Histopathological Image Analysis Of Adenocarcinoma Based On Deep Learning

Adenocarcinoma is an import part of carcinomas.Histopathological diagnosis is the"gold standard" of tumor diagnosis in modern medicine.In recent years,with the rapid development of deep learning and whole slide image scanning technology,automatic histopathological diagnosis becomes possible.This paper explores the application of deep learning in histopathological image analysis of adenocarcinoma in three aspects,mitosis detection of breast adenocarcinoma,BRAFV600E prediction of papillary thyroid carcinoma,and ISUP grade of prostate adenocarcinoma.The detection and counting of mitotic cells are important indicators in the Nottingham grading system of breast adenocarcinoma.In this work,this paper proposes a method,which automatically generates pixel-level labels of mitosis detection dataset and then uses them to train a ResNet-based FCN for mitosis detection.The proposed method surpasses the existing best method in F1-score on the three most commonly used mitosis detection datasets,AMIDA 2013,ICPR 2014,and TUPAC 2016.BRAFV600E mutation has been demonstrated to make papillary thyroid carcinoma more dangerous in many works.Its prediction is of great significance for the diagnosis and prognosis of papillary thyroid carcinoma.In the second work,this paper proposes a deep learning workflow,which includes two stages of tumor detection and mutation prediction.The proposed method reaches 0.884 of AUC on the papillary thyroid carcinoma dataset provided by Zhongshan Hospital,Xiamen University,showing that the BRAFV600E mutation can be directly predictable through histopathological images.The ISUP grading system is a new grading system for prostate adenocarcinoma proposed by the International Society of Urological Pathology,further developing the current commonly used Gleason grading system.In this work,this paper stitches the most informative patches as model input and proposes a multi-bit label for training the EfficientNet-B2 model for ISUP grade.The proposed method achieves 0.935 of QWK on the public dataset PANDA,showing its excellent performance.In the above three works,this paper has proposed new deep learning algorithms and carried out experimental verification on the corresponding datasets.All the methods achieve the state-of-art performance.The work of this paper promotes the development of automatic histopathological image analysis technology for adenocarcinoma based on deep learning,and also the advancement of tumor diagnosis and treatment technology.