Intratumoral and Peritumoral Based Radiomics for Assessment of Lymphovascular Invasion in Invasive Breast Cancer: Model Development and Validation

Background: Preoperative non-invasive assessment of lymphovascular invasion (LVI) is critical for optimizing surgical strategy and risk stratification in breast cancer patients, yet reliable preoperative predictive tools remain insufficient. Current radiomics studies for breast cancer LVI prediction predominantly focus on intratumoral region (ITR) features, while emerging evidence indicates that the peritumoral region (PTR) harbors tumor microenvironmental alterations closely related to tumor invasion and metastasis and possesses unique diagnostic potential. However, existing studies rarely explore the predictive value of MRI-based radiomics features from different peritumoral subregions for breast cancer LVI, and there is no unified consensus on the optimal effective peritumoral expansion range for LVI prediction.

Methods: This study was designed as a retrospective diagnostic prediction model development and internal validation study. A total of 291 patients with invasive breast carcinoma who received preoperative dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI) at Gansu Maternity and Child-care Hospital (Gansu Provincial Central Hospital) between December 2019 and August 2023 were enrolled. Patients were randomly divided into a training cohort (n=204, 70%) and a test cohort (n=87, 30%) using simple random sampling. Radiomics features were extracted from ITR, PTR-5mm, PTR-10mm, ITR+PTR-5mm and ITR+PTR-10mm on preoperative DCE-MRI images. Feature consistency screening, dimensionality reduction and optimization were sequentially performed via intra-class correlation coefficients (ICC), Spearman correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation. Five radiomics prediction models based on random forest algorithm were constructed corresponding to the five feature regions. The predictive performance of each model was comprehensively assessed using receiver operating characteristic (ROC) curves, from which the area under the curve (AUC), sensitivity, specificity and accuracy were calculated., calibration curve with Hosmer-Lemeshow test, and decision curve analysis (DCA). The clinically meaningful threshold ranges of model efficacy indicators were further clarified for clinical decision-making reference.

Results: Among all constructed models, the ITR+PTR-5mm radiomics model yielded the optimal and most stable predictive efficacy. The AUC values of this model were 0.920 (training cohort) and 0.794 (test cohort), with satisfactory sensitivity and specificity. The calibration curve of the ITR+PTR-5mm model showed excellent consistency between predicted probabilities and actual observed LVI status, confirmed by favorable Hosmer-Lemeshow test results. DCA verified that this model provided the highest clinical net benefit across a wide range of threshold probabilities compared with other single-region and combined models.

Conclusions: The DCE-MRI-based ITR+PTR-5mm radiomics model exhibits favorable preliminary predictive performance for breast cancer LVI. As a single-center preliminary exploratory study without external validation, this model can serve as a potential non-invasive auxiliary tool for preoperative LVI risk assessment, providing tentative evidence for individualized clinical treatment decision-making. Further multi-center external validation is required before its formal clinical application.