Prediction Of The Efficacy Of Neoadjuvant Chemotherapy For Breast Cancer Based On Histomorpholog

Breast cancer is a malignant tumor that seriously threatens women’s health worldwide.Neoadjuvant chemotherapy can make breast cancer patients achieve pathological response,its role is mainly manifested in: reducing tumor volume,making unresectable tumors resectable,improving breast-conserving rate and survival time of patient.At present,clinicians judge whether to use neoadjuvant chemotherapy according to the pathological diagnosis and immunohistochemical subtype classification of patients.However,there is a lack of accurate assessment of patient benefit from neoadjuvant chemotherapy,relying only on physicians to assess benefit with clinical information is difficult to quantify,and the assessment methods lack generalizability.For patients who do not benefit from neoadjuvant chemotherapy,there is a high risk that patients will miss the optimal time for treatment.The purpose of this paper is to quantify the microenvironment of tumor-infiltrating lymphocytes from the preoperative pathological puncture sections of patients using computerized image processing techniques,so as to predict the degree of remission of neoadjuvant chemotherapy and provide a reference for the clinical diagnosis of patients.Firstly,we used deep learning methods for tissue and cell nucleus segmentation of patient pathology puncture slides.In the tissue segmentation,GCRes2Net50 network was applied to implement the segmentation of seven types of tissues: invasive tumor,tumor-associated stroma,in-situ tumor,healthy glands,necrosis not in-situ,inflammatory stroma and other tissues.In the model training stage,2688 images(512×512)from the publicly available dataset TIGER were used for training and testing.In the model inference stage,the image patches of the patient’s pathological puncture slice were obtained using sliding window patches(512×512)and fed into the network for prediction,and the segmentation results were mapped by coordinates to obtain the tissue segmentation results of the whole puncture slice and the tissue images of the tumorassociated stroma were extracted.Then the MTUNet network was used to implement segmentation of five types of nuclei: inflammatory,epithelial,connective,neoplastic and dead,of which the tumor-infiltrating lymphocytes of interest belong to inflammatory cells.Models were trained and tested with the Pan Nuke public dataset,which included a total of 7904(256×256)images,and the models were used for nuclei segmentation of patient puncture slices.In tissue segmentation,the GCRes2Net50 model reached 0.737 MIo U and 0.811 FWIo U,and in nucleus segmentation,the MTUNet model achieved an average Dice coefficient of 72.81%Secondly,a two branch G-UNet based on graph convolutional networks were applied for response prediction of neoadjuvant chemotherapy.The work in this chapter enables neoadjuvant chemotherapy response prediction by quantitatively describing the microenvironment of tumor infiltrating lymphocytes.Firstly,the textural features,shape features and cell graph features of tumor infiltrating lymphocyte nuclei were obtained using histomorphological analysis.Then these features were converted into node features of the graph convolutional network and fed into a two branch G-UNet network for training and prediction.The ability of the two-branch G-UNet network to aggregate nodal features is utilized to more effectively characterize multidimensional cell nuclei while paying more attention to the nuclei properties that provide important information for prediction.344 breast primary and 257 axillary lymph node puncture slices from 257 patients were used to predict the response of neoadjuvant chemotherapy.The model achieved AUC values of 0.747 and 0.808 for breast primary and axillary lymph node puncture respectively,and an AUC value of 0.812 for overall patient prediction.In the prediction model for patients with different molecular,the AUC of Luminal B was 0.845 and TNBC was 0.789.In this paper,we used histomorphological methods and deep learning techniques to quantitatively analyze the tumor microenvironment of tumor infiltrating lymphocytes in patients and effectively predict the response of neoadjuvant chemotherapy,providing an objective basis for doctors to judge whether to perform neoadjuvant chemotherapy.