Pathological Analysis Of Breast Cancer And Prediction Of Drug Candidate Activity Based On Combined Model

Breast cancer is the world’s highest incidence of malignant tumors.Early diagnosis and treatment can greatly increase the survival rate of patients and prolong the life of patients.There are various methods for the diagnosis of breast cancer,including mammography,molybdenum examination and nuclear magnetic resonance.However,the gold standard in the diagnosis of breast cancer is histopathologic biopsy based on pathological images.Currently,there are still have three problems in the classification of histopathologic images based on AI-assisted diagnosis:(1)Difficulty in obtaining pathological images and lack of quality;(2)Difficulty in achieving both model accuracy and lightweight;(3)The accuracy of multi-classification is difficult to meet clinical requirements.In conclusion,although the existing researches on automatic recognition and classification of breast cancer have made great progress,there is still a gap between them and their effective application in clinical research.In addition to the early diagnosis,subsequent drug therapy also plays a crucial role in breast cancer survival.The current research and development of anticancer drugs are facing a host of big and difficult challenges,such as: rare reserve of highly talented professionals;insufficient of industrial research and development funds.It has been an urgent and tough problem about how to effectively utilize the advantages of artificial intelligence algorithm to provide powerful auxiliary tools for the research and development of current anticancer drug candidates,and reduce the cost of research and development.In view of the above problems,to further promote the clinical application of automatic diagnosis of breast cancer.Selecting Break His dataset and ERα activity dataset as the research object,this thesis has carried on research to the classification of pathological images and activity prediction of anti-breast cancer candidate drugs.The primary contents are as follows.Firstly,aiming at the complexity and insufficient robustness of the classification model of pathological images of breast cancer,a combined model of machine learning and deep learning was constructed: Xception-SVM,which could improve the prediction accuracy and reduce the training cost,and better balance the two requirements of accuracy and lightweight.(1)Firstly,in order to solve the problem of uneven distribution of image light source and detail loss,this thesis introduced a variational mode to enhance contrast.At the same time,to alleviate the problem of insufficient training set,data expansion and other data preprocessing methods were performed,including random cropping,random rotation,random flipping and image standardization.(2)Secondly,the combined Xception-SVM model was constructed for the classification of breast cancer pathological images.Xception was selected as the basic model,and Sigmoid and Softmax classification layer are replaced by SVM which has hyperplane advantage in machine learning,so as to improve the generalization ability of the model.(3)Finally,transfer learning and fine-tuning techniques were introduced to realize cross-task sharing of weights between models,which can reduce training parameters and reduce training costs.Simultaneously,the model can extract image features to a greater extent and improve classification accuracy.The experimental results show that the constructed model can better balance model accuracy and lightweight,and further promote clinical application.Secondly,in terms of the problem of AI-assisted anti-breast cancer drug candidate research and development,in order to save the time and cost,a combination model for quantitative prediction of ER-α biological activity of compounds was established:GBDT-CFPSO-SVR.The model performance was optimized from various aspects,making it a powerful auxiliary tool in drug research and development.(1)Firstly,there are many dimensions of molecular descriptor information.In order to reduce training costs,three representative feature selection methods were used: GBDT,MIC and RFECV to screen out the top 20 molecular descriptors that have the greatest impact on ERα biological activity.(2)Then,the CFPSO algorithm was used to optimize the parameters of SVR,so as to improve the prediction accuracy;(3)Finally,comparative predictive analysis was carried out to select the final combined model for predicting the activity of anti-breast cancer inhibitors: GBDT-CFPSO-SVR.In conclusion,starting with the research on the breast cancer,this thesis established an automated assisted pathological classification model and an activity prediction model for the research and development of anti-breast cancer candidate drugs.The experimental results show that the model can provide theoretical support for the diagnosis and treatment of breast cancer.