Deep-Broad Learning Based Breast Cancer Tissue Segmentation And Neoadjuvant Chemotherapy Response Prediction

Breast cancer is currently the most common malignant tumor causing cancer-related deaths in women worldwide,seriously endangering women’s health.Neoadjuvant chemotherapy has become the main treatment for reducing the stage of breast cancer tumors,but its efficacy is difficult to predict due to individual differences,which makes patients bear unnecessary treatment risks and difficult to improve their prognosis.With the development of medical technology and artificial intelligence,deep learning networks can extract rich tumor and environmental tissue(e.g.,breast fibroglandular tissue)features from dynamic contrast-enhanced magnetic resonance images(DCE-MRI),providing the opportunity for precise prediction of the response of neoadjuvant chemotherapy for breast cancer.However,current deep learning technology still faces many challenges.This thesis mainly focuses on the following two challenges:(1)The performance of existing deep learning models highly depends on high-quality data and target area annotations.However,due to factors such as MRI scanners,imaging protocols and scanning parameters,DCE-MRI has high heterogeneity,resulting in large differences in image quality.Furthermore,breast gland tissue annotation is difficult and time-consuming.In order to address the above challenge,this article mainly focuses on efficient segmentation of breast fibroglandular tissue and the main research contents are as follows:A data processing scheme based on the specific task properties of breast MRI fibroglandular tissue segmentation is proposed for breast MRI quality control and quantitative evaluation analysis.At the same time,in order to obtain relevant annotations at a low cost,a progressive growth learning strategy driven by data attributes an d guided by domain knowledge is designed based on the above data processing scheme to generate reliable breast fibroglandular tissue annotations recognized by clinicians without manual annotations.(2)Existing studies on neoadjuvant chemotherapy response prediction based on breast DCE-MRI mostly focus on a single tumor region,with few in-depth studies combining image information of its growth environment(e.g.,fibroglandular tissue)and the unique temporal characteristics(e.g.,enhancement cycles)and spatial characteristics of DCEMRI.In addition,multi-region information will increase the training burden of traditional deep learning models and affect efficiency.In order to address the above challenge,this article mainly focuses on precise prediction of neoadjuvant chemotherapy response.The main research contents are as follows: A deep-broad learning prediction method based on maximum enhancement projection oriented by multiple regions and multiple views of the breast is proposed.This method effectively integrates DCE-MRI spatial stereo and temporal enhancement deep learning features based on pre-trained models,and uses width learning for efficient training and precise prediction,effectively reducing training costs.