Prediction Of Neoadjuvant Chemotherapy Efficacy In Breast Cancer Based On Panoramic Puncture Image Analysis

Breast cancer is the highest incidence rate and mortality rate of women.Neoadjuvant chemotherapy is an effective method to the treatment of breast cancer,which can reduce tumor size and tumor stage and transform inoperable tumor into operable tumor to improve event-free survival rate of patients.Clinicians predict whether patients will receive neoadjuvant chemotherapy according to their own experience and molecular types of patients,which has strong subjectivity and cannot accurately predict the degree of pathological response after chemotherapy for a patient.It is easy to make patients with ineffective chemotherapy miss the best opportunity of treatment.Therefore,this paper aims to analyze needle biopsy images of patients with breast cancer before neoadjuvant chemotherapy quantitatively to predict the degree of pathological response by using computer image processing technology.It will provide objective basis for doctors in diagnosis and treatment of a single patients.In the first work,two models were developed which were used to segment tumor region and nucleus in needle biopsy images based on deep learning.In tumor region segmentation,UNet++ with Res Net101 as backbone was trained for tumor region segmentation,and the performance is good.In nucleus segmentation,we add residual unit and positive and negative attention module to improve the performance of UNet.The attention module makes the model not only pay attention to the information of nuclear region,but also pay attention to the information of non-nuclear region.In the independent test set,the accuracy of the model is0.9471,which achieved accurate segmentation results of nucleus.The second work is to extract the quantitative features of histomorphology based on the result of the first work to construct prediction model of neoadjuvant chemotherapy.Specifically,the segmentation models in the first work were used to segment tumor region and nucleus of370 needle biopsy images.Based on the segmentation results,global features of tumor region and local features of nucleus were extracted with 868 dimensions,in which global feature set contains tumor region distribution and morphological information of needle biopsy images,and local features contains texture,distribution and morphology of nucleus in tumor region.Furthermore,in the three types of molecular,all combination of four different feature selection algorithms(m RMR,Wilcoxon,Relief,Fisher)and four classifiers(RF,LDA,KNN and SVM)were tried to predict chemotherapy response of patients.The experimental results in the independent test set show that the AUC of HER2 and Luminal B(Her2+)was 0.7564 and TNBC was 0.8261 in the patient-level,which were higher than that in the patch-level.In addition,the experimental results of feature selection show that nucleus pixel feature and microscopic features of transformation of gray value are very important in predicting the response of neoadjuvant chemotherapy.In this paper,we extracted global features of tumor region and local features of nucleus to quantitatively analyze needle biopsy images of patients with breast cancer to predict the response of neadjuvant chemotherapy.The model can better predict the degree of pathological response of patients with neadjuvant chemotherapy,and provide objective basis for doctors to judge whether patients with breast cancer need to be treated with neoadjuvant chemotherapy.