Prediction of histological grading in ductal carcinoma in situ based on mammographic signs and clinical information...

This study aims to investigate the feasibility of constructing machine learning models based on mammographic signs and clinical information to predict histological grading in ductal carcinoma in situ (DCIS) of the breast.

A retrospective analysis was conducted on mammographic signs and clinical data from 243 patients diagnosed with breast DCIS, confirmed by pathology. The patients were divided into non-high-grade (n=110, including low- and intermediate-grade) and high-grade (n=133) groups based on histological results. Statistical analysis was performed on 10 clinical variables and key mammographic features (calcification presence, morphology, and distribution) based on the BI-RADS lexicon, and the features with significant differences were selected to develop three machine learning models: eXtreme Gradient Boosting (XGBoost), logistic regression (LR), and multinomial Naive Bayes (MNB). The models' performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC) as the primary metric, with pairwise AUC comparisons performed using the DeLong test.

The AUC values for the training sets of XGBoost, LR, and MNB were 0.788 (95% CI: 0.744-0.832), 0.796 (95% CI: 0.752-0.840), and 0.806 (95% CI: 0.761-0.851), respectively. The AUC values for the test sets were 0.763 (95% CI: 0.709-0.818), 0.756 (95% CI: 0.705-0.807), and 0.784 (95% CI: 0.735-0.833). The accuracy values were 0.761, 0.758, and 0.776; the sensitivity values were 0.726, 0.824, and 0.808; and the specificity values were 0.725, 0.692, and 0.744. Although MNB achieved the numerically highest performance, pairwise AUC comparisons showed no statistically significant differences among the three models (all p > 0.05), indicating comparable discriminative ability.

Machine learning-based models for predicting histological grading in DCIS show promising performance, with MNB demonstrating competitive predictive efficiency alongside the advantage of probabilistic interpretability. The findings highlight the potential utility of integrating mammographic features and clinical information for enhancing the accuracy of DCIS grading prediction. These results should be regarded as hypothesis-generating, pending external multi-center validation.