Background Accurate detection of axillary lymph node invasion is essential for staging breast cancer and optimizing treatment. This study aimed to develop a radiomics-based machine learning model to predict LNI and compare its performance with the Node-RADS scoring system.

Methods This retrospective study included BC patients undergoing preoperative multiparametric MRI and lymph node dissection. Stable and replicable radiomic features were extracted from subtraction and T2-weighted MRI sequences using TRACE4©. Five machine learning models were trained for binary LNI classification, using histopathology as the reference standard. The best-performing model was externally validated using an independent cohort. Performance metrics included sensitivity, specificity, positive predictive value, negative predictive value, and AUC. The Node-RADS scoring system was used for comparison in the external dataset.

Results Of 93 cases, 40 (43%) were LNI-positive; 17 stable features (ICC>0.75) were selected for model development. The best-performing model achieved an AUC of 81% (95% CI:78–85), accuracy of 75% (95% CI:70–79), sensitivity of 52% (95% CI:41–62), specificity of 92% (95% CI:86–98), PPV of 85% (95% CI:76–95), and NPV of 72% (95% CI:68–76) on the internal dataset. External validation, with 18 samples, confirmed robustness, with AUC of 94% (95% CI: 89-99), sensitivity of 89% (95% CI: 83-95), specificity of 100% (95% CI:100–100). No significant differences in accuracy, sensitivity, or specificity (p>0.05) were observed between the model and Node-RADS, with moderate agreement (Cohen’s kappa=0.72).

Conclusions The radiomics-based model demonstrated comparable performance to Node-RADS, confirming its potential as a complementary decision-support tool in BC management.