Radiomics-based MRI models for predicting breast cancer axillary lymph node involvement in comparison with Node-RADS:
Roberto Maroncelli1,2, Veronica Rizzo3, Marcella Pasculli3
1Department of Radiological, Oncological and Pathological Sciences, Sapienza-University of Rome, Rome, Italy. roberto.maroncelli@uniroma1.it.
Background:
Detection of axillary lymph node (LN) involvement is essential for staging breast cancer and optimizing treatment. This proof-of-concept two-center study explored the feasibility of magnetic resonance imaging (MRI) radiomics-based machine learning models to predict LN involvement and compare their performance with node reporting and data system (Node-RADS).
Materials And Methods:
We retrospectively included breast cancer patients undergoing preoperative multiparametric MRI and LN dissection (January 2020-September 2024). Stable radiomic features (intraclass correlation coefficient ≥ 0.75) were extracted from contrast-enhanced, subtracted, and T2-weighted sequences. Five machine learning models were trained for binary LN involvement classification, using histopathology as a reference standard. The best-performing model was externally validated on an independent cohort. Performance metrics included sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUROC). Node-RADS (scores > 2 indicating LN involvement) was used for comparison in the external dataset.
Results:
Of 93 cases, 40 (43%) were LN involvement-positive; 17 stable features were selected for model development. The best-performing model achieved 81% AUROC (95% confidence interval 78-85%), 75% accuracy (70-79%), 52% sensitivity (41-62%), 92% specificity (86-98%), 85% PPV (76-95%), and 72% NPV (68-76%) on the internal dataset. External validation (18 cases) showed promising results: 94% AUROC (89-99%), 89% sensitivity (52-100%), 100% specificity (66-100%); in this small cohort, accuracy, sensitivity, and specificity did not differ significantly versus Node-RADS, with moderate agreement (Cohen κ = 0.47).
Conclusion:
In this preliminary series, the model showed performance metrics in predicting LN involvement comparable to Node-RADS.
Relevance Statement:
Radiomics-based MRI models may represent a promising investigational tool for noninvasive axillary LN assessment in breast cancer. The performance comparable to Node-RADS suggests a potential to support clinical decision-making in the context of axillary de-escalation surgery.
Key Points:
Radiomics uses MRI to predict breast cancer LN involvement non-invasively and accurately. Radiomics and Node-RADS showed comparable performance. Radiomics could reduce invasive procedures, supporting personalized treatments in breast cancer care.
