Explainable machine learning based on habitat analysis of axillary lymph node ultrasound images for preoperatively noninvasive evaluation of axillary lymph node metastasis in breast cancer: a bi-center study

Background: Accurate evaluation of axillary lymph node (ALN) is crucial for guiding staging and treatment strategies in breast cancer (BC) patients. This study aimed to develop an optimal machine learning model for predicting ALN status by utilizing both conventional radiomics and habitat analysis based on axillary B-mode ultrasound (BMUS) images, offering a powerful, non-invasive means to quantify and map ALN heterogeneity, providing deeper insights into biological behavior of ALN metastasis. Methods: This study retrospectively analyzed the preoperative BMUS images of ALNs in 297 patients with BC from Hunan Cancer Hospital and Yueyang Central Hospital. Patients were divided into training (n=172), test (n=45), and external validation (n=80) sets. Habitat features were segmented into sub-regions using K-means clustering. Both habitat and conventional radiomics models were developed using twelve widely adopted machine learning algorithms respectively, including least absolute shrinkage and selection operator (LASSO), the support vector machine (SVM), K-nearest neighbors (KNN), random forest (RF), ExtraTrees, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), naive Bayes (NB), adaptive boosting (AdaBoost), gradient boosting (GB), logistic regression (LR) and multilayer perceptron (MLP). Model performance was assessed using receiver operating characteristic (ROC) curve, calibration curve, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) values were used to identify key features and enhance model transparency and interpretability. Results: Most habitat models exhibited better performance with higher area under the curve (AUC) values in external validation sets compared to conventional radiomics models. A comparative analysis of the diagnostic efficacy of these models demonstrated that the habitat-based MLP model exhibited superior efficacy, achieving the highest AUC value of 0.913 [95% confidence interval (CI): 0.842-0.967], an accuracy of 87.62% (95% CI: 79.97-93.75%), and an F1 score of 0.84 (95% CI: 0.73-0.93) in the external validation set. SHAP provided further insight into the contributions of each feature to the model's outcomes. Conclusions: We developed and validated machine learning models utilizing habitat-based ALN ultrasound images, demonstrating outstanding predictive performance for ALN metastasis in BC compared to conventional radiomics models. Among all machine learning models, the habitat-based MLP model showed the best predictive efficacy. The prediction process was visualized using SHAP, holding promise as a non-invasive tool for preoperative assessment of ALN status and potentially supporting surgeons in developing evidence-based, risk-stratified surgical strategies.