Study Of Model Construction For Breast Cancer Computer Aided Diagnosis And Treatment Outcome Prediction

Breast cancer is one of the most threatening cancer diseases to women’s health,and its mortality rate is among the highest in female cancers.Early detection,diagnosis,and timely treatment can effectively reduce the mortality rate and improve the prognosis of patients.However,low diagnostic accuracy and missed diagnoses still exist in the screening and diagnosis process due to clinical as well as equipment factors.For patients with locally advanced and operable breast cancer,NeoAdjuvant Chemotherapy(NAC)regimen has been widely used,which can create better physiopathological conditions for the subsequent treatment of patients.However,NAC suffers from a low Pathologic Complete Response(pCR)rate.With the development of artificial intelligence(AI)technology and the improvement of computer computing power,radiomics and deep learning algorithms have been proposed,which can mine high-dimensional and deep-level features and enhance the predictive capability of aided diagnostic models.Therefore,AI technology is expected to assist radiologists in improving diagnostic accuracy and providing assistance in the development of personalized treatment plans for patients.In this paper,we conducted a study related to the assisted diagnosis of breast lesions and the prediction of the response of NAC before treatment using AI technology.The main studies are as follows:1.An aided diagnostic model of breast microcalcification cluster lesions based on radiomicsBreast microcalcification cluster lesions are an important sign for breast cancer diagnosis.However,unlike mass lesions that can be easily quantified and analyzed,microcalcification cluster lesions are difficult to quantify in terms of intensity(brightness),shape,or diameter in digital breast tomosynthesis(DBT)images.Therefore,in this paper,we propose a radiomics model to assist radiologists in diagnosis through quantitative analysis and model prediction of microcalcification clusters.Firstly,a semi-automatic segmentation algorithm is used to obtain microcalcification cluster segmentation labels.Secondly,radiomic features are extracted from local microcalcification points and global microcalcification clusters.Finally,a radiomic classification model is constructed by feature selection and training of a random forest classifier.The experimental results verify that the intensity(brightness)of individual microcalcification points and morphological distribution of microcalcification clusters are crucial for diagnosis.In addition,robustness analysis experiments for semi-automatic segmentation verified the stability of the model framework.This study is the first to construct a diagnostic model in real clinical DBT data,which effectively helps radiologists accurately diagnose microcalcification cluster lesions in patients.2.Complementary Feature Fusion Network for Breast Lesion Aided DiagnosisExisting studies of multiple sequence Magnetic Resonance Imaging(MRI)assisted diagnosis models mainly embed different sequences of MRI information to the same feature space for correlation learning to aid diagnosis.However,learning only the correlation information of different sequences of MRI in the same feature space implies the loss of a large amount of sequence-specific information,which may limit the upper bound of feature discrimination capability.Therefore,we construct a complementary feature fusion convolutional neural network(CFF-Net)to explore the potential of complementary information incorporation for model classification capability enhancement.The model mines sequence-specific features through the design of the complementary feature extraction structure.At the same time,the complementary features are enhanced by using generative adversarial strategies.Finally,by imitating the process of radiologists’ diagnosis,the Feature Attention(FA)module is designed to adaptively weight different sequence features to improve the efficiency of feature fusion.The results of the study validate the value of complementary features in model performance improvement.In addition,comparison experiments in multiple centers and different combinations of MRI sequences validate that the model has good robustness.3.A multi-source information fusion network for predicting the response of NeoAdjuvant Chemotherapy in breast cancerNone of the single data source models based on dynamic contrast-enhanced MRI(DCE-MRI)images,pharmacokinetic parameters,or molecular information of pathology can effectively predict the pathological response of patients before NAC treatment.Therefore,in this paper,for the first time,we fuse information from these three types of data sources and design a multi-source information fusion network to achieve pCR prediction before NAC treatment.The results showed that the model’s multi-source information fusion of spatial information,tumor microvascular function,and microscopic tumor cell characteristics enabled the model to achieve better response prediction performance.The predictive ability of the model is similar to that of the model constructed based on data after multiple cycles of NAC treatment.However,the prediction before NAC treatment is more clinically valuable.The model provides important guidance value for the development of accurate personalized treatment plans for breast cancer patients.