MRI radiomics for diagnosing small BI-RADS 4 breast lesions: an interpretable model.

Abstract

Background: The early detection of breast cancer is crucial. Magnetic resonance imaging (MRI) offers significant advantages in the diagnosis of lesions. We aimed to develop and validate an interpretable MRI-based radiomics model to identify small Breast Imaging Reporting and Data System (BI-RADS) category 4 lesions to help radiologists with decision making.

Methods: In total, 561 patients (with 580 small BI-RADS category 4 lesions) from two centers (The First Affiliated Hospital of Zhejiang Chinese Medical University and The Affiliated Hangzhou First People's Hospital, Zhejiang University School of Medicine) were consecutively enrolled in this study, and the radiomics features of the intratumoral and peritumoral (3 mm) regions were extracted. After a series of feature selections, extreme gradient boosting (XGBoost) was used to construct the radiomics model, and the radiomics score (radscore) was calculated. Univariate and multivariate logistic regression analyses were performed to determine the pathological malignant-related clinico-radiological factors. Finally, a model was constructed that combined the radscore and clinico-radiological factors using logistic algorithms. Subsequently, our artificial intelligence (AI)-assisted strategy was validated in an external group (n=163), and its clinical utility was evaluated by measuring improvements in BI-RADS classification accuracy with AI support.

Results: The combined model demonstrated a robust predictive capability, and had area under the curve (AUC) values of 0.897 [95% confidence interval (CI): 0.862-0.931], 0.871 (95% CI: 0.803-0.934), and 0.869 (95% CI: 0.807-0.920) in the training, internal validation, and external validation groups, respectively. Additionally, the contribution of each feature to the radiomics and combined models was illustrated using the SHapley Additive exPlanations (SHAP) algorithm, a method for interpreting machine-learning models. Further, the AI-assisted strategy improved the two radiologists' AUC values in the two modes (the 4b+ and 4c) significantly.

Conclusions: An interpretable combined model based on MRI was developed to distinguish between benign and malignant small BI-RADS4 lesions to assist radiologists to make more accurate diagnostic decisions.