Multiple Clinical Indicators Prediction And Survival Analysis Of Breast Cancer Based On DCE-MRI Radiomics Multi-task Learning

Breast cancer has a high recurrence rate and high mortality rate.The prognosis of patients is complicated affected by many factors,and the individual efficacy varies greatly,so it is difficult to predict the medium and long-term survival status accurately.Neoadjuvant Chemotherapy(NACT)is a common treatment for advanced breast cancer.A variety of related clinical indicators such as pCR(Pathological Complete Response,pCR)and TNM Staging are closely related to the prognosis of patients,and these factors together affect the long-term survival status of patients.Dynamic contrast enhancement MRI(DCE-MRI)can provide rich tumor information and is widely used in the diagnosis of breast cancer.If the breast imaging before NACT can be used to predict the relevant indicators after NACT,the survival status of patients can be assessed in advance and precise treatment can be assisted.In this study,DCE-MRI images before NACT were used to obtain radiomics features.According to the close correlation between clinical multiple indicators,a model based on multi-task learning was established to predict these indicators,and survival analysis was carried out to explore the impact of various factors on the survival of patients,aiming to assist the accurate diagnosis and prognosis evaluation of breast cancer.The main research contents of this paper are as follows:(1)Prediction of recurrence and pCR based on multi-task learning of radiomics.Firstly,the preprocessed DCE-MRI images were used to obtain radiomics features,and a single-task prediction model was established for recurrence and pCR respectively.Because of the close relationship between recurrence and pCR,multi-task learning method was introduced,and multi-task feature selection method was used to establish multi-task prediction models of recurrence and pCR.The evaluation index was calculated,and the effects of single-task and multi-task models were compared.The results showed that by sharing the common information of related tasks,the multi-task model achieved better prediction performance than the single-task model in two different tasks,with AUC of 0.805 and0.782,respectively.(2)Multi-task prediction of TNM staging based on radiomics.TNM staging is a tumor staging system to judge the progress of tumor development and evaluate prognosis.At present,there are few studies on prognostic TNM staging.In this study,the trend of TNM stage changes in longitudinal time was analyzed,and the changes in TNM stage before and after NACT were coded to obtain theΔTNM indicator.The results showed that ΔTNM and recurrence were significantly correlated,pCR as well,which could effectively reflect the prognosis of patients.The multi-task learning method was used to construct the multi-task prediction models of ΔTNM and recurrence,ΔTNM and pCR,respectively.Finally,the single-task prediction model of ΔTNM was supplemented to compare the prediction performance of different models.The best prediction result of ΔTNM was obtained in the multi-task model,with an AUC of 0.805.(3)Breast cancer survival analysis with multiple clinical indicators based on medical imaging.The generalization performance of the radiomics multi-task model was verified in the independent dataset,which proved the good generalization performance of the multi-task model.Then survival analysis was carried out by using predictive indicators: the survival analysis of single clinical indicators and multiple clinical indicators in pairs was performed by Kaplan-Meier(K-M).A multivariate Cox proportional hazard regression model with multiple clinical indicators combined with clinicopathological information and imaging features was established.Imaging features were studied in recurrence and different indicators groups to increase the interpretability of the impact of imaging features on survival.The experimental results show that ΔTNM and other clinical indicators have a significant impact on survival,and the combined analysis of multiple clinical indicators can predict the prognosis of patients more accurately.The results also showed that some imaging features had a significant impact on survival,and important imaging features could provide more valuable auxiliary information for survival prediction.In this paper,based on DCE-MRI radiomics multi-task learning prediction model,the prognosis and survival of breast cancer patients were studied.The results show that multiple clinical indicators of prognosis can be effectively predicted by breast imaging before NACT.The combination of ΔTNM proposed in this study and multiple indicators can predict the survival of patients more accurately,and the imaging features can also be used as a marker for auxiliary prognostic evaluation to provide more information for the survival evaluation of breast cancer patients.