Interpretable Multiregional Radiomics of Cone-Beam Breast CT for Predicting Lymphovascular Invasion and Axillary Lymph Node Metastasis In Breast Cancer: A Multi-Center Study
Abstract
Purpose
To develop and validate a cone-beam breast CT (CBBCT)-based multiregional radiomics model for identification of lymphovascular invasion (LVI) and fuse it into Memorial Sloan-Kettering Cancer Center (MSKCC) nomogram for predicting axillary lymph node (ALN) metastasis.
Methods
This retrospective study included 583 breast cancer patients from two institutions who received CBBCT examination from 2012 to 2023. Patients from Hospital #1 were allocated to training (n=327) and validation (n=109) set, while patients from Hospital #2 constituted test (n=147) set. Radiomics feature extraction was performed in intra- and peritumoral regions. Radiomics model was constructed using the best-performing RadScore derived from the two regions and their combination, followed by interpretion with SHapley Additive exPlanations (SHAP) framework. Clinical model was built on independent factors. Integrated model was developed by combining the best-performing RadScore and independent clinicopathologic predictors. The model with highest AUC was chosen as LVI radiomics biomarker (LVIRB). The predictive value of LVIRB-based MSKCC nomogram was assessed in total cohort.
Results
RadScore of combined intra- and peritumoral regions performed best. SHAP analysis showed that four intratumoral and four peritumoral features contributed most. Lesion presence, size, focality, and pathologic grade were selected as independent clinical predictors. Integrated model demonstrated the best performance with AUCs[training/validation/test] of 0.836/0.796/0.795 (vs. radiomics model 0.811/0.746/0.734, vs. clinical model 0.742/0.734/0.720, all p<0.05) and was selected as LVIRB. LVIRB-based MSKCC nomogram discriminated ALN metastasis or not with an AUC of 0.773.
Conclusion
CBBCT-based LVIRB had potential in predicting LVI status and LVIRB-based MSKCC nomogram could distinguish ALN positive from negative.