Identification of sentinel lymph node macrometastasis in breast cancer by deep learning based on clinicopathological characteristics

Abstract

The axillary lymph node status remains an important prognostic factor in breast cancer, and nodal staging using sentinel lymph node biopsy (SLNB) is routine. Randomized clinical trials provide evidence supporting de-escalation of axillary surgery and omission of SLNB in patients at low risk. However, identifying sentinel lymph node macrometastases (macro-SLNMs) is crucial for planning treatment tailored to the individual patient. This study is the first to explore the capacity of deep learning (DL) models to identify macro-SLNMs based on preoperative clinicopathological characteristics. We trained and validated five multivariable models using a population-based cohort of 18,185 patients. DL models outperform logistic regression, with Transformer showing the strongest results, under the constraint that the sensitivity is no less than 90%, reflecting the sensitivity of SLNB. This highlights the feasibility of noninvasive macro-SLNM prediction using DL. Feature importance analysis revealed that patients with similar characteristics exhibited different nodal status predictions, indicating the need for additional predictors for further improvement.

Keywords: Breast cancer; Clinical decision support; Deep learning; Lymphatic metastasis; Sentinel lymph node.

Fig. 1

Deep learning workflow. The overall study cohort was divided into a development set (patients diagnosed between 2014 and 2016) and a test set (patients diagnosed in 2017). Five multivariable machine learning (ML) algorithms (logistic regression [LR], multilayer perceptron [MLP], ResNet, Transformer, and CatBoost) were trained on the development set to predict sentinel lymph node macrometastasis (macro-SLNM, a metastasis > 2 mm). Several powerful deep learning (DL) strategies were employed to take full advantage of the prediction models, including feature tokenizers for efficient feature embedding; weighted binary cross-entropy loss, focal loss, and triplet loss to address the imbalanced distribution of macro-SLNMs; and Bayesian optimization of the hyperparameter search. Internal validation was performed using fivefold cross validation. The trained fivefold models of each multivariable algorithm and the univariable model using only tumor size (T-size) were evaluated on the test set to estimate predictive performance. Performance metrics, including the area under the receiver operating characteristic (ROC) curve (AUC) and the precision recall (PR) AUC, were calculated. In addition, when the sensitivity was set to at least 90%, the specificity, negative predictive value, and positive predictive value were calculated based on the average performance across all five folds. Finally, Shapley Additive exPlanations was applied to evaluate the feature importance for each of the five multivariable algorithms and for individual patients.

Fig. 2

Logistic regression (LR) and multilayer perceptron (MLP) exhibit weak advantages over Transformer and outperform the remaining models on overall performance. (A) Receiver operating characteristic (ROC) and (B) precision recall (PR) curves for fivefold models of all multivariable algorithms on the test set. The ROC/PR curves of the univariable model based on only tumor size (T-size) serve as a shared benchmark (dashed line in black). Presented at the top is the mean area under the curve (AUC) and standard deviation across all 5 folds.

Fig. 3

Tumor size shows a significant lead over the other predictors. A heat map of the feature importance assessed by Shapley Additive exPlanations (SHAP) is presented for each of the five multivariable models. The predictors are ranked by average importance across all models, with decreasing values from top to bottom. LR, logistic regression; MLP, multilayer perceptron; PgR, progesterone receptor; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2.

Fig. 4

Individual interpretations indicate data limitations. Shown here are the individual Shapley Additive exPlanations (SHAP) of four different patients for the prediction of sentinel lymph node macrometastasis (macro-SLNM) based on the Transformer model. (A) True-positive, (B) false-positive, (C) true-negative, and (D) false-negative predictions were randomly selected from the test set. Here, f(x) is the predicted probability of macro-SLNM for the selected patient based on the transformer model, and E[f(x)] is the expectation (mean value) of predictions across the entire test set (0.145), as well as the threshold for positive and negative predictions. For each individual, the prediction starts from E[f(x)], and each predictor contributes positively (red) or negatively (blue) to the final prediction f(x). The predictors are sorted according to the absolute feature importance (contribution) of each predictor. The 26 redundant predictors are shown directly because there may be both negative and positive values within one redundant group (see Supplementary Table S1). LumA, luminal A-like; PgR, progesterone receptor; ILC, invasive lobular carcinoma; NST, no special type; C1, classification 1 [NST, ILC, others]; C2, classification 2 [NST, ILC, other or mixed]; C3, classification 3 [NST or ILC, others].