Objective To develop a nomogram prediction model based on ultrasound and mammographic features for BI-RADS category 4 breast lesions and evaluate its diagnostic value in early differentiation of benign and malignant lesions. Methods A retrospective analysis was conducted on 113 pathologically confirmed BI-RADS 4 breast lesions (42 benign, 71 malignant). Ultrasound and mammographic features were extracted. Independent predictors were selected using LASSO regression and multivariate logistic regression to construct a nomogram model, which was evaluated in terms of discrimination (C-index), accuracy (AUC), sensitivity, and specificity. Additionally, three machine learning models—Extreme Gradient Boosting (XGBoost), Vision Transformer (ViT), and Graph Neural Network (GNN) —were implemented and compared using training and testing datasets. SHAP value analysis and attention mechanisms were employed to enhance model interpretability. Results Lesion margin, boundary, shape, edge, and density were identified as independent predictors. The nomogram model achieved a C-index of 0.779 and an AUC of 0.758. The XGBoost, ViT, and GNN models yielded AUCs of 0.881, 0.872, and 0.890, respectively, outperforming traditional regression. GNN achieved the best sensitivity (88.0%) and F1 score (0.857). Across all models, spiculated margins, irregular shape, and high density were key predictors of malignancy. Conclusion Deep learning models integrating multimodal features—especially GNN and ViT—can significantly improve the prediction accuracy of BI-RADS 4 breast lesion malignancy risk. These models hold great promise for clinical application and may assist in optimizing biopsy decision-making.
Key words
BI-RADS category 4 breast lesions /
multimodal imaging /
extreme gradient boosting algorithm /
vision transformer model /
graph neural network /
risk prediction model /
machine learning
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
References
[1] 邓丽贞, 吴莉. 基于多模态超声特征的卷积神经网络乳腺癌分子亚型诊断研究[J]. 湖南师范大学学报 (医学版), 2024, 21(5): 75-80.
[2] ELMI M, WAKEAM E, AZIN A, et al.Surgical morbidity of full-thickness chest wall resection for breast cancer: A retrospective study of a national database[J]. J Surg Res, 2021, 10(257): 161-166.
[3] OKAZAWA A, IIMA M, KATAOKA M, et al.Diagnostic utility of an adjusted DWI lexicon using multiple b-values to evaluate breast lesions in combination with BI-RADS[J]. Magn Reson Med Sci, 2022, 23(4): 438-448.
[4] TOGAWA R, PFOB A, BUSCH C, et al.Potential of lesion‐to‐fat elasticity ratio measured by shear wave elastography to reduce benign biopsies in BI‐RADS 4 breast lesions[J]. J Ultrasound Med, 2023, 42(8): 1729-1736.
[5] TANG W, ZHOU H, QUAN T, et al.XGBoost prediction model based on 3.0T diffusion kurtosis imaging improves the diagnostic accuracy of MRI BI-RADS 4 masses[J]. Front Oncol, 2022, 17(12): 3680-3690.
[6] EZEANA C E, HE T, PATEL T, et al.A deep learning decision support tool to improve risk stratification and reduce unnecessary biopsies in BI-RADS 4 mammograms[J]. Radiol Artif Intell, 2023, 5(6): 1259-1271.
[7] LIU Y, WANG S, QU J, et al.High-temporal resolution DCE-MRI improves assessment of intra- and peri-breast lesions categorized as BI-RADS 4[J]. BMC Med Imaging, 2023, 23(1): 58-71.
[8] 宋美娜, 何花, 王志军, 等. 集成磁共振和扩散加权成像多定量参数联合DISCO的列线图预测BI-RADS 4类乳腺肿块性病变良恶性的价值[J]. 中国医学影像学杂志, 2023, 31(10): 1035-1042.
[9] 吴慕尧, 刘奕显, 郑婵娟, 等. 一种预测三阴性乳腺癌生存以及药物敏感性的新型失巢凋亡相关基因模型[J]. 湖南师范大学学报 (医学版), 2023, 20(3): 16-23, 53.
[10] HOSSEINZADEH M, SAHA A, BRAND P, et al.Deep learning-assisted prostate cancer detection on bi-parametric MRI: minimum training data size requirements and effect of prior knowledge[J]. Eur Radiol, 2022, 32(4): 2224-2234.
[11] EGHTEDARI M, CHONG A, RAKOW-PENNER R, et al.Current status and future of BI-RADS in multimodality imaging, from the AJR special series on radiology reporting and data systems[J]. AJR Am J Roentgenol, 2021, 216(4): 860-873.
[12] BADU-PEPRAH A, OTOO O K, AMAMOO M, et al.Breast imaging reporting and data system for sonography: Positive and negative predictive values of sonographic features in Kumasi, Ghana[J]. Transl Oncol, 2024, 45: 101976.
[13] CURY S S, KUASNE H, SOUZA J S, et al.Interplay between immune and cancer-associated fibroblasts: a path to target metalloproteinases in penile cancer[J]. Front Oncol, 2022, 12: 935093.
[14] ZHOU L, ZHANG M, FU J, et al.Correlation between imaging features and expression of MMP9 and cancer-associated fibroblasts in breast cancer[J]. BMC Cancer, 2018, 18(1): 780.
[15] MA D, WANG C, LI J, et al.Analysis of the diagnostic efficacy of ultrasound, MRI, and combined examination in benign and malignant breast tumors[J]. Front Oncol, 2025, 15: 1494862.
[16] DEAN T, LI N T, CADAVID J L, et al.A TRACER culture invasion assay to probe the impact of cancer associated fibroblasts on head and neck squamous cell carcinoma cell invasiveness[J]. Biomater Scie, 2020, 8(11): 3078-3094.
[17] CUESTA C, ALCÓN-PÉREZ M, ZHENG J, et al. RAS-PI3K Pathway in CAFs Shapes Physicochemical Properties of Tumor ECM to Impact Tumor Progression[J]. bioRxiv preprint, 2025, https://doi.org/10.1101/2025.01.633776.
[18] CANNONE S, GRECO M R, CARVALHO T M A, et al. Cancer associated fibroblast (CAF) regulation of PDAC parenchymal (CPC) and CSC phenotypes is modulated by ECM composition[J]. Cancers, 2022, 14(15): 3737.
[19] WANG Z, YAN N, SHENG H, et al.Single-cell Transcriptomic Analysis Reveals an Immunosuppressive Network Between POSTN CAFs and ACKR1 ECs in TKI-resistant Lung Cancer[J]. Cancer Genomics Proteomics, 2024, 21(1): 65-78.
[20] SHUKLA N, NAIK A, MORYANI K, et al.TGF-β at the crossroads of multiple prognosis in breast cancer, and beyond[J]. Life Sci, 2022, 310: 121011.
[21] WEBER J, ZANETTI G, NIKOLOVA E, et al.Potential of non-contrast spiral breast CT to exploit lesion density and favor breast cancer detection: A pilot study[J]. Eur J Radiol, 2024, 178: 111614.
[22] DOLSKII A, ALCANTARA DOS SANTOS S A, ANDRAKE M, et al. Exploring the potential role of palladin in modulating human CAF/ECM functional units[J]. Cytoskeleton, 2025, 82(3): 175-185.
PDF(2026 KB)
