Diagnostic Value Of Small Breast Cancer Based On Multimodal MRI Radiomics

Objective:Multiple machine learning models for identifying benign and malignant small breast masses(≤2cm)by mining the radiomics features of MRI images.Exploring the potential value of multimodality MRI-based radiomics in identifying small breast cancer applications.Methods:This study retrospectively analyzed 110 patients with small breast masses(128 small masses in total)who underwent breast MRI between September 2017 and October 2020 at our institution and met the inclusion criteria.Collection of DICOM images of each sequence of MRI T1 WI,T2WI,DWI,ADC,DCE-MRI(5 unimodal groups)of the study subjects and clinical and pathological data.And with pathological results as the gold standard,the research objects are divided into two groups: benign and malignant.Segmentation of 3D regions of interest on T1 WI,T2WI and DCE-MRI images and 2D regions of interest on DWI and ADC images by texture analysis software using semi-automatic segmentation method.Extract radiomics features from the region of interest separately,and extract a total of 483 quantitative features for each patient.T1 WI and T2 WI were classified as the plain scan group,DWI and ADC images were classified as the diffusion group,and DCE-MRI as the enhancement group.Combine these 3 basic groups into 4 multi-modal groups: as plain combined enhancement group(T1WI+T2WI+DCE-MRI),plain combined diffusion group(T1WI+T2WI+DWI+ADC),diffusion combined enhancement group(DWI+ADC+DCE-MRI),and plain combined enhancement combined diffusion group(T1WI+T2WI+DWI+ADC+DCE).Correlation-based feature selection(CFS)is used to filter key features.The best subset of features was selected from the radiomic features of five unimodal groups,three basic groups,and four multimodal groups,respectively.4 types of machine learning models based on 5 groups of unimodal best subsets of radiomics features: random forest(RF),naive bayesian(NB),support vector machine(SVM)and k nearest neighbor(KNN),respectively(20 in total).And select the best machine learning algorithm by comparing the performance of each model.Applying the best machine learning algorithm to the best subset of radiomics features for 3 basic groups and 4 multimodal groups to build the model.And plot ROC curves for each model and evaluate their effectiveness.The above radiomics models were validated for diagnostic efficacy using a 10-fold cross-validation method.Results:Of the 128 small masses,60 benign and 68 malignant masses were included.Nine key features were selected from the radiomics features of five unimodal groups,three basic groups(enhanced group with unimodal DCE-MRI)and 12 key features each of four multimodal groups by CFS algorithm,respectively.By comparing 20 models with 4classes of classifiers constructed based on 5 unimodal modes,it is found that the classification models constructed based on the RF algorithm have better and stable performance for the dataset in this study.The ADC-based RF model had the highest diagnostic efficacy(AUC value of 0.837,accuracy of 79.69%).Therefore,the RF algorithm was used for all subsequent construction models.Among the three basic groups,the RF classification model based on the diffusion group had high diagnostic efficacy with an AUC of 0.847 and an accuracy rate of 79.69%.Among the four multimodal groups,the RF classification model constructed in the flat-scan combined with enhancement combined with diffusion group showed the best diagnostic efficacy among all models in this study with an AUC of 0.926 and an accuracy of 87.5%,followed by the diffusion combined with enhancement group(AUC value of 0.906 and an accuracy of82.03%).Conclusion:Quantitative features extracted from multimodal MRI images to construct radiomics models are of high value for the diagnosis of small breast cancer(≤ 2cm).The best diagnostic efficacy of the RF model was based on the plain combined enhancement combined diffusion group(T1WI+T2WI+DCE-MRI+DWI+ADC).This is important for the early diagnosis of small breast lesions.