Assessing the diagnostic value of hemodynamic distinctions between axillary lymph nodes and adjacent vessels in breast cancer axillary lymph node metastasis via breast magnetic resonance imaging.

Introduction
Breast cancer is a malignant tumor that often occurs in the breast epithelium. Its incidence rate ranks first among women, and its mortality ranks second (1). One of the main mechanisms of breast cancer metastasis is through the lymphatic system. The earliest and most common site of metastasis is the axillary lymph nodes. Whether there is axillary lymph node metastasis (ALNM) is of great value for selecting surgical methods, the formulation of adjuvant therapy, and the prognosis evaluation (2). The main clinical strategies for detecting ALNM in breast cancer are axillary lymph node dissection and sentinel lymph node biopsy (SLNB) (3-7). Although SLNB has been proven to have more advantages than lymph node dissection (8-10) with low false-negative rates (11), these invasive examination methods may lead to many complications such as lymphedema, pain, numbness, and dyskinesia (12,13). In addition, since a physical examination may be a non-invasive and cost-effective way to detect metastasis, doctors manually touch patients with their hands to examine the affected area (14). However, the accuracy of such physical examination is poor (15). Therefore, establishing a prediction model for ALNM in breast cancer based on medical images has an excellent clinical application prospect, which can reduce unnecessary pain caused by surgery (16).
In recent years, researchers have conducted a lot of research and exploration in predicting ALNM in breast cancer based on medical images. Sidibé et al. predicted ALNM in breast cancer using ultrasound images. The experimental results show that the sensitivity of the prediction model is up to 87.1%, and the specificity is up to 95.2% (16). Magnetic resonance imaging (MRI) has been widely used in examining breast cancer. Still, at this stage, doctors may misjudge lymph node metastasis only by observing MRI images, so it is impossible to predict lymph node metastasis only by the doctors’ observation of MRI in the clinic.
With the rapid rise of radiomics and artificial intelligence methods, researchers have been researching ALNM prediction in breast cancer (17). However, the prevailing prediction approaches predominantly rely on analyzing characteristics specific to the tumor area (18-22). Liu et al. predicted ALNM in breast cancer by analyzing the tumor radiomics features of breast MRI and combining them with clinical-pathological characteristics: the area under the receiver operating characteristic curve (AUC) in the independent validation set was 0.869 (18). Arefan et al. extracted the 2D and 3D quantitative radiomics features of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). They used machine learning classifiers to compare the performance of prediction models established in 2D and 3D. The AUC values were 0.81 and 0.82, and the accuracy was 0.79 and 0.80, respectively (22). Ren et al. (23) constructed a convolutional neural network (CNN) model based on MRI images with a prediction accuracy of 0.848±0.024 and an AUC value of 0.91±0.02. By constructing the CNN, Sun et al. used breast ultrasound tumor images to predict ALNM. The AUC was 0.912, with an accuracy of 0.893 (19). Zhou et al. used ultrasonic images to build and compare three different CNN models, namely, Inception V3, Inception ResNet V2, and ResNet-101, and carried out prediction experiments in independent test sets. The optimal prediction performance was achieved using the Inception V3 model with AUC =0.89, sensitivity =0.85, and specificity =0.73, respectively, which is better than the prediction results of radiologists (sensitivity is 0.73, specificity is 0.63) (24). Zhang et al. (25) conducted a comprehensive review of 13 studies involving 1,618 participants to evaluate the diagnostic performance of machine learning-based DCE-MRI imaging in predicting ALNM in breast cancer, achieving an AUC of 0.89. Mao et al. (26) constructed a nomogram model by integrating DCE-MRI radiomics features and independent risk factors to predict breast cancer ALNM, reaching an AUC of 0.90±0.05 in the test set. Similarly, Song et al. (27) used multivariate logistic regression analysis to establish a nomogram model based on radiomics features and clinical factors, with AUCs of 0.907 and 0.874 in the training and validation sets, respectively. Chen et al. (28) applied logistic regression to develop genomic, radiomic, and radiogenomic models to predict breast cancer ALNM, with the fusion model achieving a maximum AUC of 0.84. Zhang et al. (29) utilized the ResNet50 network with an ensemble approach using weighted voting, constructing a multi-parameter MRI model from T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and DCE-MRI images, which achieved an AUC of 0.913.
Although there have been some studies in predicting ALNM in breast cancer patients, their application in clinical practice is still limited due to constraints in prediction accuracy and robustness. To better address or overcome this challenge, we propose investigating a new approach in light of the existing research. The physiological rationale for investigating lymph node-vessel hemodynamics is grounded in the microcirculatory changes associated with metastasis. Metastatic involvement of a lymph node often induces significant neoangiogenesis to support tumor growth, leading to an increased microvessel density and altered vascular permeability. Concurrently, the normal lymphatic architecture and function can be disrupted, potentially reducing normal lymphatic clearance of interstitial fluid and contrast agent. We hypothesize that these combined effects—increased, disorganized blood flow and potential retention of contrast agent—would alter the time-intensity curve (TIC) profile of a metastatic lymph node, making its hemodynamic signature more closely resemble that of a high-flow blood vessel (characterized by rapid, high-amplitude enhancement) compared to a normal or reactive node. Therefore, quantifying the hemodynamic disparity (or similarity) between a lymph node and its adjacent reference vessel could serve as a non-invasive biomarker for metastatic involvement. Accordingly, we propose investigating a new approach in light of the existing research. First, we identify and extract more clinically relevant bio-imaging features or markers from the TIC of lymph nodes and adjacent axillary vessels generated from DCE-MRI. Next, we quantitatively calculate the difference between the TIC curves generated from the segmented axillary vessels and lymph nodes. Last, we compare and analyze the contribution of the identified features integrated into an existing prediction model (a classification model based on deep image features and clinical information) to optimally predict lymph node metastasis. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-47/rc).

Methods

Participants
This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Medical Ethics Committee of Liaoning Cancer Hospital (No. 20180927). Written informed consent was obtained from all participants before their inclusion in the study. From the cancer registry database, clinical data of 186 MRI image series of breast cancer patients who were treated in Liaoning Cancer Hospital were retrospectively identified and retrieved. According to the pathological results, the data were divided into 92 cases of ALNM and 94 cases of non-ALNM. Table 1 shows the specific statistical information of patients in this study. The age, tumor diameter, and age at menarche of the study subjects are continuous variables, which were compared using t-tests between the metastatic and non-metastatic groups. The histological grading is an ordinal variable, which was compared using the Wilcoxon rank-sum test. The side of the lesion, tumor location, multifocal, magnetic resonance (MR) assessment of axillary lymph node (LN), whether the skin was thickened, whether the nipple was depressed, blood type, axillary touch (defined as the pre-operative physical examination finding where the clinician palpates enlarged or hardened lymph nodes in the axilla), vascular tumor thrombus, nerve invasion, pathological type, estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), Ki-67, molecular typing, family history of breast cancer, menopause status, and history of lactation are classified variables, were compared by Chi-squared test.
Table 1 shows the statistical results of the aforementioned clinical features in the ALNM and non-ALNM groups. It can be seen that there were significant differences (P<0.05) between the metastatic group and the non-metastatic group in age, tumor location, multifocal, MR assessment of axillary LN, thickening of the skin, axillary touch, vascular cancer thrombus, pathological type, family history of breast cancer, and menopausal status. Therefore, these clinical characteristics will be used to predict the ALNM in breast cancer.

MRI protocol
MRI images used in this study were obtained from a GE 1.5 T MRI (HDX, GE Healthcare, Waukesha, WI, USA), using four-channel coils dedicated to the breast. All patients were prone to natural prolapse of both breasts. MRI scanning parameters were as follows: axial sequence (a 3D T1 weighted imaging technology, which could perform routine or dynamic enhanced scanning of bilateral breasts to obtain axial breast images with high signal-to-noise ratio and high resolution), repetition time/echo time: 6.2/3.0 ms; flip angle: 10; slice thickness: 3.2 mm; 48 slices per phase; field of view: 360 mm × 360 mm; matrix size: 256×256. During DCE-MRI image acquisition, a contrast agent (0.1 mmol·kg GdDTPA-MBA, Omniscan, GE Healthcare) was injected intravenously at the rate of 3 mL/s and then rinsed with 20 mL of normal saline. After intravenous injection, continuous scanning was performed in 8 phases, with each scanning time of 43 s and a total scanning time of 5.7 min. All breast MRI examinations included a complete axillary field of view. Two radiologists (with 5 and 10 years of work experience) reviewed all images.

Segmentation of axillary vessels and lymph nodes on the affected side
For each patient, the target lymph node for subsequent analysis was first identified by an experienced radiologist based on the following morphological criteria on DCE-MRI: cortical thickening (>3 mm), loss of the fatty hilum, and a round shape. If no lymph nodes met these criteria, the largest node by short-axis diameter was selected. The region of interest (ROI) of this target lymph node was then precisely segmented using the following automated algorithm.
The axillary region of the affected breast was located first, and then the three-dimensional Hessian matrix of each pixel in the axillary region was calculated. Finally, the ROI regions of the axillary vessels and axillary lymph nodes were extracted by determining the relationship among the eigenvalues corresponding to the three eigenvectors of the matrix. Figure 1 shows the enlarged axillary region of the affected side, where axillary lymph nodes and blood vessels can be seen.

Axillary vessel segmentation
This study used a vascular enhancement method based on the Hessian matrix to extract the axillary vessels accurately. The 3D Hessian matrix was generated by calculating the local second derivative, and the eigenvalues corresponding to the three eigenvectors were calculated. The vascular structure can be enhanced according to the relationship between the eigenvalues, and the vascular ROI can be obtained. Because the axillary vessels are relatively thin, the parameter sigma of the Gaussian kernel used in this study was 1. The specific axillary vessel extraction method is as follows: calculate the Hessian matrix corresponding to each voxel in the axillary region and the eigenvalues corresponding to the three eigenvectors, and sort the eigenvalues according to the absolute value:
When a voxel belongs to the vascular region, its eigenvalues will have the following characteristics: Since the axillary vessel and lymph node are both bright areas, λ2 and λ3 are less than 0. The absolute value of λ1 is minimal, which should be 0 under ideal conditions. λ2 and λ3 are close and far greater than λ1. To sum up, it can be defined as the following formula:
This represents an ideal vessel shape. To accurately extract the vessel region, it is still necessary to use the maximum likelihood method to confirm the area:
Here α, β, and γ are the parameters to control the sensitivity. After testing, the values of α and β were selected as 0.5, and γ was selected as 15, where:
Figure 2 shows some examples of the axillary vessel segmentation results.

Axillary lymph node segmentation
When segmenting the lymph node, we changed the Eq. [4] to the following:
Compared with axillary vessels, lymph nodes may be more prominent in volume. The preliminary segmentation region of lymph nodes can be obtained from Eq. [8], which often captures the core of the node. To ensure the complete 3D volume of the lymph node was captured for subsequent TIC analysis, the segmentation results from Eq. [8] were used as a seed. From this seed, the segmentation was expanded to include adjacent slices above and below (i.e., along the z-axis) within the 3D image volume. A 3D region growing algorithm was then conducted, using the brightest point within this expanded region as the new seed point, to segment the entire lymph node volume. In this study, the range of 3D region growing is defined as:
Where I is the pre-segmented image, and c is an adjustable coefficient, which can be adjusted according to the characteristics of the image to obtain the best segmentation result. Here, we set c to 0.1.
After the region grew, the hole-filling method was adopted to obtain an entire lymph node region. Figure 3 shows the lymph node segmentation results.

TIC analysis

Generating two TIC curves
To draw the TIC curves, we used the above method to extract the axillary vessels and lymph nodes from the images of each phase. To reduce the impact of segmentation errors on the generated TIC curve, we took the pixels with signal intensity in the top 80% of the segmented area, calculated the average signal intensity for these pixels, and then drew the corresponding TICs. Figure 4 shows the two curves. The horizontal axis represents the scanned phases, and the vertical axis represents signal intensity after subtracting the baseline.

Statistical analysis
The MedCalc software (RRID: SCR_015044) was used to perform the statistical analysis in this study.

Effectiveness analysis of TIC characteristics in the ALNM prediction
Based on the deep learning method, the effectiveness analysis of the proposed TIC characteristics in predicting ALNM in breast cancer is performed, as shown in Figure 5. Firstly, the DCE ROI images of the patient were input into the ResNet50 network for training. A prediction model for ALNM based on a single DCE image was generated. Then, the soft voting method based on the single DCE image result is used to confirm the prediction results of the corresponding case. Finally, the ensemble learning algorithm with the Stacking framework method was used to integrate the case prediction results with the proposed TIC characteristic and related clinical characteristics (tumor location, multifocal, MR evaluation of axillary LN, skin thickening, axillary location, vascular tumor thrombus, pathological type, family history of breast cancer, menopausal status). In this study, the ensemble model selected the support vector machine (SVM) with radial basis function (RBF) kernel and nomogram as the ensemble classifier.
In the image-based prediction of ALNM, this study selected several images with prominent tumor areas at the exact coordinates of each patient during the pre-enhancement phase, peak phase, and later phase, corresponding to the 1st, 3rd, and 6th phase images, respectively, to perform RGB superposition, forming three-channel images. In total, 1,243 ROIs were obtained, with 55.8% metastatic cases and 44.2% non-metastatic cases. The dataset is divided into a training set and a test set. Firstly, randomly select 10% of metastatic and 10% of non-metastatic cases as an independent test set to evaluate the model’s effectiveness. Then, the remaining 90% of all cases in the dataset were randomly divided into five groups, on average, in which the training and validation data were composed in a 4:1 ratio for a 5-fold cross-validation.
ResNet50 was used as the initial network model for predicting ALNM in ROI images. In the ResNet50 network, a CNN with a depth of 50 layers was constructed using a stack of 16 residual blocks. Therefore, this network has a more robust feature extraction ability and higher recognition accuracy. In the experiment, the Keras library of TensorFlow was used as the backend, with TensorFlow version 2.2.0 and PyCharm software used for the experiment. Transfer learning is used to solve the problem of insufficient training samples in order to achieve better performance. The pre-trained weights are obtained on the ImageNet dataset and are transferred to fine-tune our prediction model. The parameter settings during the training process are as follows: the batch size is set to 20, the number of training epochs is set to 500, the Adam optimizer is selected, and the learning rate is 0.00001.
In addition to deep imaging features, this study also incorporated clinical features and the proposed TIC feature to investigate changes in overall prediction performance and the contribution of the TIC feature. In building and validating the ALNM prediction model, this study used the SVM classifier and employed a nomogram based on binary logistic regression analysis for prediction, validation, and evaluation.

Gold standard: surgical and pathological evaluation
The axillary nodal status (ALNM vs. non-ALNM) for each patient was definitively determined by histopathological examination of surgically resected lymph nodes. The surgical approach [SLNB or axillary lymph node dissection (ALND)] followed standard clinical guidelines. SLNB was performed using a combined technique of radiotracer and blue dye. Pathological processing included serial sectioning and hematoxylin and eosin (H&E) staining of sentinel nodes, supplemented by immunohistochemistry (IHC) when necessary. Nodes from ALND specimens were evaluated with routine H&E staining. A patient was classified as ALNM-positive if any lymph node contained metastatic deposits (macro- or micro-metastases >0.2 mm). The presence of isolated tumor cells (ITCs) only was classified as negative. The final pathological report served as the gold standard for patient grouping.

Results

TIC characteristic analysis results
First, the lymph-node TIC area (LNTICA) was calculated. Theoretically, the contrast agent flow in the blood vessel is more extensive, and the signal intensity is high. Therefore, DIFF was obtained by subtracting the lymph node TIC area from the vessel TIC area. However, this TIC area difference sometimes has a negative number. Here, we set all the negative area differences to 0, obtain the non-negative area differences (NON-NEG DIFF), and calculate the ratio of area difference to lymph node TIC area (RATIO TO LYMPH). Figure 6 compares characteristic data between the metastatic and non-metastatic groups.
Based on the data above, we performed the statistical analysis using MedCalc. We conducted the arithmetic mean, 95% CI for the mean, variance, standard deviation, and standard error of the mean, and an independent variable F-test, t-test, and Kolmogorov-Smirnov test for LNTICA, DIFF, NON-NEG DIFF, and RATIO TO LYMPH four characteristics in the metastatic group and non-metastatic group, respectively. Table 2 shows the results.

Effectiveness analysis results of TIC characteristics in the ALNM prediction
In addition to using DCE imaging alone for ALNM prediction, this paper employs SVM and nomogram as ensemble classifiers, investigating incorporating clinical features and the previously mentioned TIC curve characteristics to obtain the prediction results, as shown in Table 3 and Figures 7-9.
Table 3 shows that the TIC curve features proposed in this paper contribute positively to the prediction of ALNM regardless of whether SVM or nomogram is used. On the testing set, the AUC increased by 0.038 and 0.039, respectively. Notably, the nomogram model incorporating the TIC characteristic demonstrated excellent performance on the independent test set (AUC 0.941), suggesting good generalizability within our cohort and the potential robustness of the added hemodynamic feature.
To interpret the contribution of the integrated features in the final nomogram model, we performed a SHapley Additive exPlanations (SHAP) analysis. The resulting summary plot (Figure 10) illustrates the mean absolute SHAP value for each feature, representing its overall importance. The plot confirms that the proposed ‘RATIO TO LYMPH’ TIC characteristic was among the top contributors to the model’s predictions, underscoring its significant diagnostic value alongside key clinical predictors such as ‘MR assessment of axillary LN’ and deep learning image features.

Discussion
In the metastasis group data we analyzed, most of the curves have the following characteristics: the TIC of the lymph node ROI is very close to that of the axillary vessel ROI, as shown in Figure 11A, which indicates that the blood supply in the metastatic axillary lymph node is rich, making its hemodynamic characteristics more similar to the axillary vessels. Most of the curves in the non-metastatic group show a significant area difference between the two TICs, as seen in Figure 11B. However, this experiment also found that in some patients’ data, the TIC area of the axillary lymph node ROI would be larger than that of the axillary vascular ROI, which is more common in the metastatic group. See Figure 11C. The reason may be that compared with lymph nodes, the volume of the vessel ROI is minimal, so the imaging process is more affected by the volume effect, leading to a decrease in the vessel ROI signal value.
If the hemodynamic characteristics of the axillary lymph node alone were used as a distinguishing feature, distinguishing metastatic from non-metastatic lymph nodes would be challenging. Figure 6A shows the distribution of this feature in the two data control groups. It can be seen that the two groups of data are relatively close, both in terms of mean value and data range. Suppose the area difference of TIC curves is obtained by taking the axillary vessel TIC as a reference and taking it as the characteristic. In that case, the distinguishability of the two data groups improved, as shown in Figure 6B. Considering that the negative TIC area difference can be approximately considered as the hemodynamic characteristics of the two groups are similar, the distinguishability of the two groups of data is further improved after the processing of setting both negative values to 0, as shown in Figure 6C. Divide the above results by the TIC area of the lymph node, and the ratio obtained has the best discriminability. See Figure 6D.
It is important to note that while no patients with a primary diagnosis of lymphadenitis were present in our cohort, the non-ALNM group may have included lymph nodes with reactive benign changes. The significant hemodynamic differences observed suggest that our proposed feature may aid in distinguishing metastasis from such benign processes.
Our findings align with and build upon recent advances in predicting ALNM using AI. For instance, Zhang et al. (29) achieved an AUC of 0.913 using a multi-parameter MRI model, while Mao et al. (26) and Song et al. (27) reported AUCs of 0.90 and 0.907, respectively, using nomograms based on tumor radiomics and clinical factors. Our baseline model, integrating DCE image features and clinical characteristics, achieved comparable performance (AUC 0.902, nomogram). The key incremental benefit of our study is the introduction of the lymph node-vessel hemodynamic difference feature, which, when integrated, pushed the AUC to 0.941. This suggests that features capturing the functional interplay between lymph nodes and their local vascular environment provide complementary information beyond the tumor-centric or lymph node morphologic features typically used.
From Table 2, the F-test, t-test, and Kolmogorov-Smirnov test show that the last characteristics of the two control groups have the smallest P value. Meanwhile, the values of the Arithmetic mean, 95% confidence interval (CI) for the mean, variance, standard deviation, and standard error of the mean for the last characteristics of the two control groups are also the smallest. Notably, the nomogram model incorporating this TIC characteristic demonstrated excellent performance on the independent test set (AUC 0.941), suggesting good generalizability within our cohort and the potential robustness of the added hemodynamic feature.
From a clinical perspective, the proposed hemodynamic analysis could be integrated into the routine breast MRI reporting workflow. After a standard breast MRI exam, the radiologist could use a semi-automated tool to segment the most suspicious axillary lymph node and an adjacent vessel, with the system automatically calculating the ‘RATIO TO LYMPH’ feature. This quantitative metric could then be combined with other image findings and clinical data within a nomogram (as demonstrated here) to provide a personalized, objective probability of metastasis, aiding in pre-operative planning and potentially reducing unnecessary SLNB procedures in low-probability cases. As discussed in comparison with other leading AI models (25-27,29), which typically report AUCs between 0.89 and 0.91, our baseline model performed comparably. The key incremental benefit of our study is the consistent and significant performance boost (AUC increase of 0.038–0.039 on the test set) achieved by integrating the lymph node-vessel hemodynamic feature. This underscores that capturing functional relationships within the axillary microenvironment provides valuable, complementary diagnostic information beyond the tumor or lymph node alone.
This study has several limitations. First, the retrospective, single-center design may introduce selection bias and limit the generalizability of the findings. External validation on a multi-center cohort is necessary to confirm the robustness of our proposed feature and model. Second, the manual identification of the target lymph node by a radiologist, while based on standard morphological criteria, introduces a degree of subjectivity. Future work could explore fully automated lymph node detection and pairing with adjacent vessels. Third, due to the retrospective nature of the study and standard clinical imaging protocols, there was no specific lymph node labeling on the MRI images to guarantee a one-to-one correspondence with the specific lymph nodes examined pathologically. The MRI assessment was of the axillary region as a whole, correlated with the final overall pathological nodal status.

Conclusions
This study proposed an innovative analytical method for ALNM. In this method, the ROIs for the axillary blood vessels and lymph nodes in DCE-MRI images were semi-automatically extracted to generate corresponding TICs and construct features based on TICs. The statistical values of these features in the metastatic and non-metastatic groups indicate that these features play a role in distinguishing between ALNM and non-ALNM. Subsequently, this study incorporates these features into the ALNM prediction model based on DCE imaging and clinical characteristics using ensemble learning methods. The experimental results show that the model’s predictive performance is further enhanced compared to previous methods. It is evident that using the contralateral axillary vessels as a reference can effectively eliminate individual differences, aid in predicting lymph node metastasis, and provide new imaging characteristics for clinical diagnosis.

Supplementary
The article’s supplementary files as