Research On Intelligent Diagnosis Report Of Digital Pathological Image

With the development of the field of artificial intelligence,the medical pathological diagnosis has also made great progress.In pathological diagnosis,the pathological development of malignant tumors can be mainly divided into two stages,the stage of carcinoma in situ(non-invasive type),and the stage of invasive type of cancer formed after breakthrough of the basement membrane.However,in most research work,diagnosis is performed for a single stage(such as cancer in situ)of the development of a tumor(such as gastric cancer),and only the benign and malignant diagnostic labels are given on the diagnosis conclusion.The above-mentioned problems will lead to the incompleteness of the pathologically assisted diagnosis,and fail to give a certain interpretability of the diagnosis,which will cause certain understanding problems for pathologists.In order to solve these problems,in this work,we have mainly carried out research on digital pathology-assisted diagnostic report generation technology for all stages of the development of malignant tumors,including images and text.(1)Because there is a clear cancerous boundary in the non-invasive cancerous stage,the diagnosis problem can be solved by means of semantic segmentation.Aiming at the problem that pathologists need to label data,we propose to use a semi-automatic segmentation labeling tool based on the principle of random walks,which can help doctors to obtain fine segmentation boundaries through simple line sketching,and lay solid data for future research basis.(2)To solve the problem that the conventional segmentation method does not perform well in the non-invasive cancerous stage,this paper proposes a network model for end-to-end segmentation of the whole pathological images.We innovatively explore segmentation modules that adapt to the irregular shape characteristics of pathological cancerous regions,and can integrate different shaped receptive fields for semantic segmentation and generate relevant diagnostic reports.We compare with the mainstream segmentation algorithms on MICCAI 2019 dataset,and have achieved very good performance.(3)Aiming at the characteristic that there is no clear cancerous boundary in the stage of invasive cancer,we cooperated with Shanghai International Peace Maternity and Child Health Hospital to build a new dataset,and innovatively proposed a combination of attention mechanism and multi-label classification method to solve invasive cancer.we achieve attention areas corresponding to different pathological attributes and generate structured diagnostic reports.Finally,through the LSTM diagnostic module,better-quality diagnostic conclusions can be generated.