Abstract: Objective To develop and validate a subregional radiomics model based on contrast-enhanced MRI to predict pathological complete response(pCR)after neoadjuvant chemotherapy(NAC)in breast cancer patients. Methods A total of 155 breast cancer patients who underwent NAC were retrospectively included, with 108 in the training set and 47 in the validation set. Subregion clustering was performed for each patients region of interest by using the K-means algorithm, and then high-dimensional radiomics features in the patients contrast-enhanced MRI were extracted using the Pyradiomics software package for each subregion. Mann-Whitney U test, Spearman correlation coefficient and least absolute shrinkage and selection operator(LASSO)algorithm were utilized for feature screening. The filtered optimal features were used to build each subregion single model in the training set by Logistic regression algorithm, and the same algorithm was used to fuse the subregion single models to build the subregion total model, which was visualized by using the nomogram, and the effectiveness of all models was verified in the validation set. Results The clinicopathological factors of the patients were not statistically different between the training and validation sets. The best clustering results for patients based on the K-means algorithm were 5 classes, and the area under curve(AUC)of the receiver operating characteristic(ROC)of the total subarea model built from the fusion of the 5 subarea single models in the training set and validation set was 0.898( 0.839-0.957)and 0.828(0.695-0.961), respectively. In the validation set, the ROC curves, clinical decision curves and calibration curves suggested that the subarea total model had better discriminative ability, calibration ability and clinical utility than the subarea single model. Conclusion The subregion total model can reveal tumor heterogeneity and help clinicians accurately predict pCR after NAC before treatment and has some ability to predict the post-treatment prognosis, highlighting the potential of artificial intelligence in improving individualized treatment for breast cancer patients.

Key words: Radiomcis, Neoadjuvant chemotherapy, Breast cancer, Prediction model, Subregion analysis