Malignancy Prediction of Breast Lesion with Calcifications via Radiomics Analysis Based on Intratumoral and Peritumoral Regions of Ultrasound Images.
Authors
Affiliations (3)
Affiliations (3)
- Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education/ Beijing), Breast Center, Peking University Cancer Hospital & Institute, 52 Fucheng Rd, Beijing 100142, China (X.W., N.Z.,J.M., W.Q., S.L., H.C., L.H.).
- Huafang Hanying Medical Technology Co., Ltd, Beijing, China (J.L.).
- Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education/ Beijing), Breast Center, Peking University Cancer Hospital & Institute, 52 Fucheng Rd, Beijing 100142, China (X.W., N.Z.,J.M., W.Q., S.L., H.C., L.H.). Electronic address: [email protected].
Abstract
To evaluate whether intratumoral habitats and peritumoral regions on ultrasound (US) images enhance radiomics-based diagnosis of calcified breast lesions. Mammography-confirmed calcified breast lesions were retrospectively collected from a single center and divided into training and internal test sets. An external test set of 84 cases was obtained from two public datasets. Pathology was the reference for biopsied lesions, while benignity in non-biopsied cases required ≥24-month follow-up. Intratumoral clustering defined optimal habitat numbers. Lesion ROIs were expanded by 3, 5, and 7 pixels to capture peritumoral regions. Machine-learning models were developed and evaluated internally and externally using Receiver Operating Characteristic (ROC) analysis, and compared with four experienced sonographers. A total of 510 lesions from 492 female were enrolled. Four intratumoral clusters were identified, resulting in eight region-specific radiomics models. In the internal test set, most models performed well except Habitat2_LR (Area under the Receiver Operating Characteristic Curve [AUC] = 0.819, p < 0.05 vs the lesion-only model). In the external test set, 3-pixel_LR (AUC = 0.780, p = 0.022) and 5-pixel_XGBoost (AUC = 0.763, p = 0.049) significantly outperformed the lesion-only model (AUC = 0.623). Calibration, decision-curve analysis, and reader comparisons supported the robustness and net benefit of the 5-pixel_XGBoost model, whose AUC was comparable to that of experienced sonographers (all p > 0.05). Incorporating peritumoral regions improves US-based radiomics diagnosis of calcified breast lesions, with the 5-pixel_XGBoost model providing the best overall performance.