Breast Cancer Segmentation And Neoadjuvant Efficacy Prediction Based On Magnetic Resonance Imaging

Breast cancer is currently the most prevalent cancer worldwide and the leading cause of cancer-related deaths among women.Neoadjuvant chemotherapy(NAC)combined with surgery is a major treatment approach for breast cancer,but its efficacy is difficult to predict,leading to challenges in achieving accurate prognosis.With the development of radiomics and artificial intelligence,deep learning(DL)networks can extract semantic features that better represent tumor regions,providing new opportunities for precise prediction of NAC response in breast cancer.This study primarily focuses on addressing the following three issues:(1)The existing breast cancer tumor segmentation models lack accuracy and have limited validation on largescale independent datasets.(2)With increasing awareness of privacy protection and the improvement of national laws and regulations,collecting data from multiple centers,especially sensitive data,has become increasingly challenging.However,single-center data lacks sufficient sample size and diversity to train a well-generalized deep learning model.(3)Due to significant heterogeneity in tumor regions among different patients(e.g.,size,location,shape),it is difficult for global Region of Interest(ROI)information to capture local tumor details.Resampling or cropping methods used to handle tumor regions can lead to information loss and redundancy.To address the above issues,this study primarily focuses on two aspects: the automated and precise segmentation of breast cancer tumors and the accurate prediction of NAC efficacy.The main contents of this study are as follows:(1)We propose the Channel Spatial Attention Network(CSA-Net)for the segmentation of breast cancer tumors in a single-center setting,aiming to enhance the model’s feature extraction capabilities.(2)To address challenges related to limited sample size and diversity in single-center data,as well as difficulties in collecting multi-center data,this study introduces the Adaptive Weight Federated Aggregation(AWFA)algorithm based on Federated Learning(FL).By dynamically adjusting the weights of local models during FL model aggregation,the optimal global model is obtained.(3)We propose a Multi-Region of Interest(Multi-ROI)Transformer prediction model that combines contextual information from the global ROI and local ROIs for joint prediction,enabling accurate prediction of pathologic complete response(p CR)after neoadjuvant chemotherapy for breast cancer.The results of this study demonstrate that the breast cancer segmentation Dice coefficient based on CSA-Net improves by an average of 3% compared to the U-Net model across all datasets.The proposed federated learning-based breast cancer segmentation achieves a Dice coefficient of 0.739,which is closer to the centralized learning result(0.742)compared to conventional federated learning strategy(0.713).Finally,the Multi-ROI Transformer model proposed in this study achieves an area under the receiver operating characteristic curve(AUC)of 0.72,sensitivity of 0.86,and specificity of 0.84 for the prediction of p CR after NAC in breast cancer,enabling accurate prediction of NAC efficacy.