Multichannel deep learning-based MRI model for predicting breast cancer axillary lymph node invasion: a comparative study to Node-RADS

Introduction

Breast cancer (BC) is the most frequently diagnosed cancer and a leading cause of cancer-related death among women worldwide (1). Axillary lymph node (ALN) metastasis is one of the most critical prognostic factors, as it profoundly influences disease staging, treatment decision-making, and long-term patient outcomes (2-4). Accurate preoperative assessment of ALN status is therefore essential for guiding surgical planning, particularly regarding the need for axillary lymph node dissection (ALND), a procedure associated with considerable morbidity, including lymphedema and shoulder dysfunction (5). The current clinical paradigm is shifting toward less invasive approaches, such as sentinel lymph node biopsy (SLNB); however, a reliable, non-invasive method for definitively identifying node-negative patients could spare many individuals from unnecessary surgical intervention.

Imaging plays a central role in the preoperative assessment of ALN status. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), a standard tool for BC diagnosis and local staging, provides high-resolution morphological and functional information (6). Nevertheless, conventional radiological evaluation of lymph nodes based solely on size and morphological criteria is limited by insufficient sensitivity and specificity, resulting in a substantial number of unnecessary invasive procedures (7). To standardize reporting and improve diagnostic accuracy, structured reporting systems such as the Node Reporting and Data System (Node-RADS) have recently been introduced, which provide a composite score based on size and morphological features (8). Studies have reported moderate-to-high diagnostic accuracy and good inter-reader agreement for the Node-RADS in axillary staging (9). While Node-RADS represents a significant advancement, its performance may vary by histologic type of tumor (10), and it remains reliant on qualitative radiologist interpretation, which may limit reproducibility across diverse clinical settings.

In parallel, the advent of deep learning (DL) has opened new avenues for medical image analysis. By automatically learning intricate, high-level features from imaging data, DL models can potentially detect subtle biomarkers that are imperceptible to the human eye (11). Several studies have demonstrated the feasibility of using DL on magnetic resonance imaging (MRI) to predict ALN metastasis, with most focusing on primary tumor BC patients (12,13). However, many existing approaches are limited by their reliance on single anatomical planes or the fusion of features from disparate regions and fail to fully exploit the rich, multifaceted information embedded in volumetric DCE-MRI data. Further, the direct comparative performance of such sophisticated DL models against emerging, standardized clinical tools such as the Node-RADS remains underexplored, leaving a critical gap in understanding their relative value and potential synergy in real-world clinical workflows.

To address these gaps, we proposed a novel multichannel, multiscale DL framework designed to holistically integrate information from the primary breast tumor using axial, sagittal, coronal, and multiplanar reformat DCE-MRI images. This approach, which leveraged a robust ResNet101 backbone coupled with a transformer-based fusion module, aimed to capture a more complete and nuanced representation of tumor characteristics. The primary objective of this two-center, retrospective study was to develop and externally validate this model for the preoperative prediction of ALN metastasis in BC patients. Second, and most importantly, we sought to rigorously compare its diagnostic performance against the Node-RADS in an independent external cohort. We hypothesized that this comprehensive, multi-view DL model could achieve diagnostic accuracy comparable to that of structured Node-RADS assessment, thereby offering a powerful and automated decision-support tool to assist clinicians in personalizing surgical management and reducing unnecessary ALND procedures. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0851/rc).

Methods

Patient cohorts

This two-center retrospective study was approved by the institutional review boards of The Third Affiliated Hospital of Zhengzhou University (No. 202223901) and The First Affiliated Hospital of Henan University of Science and Technology (No. 2025-03-K0026), which waived the requirement for written informed consent due to the retrospective nature of the study. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

All DCE-MRI images and corresponding clinical data included for model construction and validation were acquired strictly before any surgical intervention, SLNB, or neoadjuvant therapy initiation. Between May 2019 and June 2024, consecutive patients diagnosed with invasive breast carcinoma via histopathological examination were retrospectively enrolled in the study. All eligible cases had undergone preoperative breast MRI, with identical enrollment criteria implemented at the two collaborating institutions. Patients were excluded from the study if they had a history of primary BC or distant metastasis, poor-quality images, unavailable ALN pathological results, or incomplete clinical and imaging data. The detailed patient selection workflow is illustrated in Figure S1.

Clinicopathologic data and reference standards

Relevant clinical and pathological information was collected from electronic medical records, including patient age, histologic type and grade of the tumor, molecular subtype, MRI-reported ALN status, Breast Imaging Reporting and Data System (BI-RADS) category, and tumor size. For further details, see the Appendix 1.

For the assessment of axillary nodal status, histopathologic data from either SLNB or ALND were adopted as the reference standard. Lymph nodes were grouped into benign, isolated tumor cells (≤0.2 mm), micrometastases (>0.2 and ≤2 mm), and macrometastases (>2 mm). In this study, micrometastases and macrometastases were both categorized as positive for ALN metastasis, with isolated tumor cells categorized as negative.

MRI image processing and segmentation

All participants underwent breast MRI using 3.0-Tesla scanners; detailed scanning parameters are provided in Table S1. The second post-contrast phase of the DCE-MRI was selected for analysis. To address inconsistencies in spatial resolution across different scans, all images were resampled to an isotropic voxel dimension of 1×1×1 mm³ and subsequently standardized using Z-score normalization.

Regions of interest (ROIs) were manually segmented by consensus between two radiologists (L.S.M and L.Z, with 15 and 6 years of experience in breast MRI, respectively) using ITK-SNAP software (version 3.8.0). Both radiologists were blinded to the patients’ pathological findings throughout the segmentation process. To quantitatively assess interobserver segmentation reliability, 50 cases were randomly selected from the total dataset, and both radiologists independently performed ROI segmentation on these cases without access to pathological information. Interobserver agreement was assessed using the Dice similarity coefficient (DSC). The pooled mean DSC was 0.856 [95% confidence interval (CI): 0.8495–0.8689]. The finalized ROIs were expanded by a 5-mm peripheral margin to capture the peritumoral region, consistent with its established prognostic significance in previous literature (7). Non-tumoral structures were excluded to facilitate subsequent analyses.

Based on the maximum cross-sectional area of the tumor, the corresponding bounding cuboid of the tumor region was extracted. Using the OKT-crop_max_roi tool (https://github.com/OnekeyAI-Platform/onekey), a total of 12 images across distinct layers were subsequently generated. To comprehensively characterize tumor heterogeneity from multiple anatomical perspectives, four specialized imaging modules were developed: multiplanar imaging (MPI), coronal-plane imaging (CPI), sagittal-plane imaging (SPI), and axial-plane imaging (API) (Figure 1 and Appendix 1). All image sets were uniformly resized to 256 ×256 pixels. For each module, three relevant slices were then concatenated along the channel axis to produce standardized 256×256×3 pseudo-red-green-blue inputs suitable for DL models.

Figure 1 Overall flowchart of model development and assessment. API, axial-plane imaging; CPI, coronal-plane imaging; MPI, multiplanar imaging; SPI, sagittal-plane imaging.

Multichannel DL model construction and feature extraction

The ResNet101 framework was employed to train the multichannel DL model. During model training, the Stochastic Gradient Descent (SGD) optimizer was used with a starting learning rate of 0.01, and network parameters were initialized using pre-trained weights from ImageNet. Each training iteration employed a consistent batch size of 32, with balanced sampling maintained across batches. To mitigate overfitting, regularization methods, including dropout (rate =0.2) and L2 weight decay (λ=1×10⁻⁴), were applied. Data augmentation was performed using random cropping, horizontal flipping, and vertical flipping to improve model generalizability. Early stopping was implemented to halt training when validation performance ceased to improve. The trained models output the probability of ALN metastasis.

Upon completion of ResNet101 training, 2048-dimensional DL features were extracted from the penultimate average pooling layer for each imaging module. This process generated four distinct feature sets for subsequent analysis.

To enhance the interpretability of the multichannel DL framework, gradient-weighted class activation mapping (Grad-CAM) was employed (14,15). This technique produces localization maps that highlight the image regions most influential to the classification decision. By examining gradient information from the final convolutional layer of the last residual block, activation patterns associated with model predictions were visualized, enabling the identification of salient areas that contributed substantially to the classification results. Further, SHapley Additive exPlanations (SHAP) analysis was performed to quantitatively evaluate the interpretability of the proposed model (16).

Feature fusion model construction

After features had been extracted from every individual imaging module, the multiscale feature sets were fed into a transformer-based fusion architecture (Figure S2). This integrated framework incorporated a transformer encoder coupled with a multilayer perceptron (MLP) classification head, thereby facilitating a holistic evaluation of four distinct imaging modules; that is, MPI, CPI, SPI, and API. Notably, the fusion mechanism capitalized on the mutually complementary information inherent among these imaging sequences.

The DL models were developed and trained on the OnekeyAI platform (version 4.9.1). The implementation was based on Python 3.7.12 and the PyTorch 1.11.0 framework, with graphics processing unit acceleration provided by CUDA 11.3.1 and cuDNN 8.2.1.

Node-RADS assessment

All cases were selected exclusively from the external validation cohort for Node-RADS assessment. Two radiologists (L.S.M and L.Z.), blinded to the pathological results, independently evaluated the ALNs using the Node-RADS (8,10). Any discrepancies were resolved through consensus discussion. The maximum Node-RADS score for each patient was recorded and subsequently compared with the pathological reference standard to determine the presence of ALN metastasis. A 5-point scoring scale was adopted, on which a score of 1 indicated very low suspicion and a score of 5 indicated very high suspicion. The optimal Node-RADS score threshold for classifying ALN status was determined based on the maximum Youden index.

Statistical analysis

Categorical and continuous variables were analyzed using the chi-squared (χ²) test (or Fisher’s exact test, as appropriate) and the Mann-Whitney U test (or Student’s t-test), respectively. The diagnostic performance of each model was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Differences in AUC values between models were compared using the DeLong test. The performance of the transformer-based model was directly compared with that of the Node-RADS, and differences in paired sensitivity and specificity were evaluated using the McNemar test. A two-tailed P value <0.05 was considered statistically significant. All statistical analyses were performed using R (version 4.5.0).

Results

Patient characteristics

A total of 735 patients were included in this study, comprising 587 from center 1 and 148 from center 2. The 587 patients from center 1 were randomly divided into a training cohort (n=411) and an internal validation cohort (n=176) at an approximate ratio of 7:3. The combined center 1 cohort included 270 ALN-positive and 317 ALN-negative cases, with a median age of 50 [43, 56] years. The external validation cohort included 85 ALN-positive and 63 ALN-negative cases, with a median age of 50.5 [43, 56] years (Table 1).

Baseline demographic, clinical, and pathological characteristics were balanced between the combined training/internal validation cohort (n=587) and the external validation cohort (n=148) in relation to age, BI-RADS classification, tumor size, availability of MRI reports, estrogen receptor (ER) status, progesterone receptor (PR) status, histological subtype, Ki-67 index, and human epidermal growth factor receptor 2 (HER2) status (all P>0.05). In contrast, significant inter-cohort differences were observed in relation to tumor grade (P<0.001), with the external validation cohort exhibiting higher rates of high-grade tumors (34% vs. 16%) (Table 1).

Multi-DL model

Systematic comparison across the four modules demonstrated their moderate diagnostic performance. Specifically, the MPI model achieved AUCs of 0.826 (95% CI: 0.787–0.865), 0.785 (95% CI: 0.719–0.852), and 0.821 (95% CI: 0.751–0.891) in the training, internal validation, and external validation cohorts, respectively. The API model achieved corresponding AUC values of 0.814 (95% CI: 0.773–0.854), 0.788 (95% CI: 0.721–0.855), and 0.749 (95% CI: 0.671–0.828). The CPI model achieved corresponding AUC values of 0.756 (95% CI: 0.710–0.802), 0.710 (95% CI: 0.631–0.789), and 0.766 (95% CI: 0.685–0.846), while the SPI model achieved corresponding AUC values of 0.771 (95% CI: 0.726–0.816), 0.773 (95% CI: 0.704–0.843), and 0.815 (95% CI: 0.747–0.883). The predictive performance of these models was further supported by calibration plots, DeLong test comparisons, and confusion matrices, all of which are presented in Table 2, Figure 2, and Figures S3,S4.

Figure 2 Performance of multichannel DL models in the API, CPI, SPI, and MPI modules. (A-C) Receiver operating characteristic curves of the four modules in the training, internal validation, and external validation cohorts, respectively. (D-I) Corresponding DeLong test results and calibration curves. API, axial-plane imaging; CPI, coronal-plane imaging; DL, deep learning; MPI, multiplanar imaging; ROC, receiver operating characteristic; SPI, sagittal-plane imaging.

Fusion model

To alleviate information loss caused by single-scale representation and simultaneously reduce computational burden, features were derived from four multichannel imaging modules. These features were subsequently assigned dynamic weights across planes via a transformer architecture equipped with multi-head attention, thereby facilitating fusion and leading to improved predictive precision. The transformer fusion model exhibited more favorable diagnostic efficacy, with AUC values of 0.959 (95% CI: 0.942–0.976), 0.885 (95% CI: 0.839–0.931), and 0.908 (95% CI: 0.862–0.954) in the training, internal validation, and external validation cohorts, respectively (Table 2 and Figure 3). Diagnostic performance was assessed by confusion matrices across all cohorts (Figure S5). Subgroup analyses showed that patient age, BI-RADS category, and tumor size did not significantly affect the performance of the proposed model (Figure 3 and Figure S6).

Figure 3 Performance of the transformer model for ALN metastasis prediction. (A) ROC curves of the transformer model in different cohorts. (B-D) Comparisons of the ROC curves of the transformer model among patients with different ages (years) (B), BI-RADS categories (C), and tumor sizes (mm) (D) in the external validation cohort. The P value was calculated using the Delong test. ALN, axillary lymph node; AUC, area under the curve; BI-RADS, Breast Imaging Reporting and Data System; ROC, receiver operating characteristic.

Insights into the DL model

As illustrated in Figure 4A, our findings indicated that the Grad‑CAM heatmap effectively captured finegrained image characteristics in the identifiable tumor region of an ALN-positive case. This finding supports the reliability of multichannel DL in feature extraction. Figure S7A displays the corresponding visualization for an ALN-negative patient, enabling comparison of the model’s feature learning process. By visualizing the network’s response patterns, we gained insights into how DL models achieve accurate classification of ALN status in BC patients.

Figure 4 Visualization in a patient with ALN-positive status. ALN, axillary lymph node; API, axial-plane imaging; CPI, coronal-plane imaging; MPI, multiplanar imaging; SHAP, SHapley Additive exPlanations; SPI, sagittal-plane imaging.

The interpretability of the model was assessed using SHAP analysis applied to the ExtraTrees classifier, enabling quantitative evaluation of each feature’s influence on the predictive results. Further evaluation involved systematic analysis of SHAP values in the full cohort (Figure 4B), a representative ALN-positive case (Figure 4C,4D), and a representative ALN-negative case (Figures S7B,S7C), collectively illustrating the transparency of the predictive model.

Comparison with Node-RADS

A Node-RADS score > 2 was identified as the optimal cutoff for differentiating ALN-positive and ALN-negative status. The performance of the transformer model was compared with that of the Node-RADS using the external validation cohort. No statistically significant differences were observed between the transformer model and the Node-RADS in terms of the AUC (P=0.971), diagnostic accuracy (P=0.176), sensitivity (P=0.134), or specificity (P>0.99) (Table 3 and Figure 5). The weighted kappa coefficient was 0.523 (95% CI: 0.384–0.660), indicating moderate agreement between the two methods.

Figure 5 Performance comparison between the transformer model and Node-RADS for axillary lymph node metastasis prediction. Bar charts show the sensitivity, specificity, accuracy, and AUC for each method. AUC, area under the curve; Node-RADS, Node Reporting and Data System.

Discussion

This two-center retrospective study presented a multichannel, multiscale DL model for the preoperative prediction of ALN metastasis in BC patients using DCE-MRI. The proposed model achieved excellent diagnostic performance, with AUC values of 0.959 (95% CI: 0.942–0.976) in the training cohort, 0.885 (95% CI: 0.839–0.931) in the internal validation cohort, and 0.908 (95% CI: 0.862–0.954) in the external validation cohort. Notably, subgroup analyses showed that the model’s performance was not significantly influenced by patient age, BI-RADS classification, or tumor size, supporting its robustness across diverse patient populations. Further, the model demonstrated diagnostic accuracy comparable to that of the Node-RADS, highlighting its potential as a reliable clinical decision-making tool.

Several clinical models for predicting ALN metastasis have been reported in previous studies. Isaac et al. (17) introduced a multiparametric scoring system aimed at evaluating non-sentinel lymph node (NSLN) status in patients with clinically node-negative BC who had positive sentinel lymph nodes (SLNs) after systemic therapy. Meanwhile, a widely recognized nomogram developed by Memorial Sloan Kettering Cancer Center has been frequently used across diverse populations (18). However, its clinical utility may be limited by its reliance on postoperative parameters, such as histological tumor size and lymphovascular invasion. In contrast, in this study, all features incorporated into the proposed model were extracted from preoperative DCE-MRI, and the model maintained robust predictive accuracy in the external validation cohort.

Previous studies have demonstrated the value of DL models in predicting ALN metastasis. Gao et al. (19) developed a DL-based approach to predict overall ALN metastasis by integrating imaging data from both the primary tumor and lymph nodes, achieving an AUC of 0.852 in the external validation cohort. However, a limitation of the study by Gao et al. was that conventional DCE-MRI scans may fail to encompass all ALNs, which may compromise the generalizability of ALN metastasis prediction based on ALN images. Chen et al. (20) constructed an MRI-based DL model to identify the SLN and NSLN status of BC patients. However, the model demonstrated variable performance across external validation sets, with AUC values ranging from 0.728 to 0.885. Such substantial variability suggests that the generalizability of the model may be inconsistent across different external populations.

In the present study, the proposed model demonstrated excellent diagnostic performance in both internal and external validation cohorts. The superior performance of our model may be attributed to several key design features. First, the integration of MPI modules (MPI, CPI, SPI, and API) enables the comprehensive capture of tumor spatial heterogeneity from multiple anatomical perspectives. Second, the transformer-based feature fusion architecture plays a pivotal role in synthesizing complementary information across different imaging modules. By assigning dynamic weights to each module via multi-head attention, the model effectively prioritizes clinically relevant features and reduces information redundancy, addressing a major limitation of traditional feature concatenation methods (21,22). This fusion strategy not only enhances predictive accuracy but also improves model generalizability, as evidenced by its consistent performance in the external validation cohort.

Another important strength of our study is the emphasis on model interpretability. Through Grad-CAM visualization, we identified the image regions most influential to the model’s predictions, providing insights into the biological basis of ALN metastasis (14,15). Additionally, SHAP analysis quantitatively elucidated the contribution of individual DL features to each prediction, enhancing transparency and facilitating clinical trust (16). This dual approach to interpretability addresses a major barrier to the adoption of DL in clinical oncology, where “black box” models are often viewed with skepticism (23).

The Node-RADS has been validated across multiple tumor studies (9,24-26). A previous study demonstrated that the Node-RADS effectively evaluated ALN status in BC patients without neoadjuvant therapy, with scores >2 indicating a high risk of metastasis (9). A multicenter study demonstrated that an MRI-based model integrating Node-RADS, DL, and radiomics features achieved superior predictive performance for ALN metastasis following neoadjuvant chemotherapy in BC compared with single-modality approaches (27). Motivated by these findings, we developed and validated a multichannel, multiscale DL model. Consistent with previous findings (28), our model achieved comparable AUC, sensitivity, specificity, and diagnostic accuracy to the Node-RADS. The moderate agreement between the two methods (kappa =0.523) suggests that they may capture complementary aspects of ALN pathology. While the Node-RADS relies on radiologists’ interpretation of lymph node morphology, our model leverages tumor-related imaging features to predict metastatic risk, offering an alternative or adjunctive tool in cases in which Node-RADS results are inconclusive. This approach may be particularly valuable in clinical settings with limited access to experienced radiologists or for patients with ambiguous lymph node findings.

Our study had several limitations that warrant consideration. First, the retrospective design may have introduced selection bias, despite consistent enrollment criteria across the two centers. Second, the external validation cohort was derived from a single institution, and future studies should validate the model in more diverse, multicenter populations to further confirm its generalizability. Third, we focused exclusively on DCE-MRI, and future research should explore the integration of other imaging sequences (e.g., diffusion-weighted imaging and T2-weighted imaging) to further improve predictive performance (29). Fourth, the model was trained on patients who did not receive neoadjuvant chemotherapy, and its applicability in this specific subgroup requires further investigation. Fifth, the external validation cohort showed a statistically higher rate of high-grade tumors. The uneven tumor grade distribution across cohorts may have influenced external validation performance, and future balanced multicenter studies are needed to further improve model generalizability.

Conclusions

Our multichannel, multiscale DL model demonstrated excellent performance in the preoperative prediction of ALN metastasis in BC, with robust performance across subgroup analyses and diagnostic accuracy comparable to that of the Node-RADS. The model’s interpretability, facilitated by Grad-CAM and SHAP analysis, further enhances its clinical applicability. By providing accurate, objective, and interpretable predictions of ALN metastasis, this model has the potential to optimize treatment decision-making, reduce unnecessary ALND, and improve patient outcomes. Future prospective studies are warranted to validate these findings and explore the integration of additional data modalities to further enhance model performance.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0851/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0851/dss

Funding: This work was supported by the Science and Technology Research Project of Henan Province (No. 252102310061), Key Scientific Research Projects of Higher Education Institutions of Henan Province (Nos. 25B320031, 26B320017), Science and Technology Program of Shangqiu City (Nos. 2024075, 2025068), and Doctoral Scientific Research Initiation Project of Shangqiu Medical College (No. BSJH004). Additional support was provided by the Henan Provincial Health Commission Key Laboratory of Imaging Precision Diagnosis.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0851/coif). All authors report the funding from the Science and Technology Research Project of Henan Province (No. 252102310061), Key Scientific Research Projects of Higher Education Institutions of Henan Province (Nos. 25B320031, 26B320017), Science and Technology Program of Shangqiu City (Nos. 2024075, 2025068), and Doctoral Scientific Research Initiation Project of Shangqiu Medical College (No. BSJH004). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional review boards of The Third Affiliated Hospital of Zhengzhou University (No. 202223901) and The First Affiliated Hospital of Henan University of Science and Technology (No. 2025-03-K0026), and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin 2023;73:17-48. [Crossref] [PubMed]
2. Jatoi I, Hilsenbeck SG, Clark GM, Osborne CK. Significance of axillary lymph node metastasis in primary breast cancer. J Clin Oncol 1999;17:2334-40. [Crossref] [PubMed]
3. Mamounas EP, Kuehn T, Rutgers EJT, von Minckwitz G. Current approach of the axilla in patients with early-stage breast cancer. Lancet 2017;S0140-6736(17)31451-4.
4. Yu Y, Tan Y, Xie C, Hu Q, Ouyang J, Chen Y, et al. Development and Validation of a Preoperative Magnetic Resonance Imaging Radiomics-Based Signature to Predict Axillary Lymph Node Metastasis and Disease-Free Survival in Patients With Early-Stage Breast Cancer. JAMA Netw Open 2020;3:e2028086. [Crossref] [PubMed]
5. Langer I, Guller U, Berclaz G, Koechli OR, Schaer G, Fehr MK, Hess T, Oertli D, Bronz L, Schnarwyler B, Wight E, Uehlinger U, Infanger E, Burger D, Zuber M. Morbidity of sentinel lymph node biopsy (SLN) alone versus SLN and completion axillary lymph node dissection after breast cancer surgery: a prospective Swiss multicenter study on 659 patients. Ann Surg 2007;245:452-61. [Crossref] [PubMed]
6. Mann RM, Cho N, Moy L. Breast MRI: State of the Art. Radiology 2019;292:520-36. [Crossref] [PubMed]
7. Liu Z, Hong M, Li X, Lin L, Tan X, Liu Y. Predicting axillary lymph node metastasis in breast cancer patients: A radiomics-based multicenter approach with interpretability analysis. Eur J Radiol 2024;176:111522. [Crossref] [PubMed]
8. Elsholtz FHJ, Asbach P, Haas M, Becker M, Beets-Tan RGH, Thoeny HC, Padhani AR, Hamm B. Introducing the Node Reporting and Data System 1.0 (Node-RADS): a concept for standardized assessment of lymph nodes in cancer. Eur Radiol 2021;31:6116-24. [Crossref] [PubMed]
9. Pediconi F, Maroncelli R, Pasculli M, Galati F, Moffa G, Marra A, Polistena A, Rizzo V. Performance of MRI for standardized lymph nodes assessment in breast cancer: are we ready for Node-RADS? Eur Radiol 2024;34:7734-45. [Crossref] [PubMed]
10. Kim HJ, Chae EY, Eom HJ, Choi WJ, Shin HJ, Cha JH, Kim HH. Node Reporting and Data System Evaluation of Axillary Nodes in Invasive Ductal and Lobular Carcinoma. Radiology 2025;316:e243823. [Crossref] [PubMed]
11. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. [Crossref] [PubMed]
12. Li X, Yang L, Jiao X. Comparison of Traditional Radiomics, Deep Learning Radiomics and Fusion Methods for Axillary Lymph Node Metastasis Prediction in Breast Cancer. Acad Radiol 2023;30:1281-7. [Crossref] [PubMed]
13. Chen Y, Wang L, Dong X, Luo R, Ge Y, Liu H, Zhang Y, Wang D. Deep Learning Radiomics of Preoperative Breast MRI for Prediction of Axillary Lymph Node Metastasis in Breast Cancer. J Digit Imaging 2023;36:1323-31. [Crossref] [PubMed]
14. Niu W, Yan J, Hao M, Zhang Y, Li T, Liu C, Li Q, Liu Z, Su Y, Peng B, Tan Y, Wang X, Wang L, Zhang H, Yang G. MRI transformer deep learning and radiomics for predicting IDH wild type TERT promoter mutant gliomas. NPJ Precis Oncol 2025;9:89. [Crossref] [PubMed]
15. He J, Xu J, Chen W, Cao M, Zhang J, Yang Q, Li E, Zhang R, Tong Y, Zhang Y, Gao C, Zhao Q, Xu Z, Wang L, Cheng X, Zheng G, Pan S, Hu C. Development of a deep learning model for T1N0 gastric cancer diagnosis using 2.5D radiomic data in preoperative CT images. NPJ Precis Oncol 2025;9:249. [Crossref] [PubMed]
16. Zhu J, Xu B, Fan T, Ji S, Gu K, Ding J, Lu H, Ma J, Zhou Y. Sub-regional radiomics combining multichannel 2-dimensional or 3-dimensional deep learning for predicting neoadjuvant chemo-immunotherapy response in esophageal squamous cell carcinoma: a multicenter study. NPJ Precis Oncol 2025;9:248. [Crossref] [PubMed]
17. Cebrecos I, Mension E, Alonso I, Castillo H, Sanfeliu E, Vidal-Sicart S, Ganau S, Vidal M, Schettini F. Nonsentinel Axillary Lymph Node Status in Clinically Node-Negative Early Breast Cancer After Primary Systemic Therapy and Positive Sentinel Lymph Node: A Predictive Model Proposal. Ann Surg Oncol 2023;30:4657-68. [Crossref] [PubMed]
18. Bevilacqua JL, Kattan MW, Fey JV, Cody HS 3rd, Borgen PI, Van Zee KJ. Doctor, what are my chances of having a positive sentinel node? A validated nomogram for risk estimation. J Clin Oncol 2007;25:3670-9.
19. Gao J, Zhong X, Li W, Li Q, Shao H, Wang Z, Dai Y, Ma H, Shi Y, Zhang H, Duan S, Zhang K, Yang P, Zhao F, Zhang H, Xie H, Mao N. Attention-based Deep Learning for the Preoperative Differentiation of Axillary Lymph Node Metastasis in Breast Cancer on DCE-MRI. J Magn Reson Imaging 2023;57:1842-53. [Crossref] [PubMed]
20. Chen M, Kong C, Lin G, Chen W, Guo X, Chen Y, Cheng X, Chen M, Shi C, Xu M, Sun J, Lu C, Ji J. Development and validation of convolutional neural network-based model to predict the risk of sentinel or non-sentinel lymph node metastasis in patients with breast cancer: a machine learning study. EClinicalMedicine 2023;63:102176. [Crossref] [PubMed]
21. Wang W, Liang H, Zhang Z, Xu C, Wei D, Li W, Qian Y, Zhang L, Liu J, Lei D. Comparing three-dimensional and two-dimensional deep-learning, radiomics, and fusion models for predicting occult lymph node metastasis in laryngeal squamous cell carcinoma based on CT imaging: a multicentre, retrospective, diagnostic study. EClinicalMedicine 2024;67:102385. [Crossref] [PubMed]
22. Wagner SJ, Reisenbüchler D, West NP, Niehues JM, Zhu J, Foersch S, et al. Transformer-based biomarker prediction from colorectal cancer histology: A large-scale multicentric study. Cancer Cell 2023;41:1650-1661.e4. [Crossref] [PubMed]
23. Qamar T, Bawany NZ. Understanding the black-box: towards interpretable and reliable deep learning models. PeerJ Comput Sci 2023;9:e1629. [Crossref] [PubMed]
24. Bai X, Peng C, Liu B, Zhou S, Liu H, Chen Y, Guo H, Hao Y, Liu X, Zhao J, Ding X, Li L, Zhang X, Ye H, Ma X, Wang H. Diagnostic and prognostic value of MRI-based Node-RADS for the assessment of regional lymph node metastasis in renal cell carcinoma. Diagn Interv Imaging 2025;106:117-25. [Crossref] [PubMed]
25. Fang Y, Chen M, Zheng X, Yao Y, Huang K, Chen S, Xu T, Xu Z, Lin D. Validation of the Node Reporting and Data System (Node-RADS) for standardized CT evaluation of regional lymph nodes in esophageal squamous cell carcinoma patients. Eur Radiol 2025;35:2999-3009. [Crossref] [PubMed]
26. Riccardi S, Ninkova RV, Calabrese A, Curti F, Gennarini M, Miceli V, Cupertino A, Cutonilli C, Di Donato V, Pernazza A, Rizzo SMR, Panebianco V, Catalano C, Manganaro L. The Node Reporting and Data System (Node-RADS) for standardized MRI evaluation of lymph nodes in endometrial cancer, integrated with clinicopathological and molecular data. Eur J Radiol 2025;187:112079. [Crossref] [PubMed]
27. Lai X, He H, Liang B, Xu Z, Yang L, Wu T, Han K, Li W, Liu Q, Zhu C, Zhao R, Cai G, Dong H, Yang Y. Interpretable MRI-Based Machine Learning Model for Noninvasive Prediction of Axillary Lymph Node Metastasis After Neoadjuvant Chemotherapy in Breast Cancer. Acad Radiol 2026;33:1454-65. [Crossref] [PubMed]
28. Maroncelli R, Rizzo V, Pasculli M, Coppola S, De Nardo C, Moschetta M, Catalano C, Pediconi F. Radiomics-based MRI models for predicting breast cancer axillary lymph node involvement in comparison with Node-RADS: a proof-of-concept study. Eur Radiol Exp 2025;9:117. [Crossref] [PubMed]
29. Wang Z, Sun H, Li J, Chen J, Meng F, Li H, Han L, Zhou S, Yu T. Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer Using CNN Based on Multiparametric MRI. J Magn Reson Imaging 2022;56:700-9. [Crossref] [PubMed]