DCE-MRI-based machine learning model for predicting axillary lymph node metastasis in breast cancer

Introduction

Accurate identification of axillary lymph node (ALN) metastasis in breast cancer is crucial for staging and treatment planning (1,2). Clinical axillary staging and treatment for early breast cancer have evolved from ALN dissection (ALND) to sentinel lymph node biopsy (SLNB), which effectively avoids postoperative drainage and reduces haematoma and pain, neuropathy, limited arm abduction, lymphedema and cellulitis risk (3,4). However, 43–65% of sentinel lymph node (SLN)-positive patients undergo unnecessary axillary surgery because they have no other non-SLN metastases (5,6). Considering that most patients with early-stage breast cancer do not have ALN metastases, accurate assessment of ALN may avoid non-essential SLN biopsies (7).

Breast axillary ultrasonography is an important screening tool for assessing whether ALN is metastatic (8), which can accurately evaluate lymph node morphology and detect metastatic deposits larger than 5 mm in diameter; however, ultrasound is limited by its subjectivity. Magnetic resonance imaging (MRI) is a multiparameter, multidirectional method that can provide biological information about the primary focus and ALN (9,10). Dynamic contrast enhanced MRI (DCE-MRI) can describe the morphological features of the tumour in detail, reveal the kinetics of tumour enhancement and respond to angiogenesis. Distant metastasis and survival in breast cancer have been associated with tumour heterogeneity measured in DCE-MRI images (11,12); therefore, information suggestive of patient prognosis appears to be present in DCE-MRI images of breast cancer.

Radiomics, which provides methods for identifying lesions at the macroscopic and microscopic levels by transforming medical images into quantifiable multidimensional data, has been increasingly used for disease detection and classification in recent years (13,14). The goal of this study was to develop a radiomics-based predictive model to predict ALN status using DCE-MRI images in breast cancer. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2024-495/rc).

Methods

Clinical information

This retrospective study was approved by the Ethics Committee of The First Central Hospital of Baoding (approval No. 2022093) and individual consent for this retrospective analysis was waived. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). Clinical and imaging data of 166 cases of pathologically confirmed breast cancer were collected from January 2017 to August 2020, and the inclusion criteria were (I) pathologically confirmed breast cancer; (II) ALN status determined by ALN biopsy and/or ALN stripping. Exclusion criteria: (I) combination of other malignancies; (II) loss of clinical or imaging data; (III) multiple or bilateral lesions; (IV) poor image quality. According to the ALNB or ALN stripping results, patients were divided into ALN metastasis group and non-metastasis group. A total of 166 patients were included in this study, including 133 patients in the training group with a mean age of 48.1±10.2 years, and 33 patients in the test group with a mean age of 47.8±10.8 years.

MRI was performed using an Ingenia MR 3.0 T (Philips, The Netherlands) scanner equipped with a 7-channel phased array body coil, images were acquired in the prone position with both breasts naturally descending in the coil, and the scanning area was the breasts and axillae bilaterally. The contrast agent was a single dose of 0.1 mmol/kg (equivalent to 0.1 mL/kg) of gadopentetate dextran at a flow rate of 2.0 mL/s, diluted with 20 ml of saline. Collection sequences were DCE sequences, diffusion-weighted imaging (DWI; b=0, 800s/mm²) and apparent diffusion coefficient (ADC) maps derived from DWI, with enhanced imaging at 61s intervals.

Image segmentation and feature extraction

Segmentation of the lesions was done manually on the enhanced first-phase T1-weighted images (Figure 1). The cross-sectional T1+C images were imported into three-dimensional (3D) Slicer (version 4.10.2, www.slicer.org) software, and a physician with 10 years of experience in diagnostic breast imaging sketched layer by layer along the edge of the lesion without knowledge of the clinical results, and fused to obtain the 3D volume of interest (VOI). The VOI will be corresponded to other images, and when the lesion area is difficult to define, a senior physician (30 years of experience in breast imaging) will be sought to determine the boundary. The VOIs and corresponding images were imported into PyRadiomics software (14), and 1,409 3D radiomics features (3D-VOIs) were extracted for each period, totalling 9,863. Subsequently, the tumour area of each layer was calculated based on the VOI to obtain the maximum tumour dimension, and two-dimensional features (2D-ROI) were extracted for different periods based on the maximum tumour dimension, with a total of 7,224.

Figure 1 Manual tumour segmentation for physicians based on MRI images. MRI, magnetic resonance imaging.

Thirty patients were selected by complete randomisation, and two physicians with 15 and 10 years of experience in breast diagnosis respectively sketched the VOIs again and carried out feature extraction for intraclass correlation coefficient (ICC) consistency evaluation within and between groups.

Feature selection and classification by machine learning methods

In this study, the least absolute shrinkage selection operator (LASSO) algorithm was used to build the model. The features were firstly homogenised, and then the number of features was reduced by feature dimensionality reduction to reduce the risk of model overfitting and avoid statistical bias; the Mann-Whitney U test was used to select the features that were highly correlated with the ALN transfer, and the significance level was set to 0.05 (P<0.05); the coefficient between features (R) was then calculated to avoid redundancy of the features, and this study indicated that the features were highly correlated. correlation cut-off value R was set to 0.8; feature pairs with lower P values in a strongly correlated feature pair will be discarded. The same dimensionality reduction method was used for all 3D-VOI features and 2D-ROI features, and finally, non-zero coefficient features were filtered using the LASSO algorithm to calculate Rad-score for all patients. The feature selection process is shown in Figures 2,3.

Figure 2 Optimal values of LASSO algorithm. LASSO, least absolute shrinkage selection operator.

Figure 3 Characteristics of non-zero coefficients of LASSO algorithm and their corresponding coefficients. NA, not available; LASSO, least absolute shrinkage selection operator.

To illustrate the performance of the models the model based on multi-period 3D features was compared with the model built based on multi-period 2D features, pre-enhanced MRI features (Pre) and enhanced first-period MRI features (1^st); the same dimensionality reduction and modelling method was used for all models for a fair comparison. Based on the best-performing model, a comprehensive model for predicting ALN metastasis in breast cancer was established by combining the semantic features of patient image reports.

Model visualisation

The importance of features was ranked to indicate the importance of each feature in decision making, and in order to investigate the contribution of each imaging feature to the model output, heatmaps were used to depict the contribution of different features in individual decision making.

Statistical analysis

Statistical analysis was performed using R (version 3.6.2; https://www.r-project.org) and Python (version 3.8; https://www.python.org/). The model was evaluated using the following metrics: area under the curve (AUC), accuracy, sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), and AUC comparison was done using the Delong test with a 10-fold cross-validation to assess model accuracy.

Results

The basic information of the patients was shown in Table 1, with 72 ALN metastases in the training group and 17 ALN metastases in the test group.

Before feature dimensionality reduction, inter-observer and intra-observer consistency were first assessed, and the results showed that the intra-group feature consistency was 0.85, and the inter-group feature consistency was 0.82 and 0.84, and the ICC were both >0.8, which indicated that the ROIs had good consistency. During the dimensionality reduction process, a total of 4,892 3D features were deleted due to irrelevance, and a total of 4,496 3D features were deleted due to high redundancy, resulting in 152 3D features. The dimensionality reduction features were input into the LASSO algorithm, and the non-zero coefficient features were screened to establish the ALN metastasis prediction model for breast cancer, and the feature screening process was shown in Figures 2,3. The radiomics score (Radscore) of each patient was calculated according to the LASSO method, and the distribution of Radscore in the training and test groups was shown in Figures 4,5; Table 2 and Figures 6,7 showed the performance of the imaging histology model based on 3D MRI features to predict ALN metastasis in breast cancer. The results showed that the AUC, accuracy, sensitivity, and specificity in the training group were 0.82, 0.74, 0.67, and 0.85, respectively; and the AUC, accuracy, sensitivity, and specificity in the test group were 0.80, 0.79, 0.85, and 0.75, respectively. The 10-fold cross-validation average AUC, accuracy, sensitivity, and specificity were 0.82, 0.82, and 0.83, respectively, 0.81 (Figure 8), and the Delong test result of P>0.05 showed good predictive stability. Figure 9 showed the contribution of different radiomics features in individual decision-making, which shows that the radiomics-based approach using pre-treatment 3D MRI image images can accurately predict ALN metastasis in breast cancer.

Figure 4 Radscore distribution cloud and rain plot for training group. X-axis: the model predictive value; Y-axis, group; group 0: non lymph node metastasis; group 1: lymph node metastasis.

Figure 5 Cloud and rain plot of Radscore distribution for test group. X-axis: the model predictive value; Y-axis, group; group 0: non lymph node metastasis; group 1: lymph node metastasis.

Figure 6 ROC curve for training group. ROC, receiver operating characteristic; AUC, area under the curve.

Figure 7 ROC curve for testing group. ROC, receiver operating characteristic; AUC, area under the curve.

Figure 8 3D model 10-fold cross-validation results. AUC, area under the curve; 3D, three-dimensional.

Figure 9 Heatmap of feature visualisation.

In order to further validate the performance of the 3D radiomics model for predicting breast cancer ALN, input different images were selected for side-by-side comparison, and the input features included multi-period radiomics features based on the largest surface of the tumour (2D features), 3D features based on pre-enhanced MRI images (Pre features) and 3D features based on the first-period images after enhancement (1^st features). The patient dataset was identical to the 3D image radiomics model, and the different features were input into the model, which was divided into training group and test group according to 7:3 for model building and testing, and the results were shown in Table 2 and Figures 10,11. The results showed that the radiomics model based on 3D multi-period MRI images has the best performance in predicting ALN metastasis in breast cancer.

Figure 10 Comparison of ROC of different models’ performance in predicting axillary lymph node metastasis of breast cancer in the training group. Group: Groups 1, 3, 4, 5 represent the integrated model and 2D, 3D model based on pre-enhanced MRI images (Pre features) and 3D model based on the first-period images after enhancement (1^st features). ROC, receiver operating characteristic; 2D, two-dimensional; 3D, three-dimensional; MRI, magnetic resonance imaging.

Figure 11 Comparison of ROC of different models’ performance in predicting axillary lymph node metastasis of breast cancer in the test group. Group: Group 1, 3, 4, 5 represent the integrated model and 2D, 3D model based on pre-enhanced MRI images (Pre features) and 3D model based on the first-period images after enhancement (1^st features). ROC, receiver operating characteristic; 2D, two-dimensional; 3D, three-dimensional; MRI, magnetic resonance imaging.

By summarising the semantic features of the image report, incorporating including the breast gland background enhancement pattern, DWI performance, whether it invades the skin, lymph node MRI performance, tumour size and age. The combined radiomics predictive values and MRI semantic features were established to predict breast cancer ALN metastasis in a comprehensive column line diagram, as shown in Figure 12, where the position of each clinical and imaging histology variable on the corresponding axis was sought separately, and a plumb line was drawn on the point axis to indicate the number of points corresponding to that variable; the number of points of each variable was summed up and a plumb line was drawn on the total score axis, and the corresponding probability value was the column line diagram diagnosis of breast cancer ALN metastasis the probability of the diagnosis. Figures 13,14 demonstrate the performance of the integrated model, and the results show that the integrated model has a good performance in predicting ALN metastasis in breast cancer, with a C-index of 0.90, an AUC of 0.90 [95% confidence interval (CI): 0.86–0.95], a specificity of 0.91, a sensitivity of 0.77, an accuracy of 0.84, an NPV of 0.78, and a PPV of 0.91.

Figure 12 Performance of integrated model in predicting axillary lymph node metastasis in breast cancer. DWI, diffusion-weighted imaging; LN, lymph node.

Figure 13 ROC curve for column-line diagram. ROC, receiver operating characteristic; AUC, area under the curve.

Figure 14 Calibration curves for column line graphs.

Discussion

ALN status determines the surgical extent, reconstruction and postoperative treatment options, and preoperative ALN staging is crucial in determining the treatment strategy for early-stage breast cancer. In this study, we developed an artificial intelligence (AI) model for predicting breast cancer ALN status based on 3D MRI radiomics features with an AUC of 0.90, which achieved non-invasive and accurate identification of breast cancer ALN status.

MRI-based AI shows promise in identifying lymph node metastasis. Nguyen et al. (15) developed a four-dimensional convolutional neural network (4D CNN) to predict ALN metastasis by fusing spatiotemporal features of primary tumours in DCE-MRI. The best prediction model developed in this study had a sensitivity of 72% and an AUC of 71%. Zhang et al. (16) used a weighted voting method to predict ALN metastasis by constructing a multi-parameter ResNet50 model based on T2 weighted image (T2WI), DWI and DCE-MRI, which had an AUC of 91% (95% CI: 0.799–0.974), however the prediction model incorporating DCE-MRI images had an AUC of only 57% (95% CI: 0.424–0.711). Gao et al. (17) built a predictive model containing a 3D deep residual network (ResNet) architecture for the prediction of ALN metastasis with three ResNet architecture and a convolutional block attention module (RCNet) models based on tumour, ALN and tumour-ALN, which had an AUC of 85%, accuracy of 83%, sensitivity of 79% and specificity of 85% in an external test cohort. Some studies have used the same methodology, utilising multi-parameter MRI radiomics and deep learning features, without considering temporal relationships despite the use of multi-parameter or multi-scanning loci (18,19). Compared with the above research methods, this study achieved accurate prediction of breast cancer ALN status from multiple perspectives in both temporal and spatial dimensions and obtained higher comprehensive diagnostic accuracy by making full use of dynamic contrast-enhanced multi-period images of DCE-MRI in conjunction with tumour 3D features.

Although breast MRI can assess ALN status directly by observing lymph nodes, there are limitations. The breast is located in the core of the coil, whereas the ALN may be located at the edge of the coil affecting the accuracy of the examination; in addition, some lymph nodes may not show signs of metastasis when they are in the early stages of metastasis. This study makes predictions directly from the primary lesion to compensate for the limitations of MRI in showing axillary status.

There are limitations in this study. Firstly, for patients with multifocal and bilateral lesion breast cancer, this study was unable to locate the specific tumour causing ALN metastasis; therefore, this group of patients was excluded from this study, and a model to predict the status of multifocal and bilateral lesion ALN needs to be established in future studies. Secondly, the predictive model developed in this study is based on the primary breast cancer foci rather than the lymph nodes themselves, whereas the inclusion of ALN images may be an important contribution to the improvement of prediction results, and in future work it may be possible to identify suspected metastatic lymph nodes with the help of localisation clips and to match them to the pathology results in order to make direct predictions based on lymph node images. In addition, the patients included in this study could potentially bias the results due to the limitations of the examination equipment and region, and in the future, we will collect patients from multiple centres, different equipment and prospective database to further validate our model. Finally, this study was retrospective and lacked prospective validation, and a multicentre prospective study will be conducted in future work to more fully validate the model.

Conclusions

In summary, machine learning is an emerging computer-aided diagnostic technology that has emerged in recent years, and AI based on machine learning has an inherent advantage. Although the research on breast tumour imaging is in its infancy, it is believed that in the near future, with the continuous development of algorithms, it can be fully integrated into the clinic and become a routine practice for clinicians.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2024-495/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2024-495/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2024-495/prf

Funding: This study was supported by the National Natural Science Foundation of China (No. 62201386), Science and Technology Programme of Baoding (No. 2441ZF219).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2024-495/coif). The authors have no 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 retrospective study was approved by the Ethics Committee of The First Central Hospital of Baoding (approval No. 2022093) and individual consent for this retrospective analysis was waived. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

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. Chang JM, Leung JWT, Moy L, et al. Axillary Nodal Evaluation in Breast Cancer: State of the Art. Radiology 2020;295:500-15. [Crossref] [PubMed]
2. Chung HL, Le-Petross HT, Leung JWT. Imaging Updates to Breast Cancer Lymph Node Management. Radiographics 2021;41:1283-99. [Crossref] [PubMed]
3. Jatoi I, Kunkler IH. Omission of sentinel node biopsy for breast cancer: Historical context and future perspectives on a modern controversy. Cancer 2021;127:4376-83. [Crossref] [PubMed]
4. Jatoi I, Benson JR, Toi M. De-escalation of axillary surgery in early breast cancer. Lancet Oncol 2016;17:e430-41. [Crossref] [PubMed]
5. Chu KU, Turner RR, Hansen NM, et al. Sentinel node metastasis in patients with breast carcinoma accurately predicts immunohistochemically detectable nonsentinel node metastasis. Ann Surg Oncol 1999;6:756-61. [Crossref] [PubMed]
6. Kamath VJ, Giuliano R, Dauway EL, et al. Characteristics of the sentinel lymph node in breast cancer predict further involvement of higher-echelon nodes in the axilla: a study to evaluate the need for complete axillary lymph node dissection. Arch Surg 2001;136:688-92. [Crossref] [PubMed]
7. Kim GR, Choi JS, Han BK, et al. Preoperative Axillary US in Early-Stage Breast Cancer: Potential to Prevent Unnecessary Axillary Lymph Node Dissection. Radiology 2018;288:55-63. [Crossref] [PubMed]
8. Le Boulc'h M, Gilhodes J, Steinmeyer Z, et al. Pretherapeutic Imaging for Axillary Staging in Breast Cancer: A Systematic Review and Meta-Analysis of Ultrasound, MRI and FDG PET. J Clin Med 2021;10:1543. [Crossref] [PubMed]
9. Pesapane F, Downey K, Rotili A, et al. Imaging diagnosis of metastatic breast cancer. Insights Imaging 2020;11:79. [Crossref] [PubMed]
10. Nicolau P, Gamero R, Rodríguez-Arana A, et al. Imaging and pathology features to predict axillary tumor load in breast cancer. J Obstet Gynaecol Res 2018;44:331-6. [Crossref] [PubMed]
11. Kim JY, Kim JJ, Hwangbo L, et al. Kinetic Heterogeneity of Breast Cancer Determined Using Computer-aided Diagnosis of Preoperative MRI Scans: Relationship to Distant Metastasis-Free Survival. Radiology 2020;295:517-26. [Crossref] [PubMed]
12. Kim JJ, Kim JY, Kang HJ, et al. Computer-aided Diagnosis-generated Kinetic Features of Breast Cancer at Preoperative MR Imaging: Association with Disease-free Survival of Patients with Primary Operable Invasive Breast Cancer. Radiology 2017;284:45-54. [Crossref] [PubMed]
13. Zheng X, Yao Z, Huang Y, et al. Deep learning radiomics can predict axillary lymph node status in early-stage breast cancer. Nat Commun 2020;11:1236. [Crossref] [PubMed]
14. Bae MS, Kim HG. Breast Cancer Risk Prediction Using Deep Learning. Radiology 2021;301:559-60. [Crossref] [PubMed]
15. Nguyen S, Polat D, Karbasi P, et al. Preoperative Prediction of Lymph Node Metastasis from Clinical DCE MRI of the Primary Breast Tumor Using a 4D CNN. Med Image Comput Comput Assist Interv 2020;12262:326-34.
16. Zhang X, Liu M, Ren W, et al. Predicting of axillary lymph node metastasis in invasive breast cancer using multiparametric MRI dataset based on CNN model. Front Oncol 2022;12:1069733. [Crossref] [PubMed]
17. Gao J, Zhong X, Li W, et al. 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]
18. Chai R, Ma H, Xu M, et al. Differentiating axillary lymph node metastasis in invasive breast cancer patients: A comparison of radiomic signatures from multiparametric breast MR sequences. J Magn Reson Imaging 2019;50:1125-32. [Crossref] [PubMed]
19. Wang D, Hu Y, Zhan C, et al. A nomogram based on radiomics signature and deep-learning signature for preoperative prediction of axillary lymph node metastasis in breast cancer. Front Oncol 2022;12:940655. [Crossref] [PubMed]