Breast cancer is the most common malignant tumor among women worldwide, having overtaken lung cancer in incidence since 2020. Early detection of axillary lymph node metastasis (ALNM) is critical for breast cancer management and prognosis, as it determines the degree of metastasis and influences overall survival outcomes.
Sentinel lymph node biopsy (SLNB) is currently the standard method for assessing lymph node involvement, but it is an invasive surgical procedure with contraindications and potential complications. There is therefore an urgent need for noninvasive, accurate methods to predict sentinel lymph node metastasis (SLNM) preoperatively using imaging techniques such as ultrasound.
Several studies have demonstrated a link between ultrasonographic manifestations and axillary lymph node metastasis in breast cancer. However, most traditional radiomics approaches rely on manual extraction of image features by experienced radiologists, which is subjective and time-consuming.
Deep learning techniques, particularly convolutional neural networks (CNNs), offer a major advantage by automatically learning high-dimensional abstract image features from data, reducing the burden of manual feature selection. This study aimed to develop deep learning models that allow fully automated prediction of SLNM using different types of ultrasound images.
Clinical and ultrasound data from 317 patients diagnosed with breast cancer at the Second Affiliated Hospital of Nanchang University were collected from January 2018 to December 2021. These patients were randomly divided into training and internal validation cohorts at a ratio of 7:3, yielding 222 training and 95 internal validation cases.
An external validation cohort of 42 patients was collected from Nanchang Third Hospital using the same methodology. Among all 359 patients, 170 (47.4%) had SLNM and 189 (52.6%) did not. Inclusion criteria required ultrasound examination within 14 days before surgery, clearly visible lesions, mastectomy with SLNB, and pathological confirmation of breast cancer.
Three deep learning models were proposed to predict SLNM by analyzing different ultrasound image types: DL-grayscale (using grayscale ultrasound), DL-CDFI (using color Doppler flow imaging), and DL-elastography (using elastography images). All three models adopted the same ResNet-18 architecture, which consists of a convolutional layer, pooling layer, and fully connected layer.
Images were preprocessed by converting from DICOM to JPEG format, with regions of interest manually cropped using the LabelImg annotation tool. All images were normalized to 448 by 448 pixels. The models were trained using a stochastic gradient descent optimizer with the Adam optimizer for parameter updates, cross-entropy loss function, and a learning rate decayed from 0.01 to 0.00001 over 100 epochs with a mini-batch size of 8.
The DL-elastography model achieved the best diagnostic performance among all three models. In the internal validation cohort, its AUC was 0.879 (95% CI: 0.839-0.920) with 91.60% sensitivity and 82.79% specificity. In the external validation cohort, its AUC remained high at 0.876 (95% CI: 0.813-0.970) with 88.46% sensitivity and 82.35% specificity.
The DL-grayscale model achieved AUCs of 0.855 and 0.788 in the internal and external validation cohorts respectively. The DL-CDFI model performed lowest with AUCs of 0.761 and 0.728. The overall classification accuracy of DL-elastography was 86.13% internally and 85.00% externally, confirming strong generalization ability.
Confusion matrices for both internal and external validation cohorts revealed the classification accuracy of each model. In the internal validation cohort, DL-elastography correctly classified 120 true positives and 178 true negatives, while misclassifying 37 false positives and 11 false negatives.
The positive predictive value of DL-elastography was 76.43% and the negative predictive value was 94.18% in the internal validation cohort. The high negative predictive value is particularly important clinically, as it means the model is reliable at identifying patients who do not have sentinel lymph node metastasis, potentially sparing them from unnecessary invasive biopsy procedures.
Ultrasound elastography is a novel imaging technique for noninvasive assessment of tumor tissue elasticity, which has shown significant progress in breast cancer diagnosis. The underlying mechanism is the characterization of tissue stiffness using ultrasound imaging, and biological deformation in tumor tissues may influence radiological characteristics that can be encoded by deep learning algorithms.
The DL-CDFI model performed worst among the three, which the authors attributed to the characteristics of breast cancer being abundant with blood supply, causing a small discrepancy in blood flow signals that may have limited the discriminative power of CDFI images. In contrast, elastography images provided added value by capturing mechanical tissue properties that differ between metastatic and non-metastatic sentinel lymph nodes.
The proposed DL-elastography model holds good potential for clinical application as a noninvasive preoperative screening tool. With its high negative predictive value, the model could help identify patients who are unlikely to have SLNM, potentially reducing the number of unnecessary sentinel lymph node biopsies and their associated surgical complications.
Compared to traditional radiomics approaches that require experienced radiologists to manually select and extract features, the deep learning approach automates the entire feature extraction and classification process. This makes it more objective, reproducible, and potentially deployable across different clinical settings where expert radiologists may not be readily available.
The study had several limitations. It was a retrospective study based on a limited database, and multicenter prospective studies with unbalanced databases are needed to validate the robustness of the radiomics models. Additionally, there was no direct comparison with traditional radiomics methods or other established prediction tools.
The authors noted that images of the breast lesions were not combined with axillary lymph node images in the current models, which could potentially improve prediction accuracy. Future studies should address these issues by incorporating multimodal imaging data and validating results across larger, more diverse patient populations.
This study designed three deep learning models based on preoperative grayscale ultrasound, CDFI, and elastography images to predict sentinel lymph node metastasis in breast cancer. The DL-elastography model based on elastography images consistently achieved the best diagnostic performance across both internal and external validation cohorts.
The results demonstrate that deep learning radiomics applied to ultrasound images, particularly elastography, can be an effective noninvasive tool for SLNM prediction. This approach has the potential to benefit subsequent treatment planning for breast cancer patients by providing preoperative information about sentinel lymph node status without the need for invasive biopsy procedures.