosis and analysis research.This thesis promotes the accuracy and intelligence of pathological diagno-sis and pathological analysis through the integration of computer intelligence science,medical imaging,and digital pathology.In this study,by analyzing the specific problems of digital pathology,we introduce and improve the algorithms of pattern classification,semantic segmentation,and object detection based on deep neural network algorithms to solve the related problems of digital pathol-ogy in different sub-fields.For example,cell detection and segmentation at the cytological observation level,tumor region segmentation at the histological observation level,and classification of protein-negative positive image blocks at the histochemical observation level based on immunohistochemical staining for image analysis.This thesis conducts specific research on these sub-problems just raised.The main work and results include:1)A multi-scale fusion and deformable convolution-based segmentation model of gastric cancer based on deep learning is proposed.In these works,we first established a clinical-based gastric cancer pathological image dataset with our partner hospital.The dataset was fine-labeled by a number of medical ex-perts.Secondly,we proposed multi-scale embedded networks for different sizes.We integrated an encoder based on cascade dilated convolution and a spatial pyramid pooling module,thereby solving the problem of image segmentations under reduced size.At the same time,we embed the deformable convolution in the middle layer of the encoder to use it to perceive richer pathological region features,especially the "fluid" and "non-rigid" features of the gastric cancer region.Finally,in order to compensate for the "gridding" problem caused by cascading dilated convolutions,we used Dense Upsampling Convolution(DUC)modules at the end of our entire architecture,which are mainly used to refine fuzzy and gridded boundaries.After several sets of experiments,we compared our method with the natural image segmentation algorithms that were prominent in the field of computer vision semantic segmentation.We also com-pared the impact of different modules on the accuracy of the model under the influence of different hyper-parameters.Through our model,we can segment the pathological region of gastric cancer more accurately,which can provide better,faster and more reliable auxiliary information for clinical diagnosis than the compared models.After a large amount of research on published literature,this sub-research belongs to the first application of deep learning models in the field of pathological images of gastric cancer,which has certain promotion significance for simplifying medical procedures,improving medical efficiency,and achieving precision medicine.2),A cell/nucleus detection and segmentation application framework based on deep learning instance segmentation model is proposed.In this work,we segmented the nucleus detection and counting task instance based on patho-logical images to separate each cell separately,which can provide more detailed quantitative information about cells for the clinic.Methodologically,we applied a Mask R-CNN architecture based on multi-task integration to simultaneously solve cell detection and cell segmentation tasks.We applied the Faster R-CNN and DeepLab series of algorithms as the backbone networks with the most sta-ble detection,segmentation performance on different task branches.In terms of network structure,we modified a multi-scale structured Region Proposal Network(RPN),which uses the Feature Pyramid Network(FPN)to detect tiny cell nuclei.In particular,we used the "anchor" mechanism to adapt the model to a specific data set,to detect the nucleus like "round".And for the detection of fuzzy nucleus,we changed the Non-Maximum Suppression(NMS)threshold to increase the effect.In the training strategy,the traditional method first divides all the cell nuclei and then separates each nucleus by post-processing.Unlike the traditional method,we performed end-to-end training on the network.Through experimental comparisons on public data sets,our method achieves satisfactory results.At the same time,we verified the effectiveness of our model by experi-mentally comparing different modules,different hyper-parameter settings,and different structural thresholds.3)An intelligent and efficient model of Positive-Negative classification of BAP-1 gene protein expression on uveal melanoma pathological images based on a deep learning classification framework was proposed.Research is more bi-ased towards qualitative analysis in the medical field,which is mainly applied to the transferability and robustness of deep learning algorithms.In the study,we first established an ophthalmological uveal melanoma tumor suppressor protein expression pathology data set with a partner hospital.In essence,we designed a densely connected deep neural network and identified the expression of BAP1 in the nucleus.Through a large number of comparative experiments and clinical verifications,our deep network algorithm performs well and can prove effective with ophthalmologists.While accurately and effectively detecting protein ex-pression,our algorithm also provides more reliable information for subsequent clinical diagnosis and prognosis.Through the study of a large number of litera-tures,this study is the first use of a tightly connected deep classification network to identify the nuclear BAP-1 protein expression in tissues of uveal melanoma immunostained tissuesIn summary,our research shows that artificial intelligence methods are promising in digital pathology image analysis.Through the integration of computer intelligence science and digital pathology,it can promote pathological diagnosis and pathological analysis,and promote the intelligent development of clinical medicine,which can save medical workers more repetitive working hours and encourage them put more energy into more meaningful work.Our research has certain promotion significance for the development and integration of precision medicine and smart medicine.