Value of chest Computed Tomography imaging radiomics features in predicting breast density classification

ObjectiveTo explore the correlation between the radiomic characteristics derived from chest CT scan imaging and breast density classification, to construct an imaging radiomics model that can automatically achieve CT breast density classification, and to evaluate its diagnostic efficacy, so as to provide a valuable reference for breast density AI classification through chest CT scanning images. Methods393 patients who had negative results of breast ultrasound and mammography in our hospital and had undergone CT scanning of the chest within one year were collected for retrospective analysis, and 330 patients who had consistent results of breast density classification judged by two radiologists based on mammography images were used for the construction of imaging radiomics prediction models, which involved mapping of three-dimensional regions of interest (ROI) of the breast, extraction of imaging radiomics features and dimensionality reduction, screening of dominant radiomics features, establishment of multiple four-classification prediction models, and evaluation the effectiveness of each model. ResultsA U-net neural network segmentation model was trained to automatically delineate breast ROI, extract 1427 radiomics features, and screen out 28 dominant features closely related to breast density for AI automatic classification of breast density. Among the four types of four-classification prediction models constructed based on classifier including {sscircled1} Xgboost Classifier, {sscircled2} One Vs Rest Classifier(Logistic Regression in the One Vs Rest framework), {sscircled3} Gradient Boosting, {sscircled4} Random Forest Classifier, Xgboost four-classification prediction model has the best prediction performance, and after its parameter tuning, the classification accuracy for the test set reaches 0.866, and the area beneath the curve (AUC) derived from the receiver operating characteristic curves (ROC) of the four categories of classification labels are 1.00, 0.93, 0.93,and 0.99, respectively, and the AUC of the micro-averaged ROC of each classification label is 0.97, and the AUC of the macro-averaged ROC is 0.96, indicating the best prediction performance. ConclusionChest CT plain images can provide breast density classification, a valuable information that reveal significant insights pertaining to breast cancer risk, and can be used to automatically achieve breast density four-classification through imaging radiomics model, which lays the foundation for precise and individualized breast screening programs. Introducing AI methods for automatic breast density classification can help clinical diagnosis. Author summaryWeipeng Zhou is the scheme designer of this paper, and is responsible for collecting relevant data, conducting data analysis and modeling, writing and revising the manuscript. Huimao Zhang is the corresponding author of the paper, responsible for scheme design and paper revision. This paper is funded by the fund project hosted by Huimao Zhang. Jianhua Liu is the co-author of this paper, responsible for data analysis.