Integrating multimodal ultrasound imaging for improved radiomics sentinel lymph node assessment in breast cancer

Abstract

Background: Accurate preoperative assessment of sentinel lymph node (SLN) is critical for treatment planning in breast cancer (BC). While SLN biopsy (SLNB) remains the gold standard, it is invasive and may be unnecessary for all patients, particularly those with clinically node-negative disease. Combining conventional B-mode ultrasound (BMUS) and color Doppler ultrasound (CDUS) with new techniques like radiomics and deep learning may improve SLN prediction, but this approach has not been widely studied yet. This retrospective study aims to develop and validate a deep learning radiomics model that combining BMUS and CDUS imaging to noninvasively predict SLN metastasis in patients with BC.

Methods: A total of 450 women with invasive BC who were treated at 2 hospitals between October 2021 and March 2025 were retrospectively analyzed. Patients were divided into training (n=276), external validation (n=105), and testing (n=69) sets. Handcrafted features were extracted from the breast lesion areas and its surrounding areas in BMUS images. Deep learning-based features were derived by applying a fine-tuned dual-stream MobileNetV2-based model, ultrasound and color doppler network, to both BMUS and CDUS images. The extracted deep features were then subjected to dimensionality reduction using principal component analysis. Following this, both handcrafted and deep learning features underwent further feature selection and dimensionality reduction process via application of inter- and intraclass correlation coefficient filtering, Pearson correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. Three models were constructed: only handcrafted features (ONLY_HF), only deep-learning features (ONLY_DF), and combined features (COMB). Each model's performance was evaluated using the area under the curve (AUC).

Results: The COMB model integrated ten features (six handcrafted and four deep learning) following LASSO regression. In predicting SLN metastasis between N0 and N≥1, COMB achieved a higher AUC (0.888, 0.861, and 0.837 in the training, validation, and testing sets, respectively) compared to ONLY_HF (0.792, 0.765, and 0.739) and ONLY_DF (0.781, 0.748, and 0.717). The negative prediction value of COMB was the highest (88.89%, 76.60%, and 71.23%), followed by ONLY_HF (83.33%, 72.00%, and 43.10%), and ONLY_DF (78.38%, 67.57%, and 52.69%).

Conclusions: By integrating BMUS and CDUS imaging with advanced deep learning techniques, the COMB model achieved a high negative predictive value, which could guide axillary treatment decisions and reducing unnecessary invasive procedures. These findings highlight the potential of multimodal imaging and machine learning strategies to serve as noninvasive, supplementary tools for personalized BC management.

Keywords: Sentinel lymph node metastasis (SLN metastasis); breast cancer (BC); deep learning; ultrasound.

Figure 1

Overview of the study design, model architecture, and classification outcomes. (A) Patient recruitment and model construction workflow. (B) Architecture of the proposed dual-modality deep learning model (USCD-Net). (C) Quantitative evaluation and interpretability of the USCD-Net model. Model performance is illustrated using ROC curves (shown as illustrative examples). CAM visualization highlights key regions in ultrasound images to enhance interpretability. CAM, class activation map; ONLY_DF, only deep-learning features; ONLY_HF, only handcrafted features; RES-COMB, deep-learning features and handcrafted features; ROC, receiver operating characteristic; SVM, support vector machine; USCD-Net, ultrasound and color Doppler network.

Figure 2

Patient enrollment workflow.

Figure 3

Radiomic feature selection and LASSO regression analysis. (A) Heatmap of the 22 handcrafted features selected after Pearson correlation analysis (threshold: 0.70), illustrating the inter-feature correlation matrix and identifying potential redundancies. (B) LASSO cross-validation path (MSE), displaying the selection of the optimal alpha value (0.03199) and the corresponding mean absolute error across multiple folds. (C) Coefficient path of the six radiomic features selected by LASSO regression, highlighting the retention of first-order, GLCM, GLSZM, and NGTDM features and their respective coefficients. GLCM, gray level co-occurrence matrix; GLSZM, gray level size zone matrix; LASSO, least absolute shrinkage and selection operator; MSE, mean square error; NGTDM, neighboring gray tone difference matrix.

Figure 4

Diagnostic performance of the three models used to (A) classify N0, N1–2, and N ≥3 in the test cohort and (B) their 5-fold cross-validation results. AUC, area under the curve; COMB, combined both traditional and deep features via the USCD-Net; N, node; ONLY_DF, only deep-learning features; ONLY_HF, only handcrafted features; ROC, receiver operating characteristic; USCD-Net, ultrasound and color Doppler network.

Figure 5

Visualization of five examples with SLN metastasis and non-metastasis each. Each example presents a conventional ultrasound image and its corresponding heat map, where the red areas indicate higher weights, as illustrated by the color bar on the right. The images highlight the importance of the lobulated areas and tumor boundary in predicting SLN metastasis status. CAM, class activation map; SLN, sentinel lymph node; USCD-Net, ultrasound and color Doppler network.