Deep Multi-task Learning Based On Multiparametric Magnetic Resonance Image Fusion Method For Joint Prediction Of Histological Information Of Breast Cancer

Breast cancer has high morbidity and mortality.It is the most common cancer in women.At present,the diagnosis and treatment of breast cancer has been gradually transformed into a targeted and individualized treatment plan.The Ki-67 expression level,molecular classification and histological grade in breast cancer histological information can provide prognostic information for patients.These indicators are helpful for the formulation of treatment plans and predict patient prognosis information.It is an important reference for individualized treatment of breast cancer.Medical imaging examination is a necessary method before the diagnosis and treatment of breast cancer.Multiparameter magnetic resonance imaging can comprehensively analyze the images of different parameters.Multi-task learning can share some parameters for related tasks to achieve the purpose of mutual learning,thus improving the performance of single task learning.Most of the traditional breast cancer research methods use machine learning to study single parameter magnetic resonance imaging,and most of the studies only predict one or two histological information separately.In this study,we used multiparameter magnetic resonance imaging to predict Ki-67 expression,molecular typing and histological grading of breast cancer.The specific research contents are as follows:(1)Collect,process and analyze patient histological information: Analyze 202 preoperative and pre-chemotherapy breast MRI data of patients with invasive ductal carcinoma,sort out the pathological information data of breast cancer patients,and count the age of the patients Basic information about menopause,Ki-67 expression,molecular typing and histological grading.Use chi-square test and analysis of variance to perform statistical analysis on patient histological information.Obtain the original multiparameter images and preprocess them,use artificial methods to segment the tumor-containing gland images,and expand the processed image data.(2)Research on breast cancer histological information prediction based on singletask deep learning: In the establishment of breast cancer histological information prediction model,the model weight parameters pre-trained on the large data set Image Net are used as transfer learning,and its classification layer is redesigned.And use the fine-tuning method to retrain the thawing layer and the classification layer to predict the molecular classification,histological classification and Ki-67 expression of breast cancer respectively.In single-task learning research,the three tasks of predicting Ki-67 expression,molecular typing,and histological grading are carried out independently,and there is no connection between them.(3)Research on breast cancer histological information prediction based on multitask deep learning: In order to improve the performance of the single-task learning histological information prediction model,we use a multi-task learning method based on partial parameter weight sharing to simultaneously learn significantly related prediction tasks.Establish a deep multi-task learning breast cancer pathological information prediction model,predict the expression of Ki-67,Luminal A type and histological grade of breast cancer at the same time,and evaluate the predictive ability of the model.The single-task learning method and the multi-task learning method are compared and analyzed,and the best learning method is evaluated.(4)Research on breast cancer histological information prediction based on multitask and multiparameter image fusion: In order to make full use of image information,this article mainly proposes three image fusion methods to predict breast cancer pathological information,which are based on the average fusion of decision-making layers.Algorithm,feature layer fusion algorithm based on feature series and feature layer fusion algorithm based on feature parallel.The above three image fusion methods were used to study the histological information prediction of multiparameter images.Finally,all histological information prediction models are evaluated and the experimental results are analyzed.The experimental results show that the best AUC of deep multi-task learning combined with multiparameter image model to predict Ki-67 expression,Luminal A and histological grade is 0.819,0.799,0.747.Deep multi-task learning methods combined with multiparameter images can significantly improve the predictive ability of Ki-67 expression,Luminal A type and histological grade,which is important for breast cancer diagnosis and the selection of personalized treatment options significance.