A preoperative prediction model for sentinel lymph node status using artificial intelligence on mammographic images and clinicopathological variables in patients with clinically node-negative breast cancer

doi:10.21203/rs.3.rs-2590918/v1

Background:

Cancer detection systems using artificial intelligence (AI) are a rapidly growing research area, in breast cancer. Sentinel lymph node biopsy (SLNB) is the recommended surgical axillary staging method in patients with clinically node-negative breast cancer, although approximately 75–80% have a negative sentinel lymph node (N0) status in the final pathology report. Previous prediction models for N0 status included variables only available postoperatively, thus defeating their purpose. Therefore, we aimed to investigate whether two AI systems, Transpara (Screenpoint Medical) and Laboratory for Individualized Breast Radiodensity Assessment (LIBRA), on mammographic images can be used to improve a previous prediction model for N0 status using only preoperatively available variables. To our knowledge, this is the first preoperative prediction model for N0 status combining AI detection on mammographic images with clinicopathological variables.

Methods:

This retrospective cohort study included 755 women with primary breast cancer treated at Lund University Hospital between 2009 and 2012. Mammographic images were analyzed using Transpara and LIBRA. Preoperative clinicopathological and radiological variables were used in a multivariable logistic regression analysis to predict N0 status with multiple imputation. The area under the receiver operating curve (AUC) was used to assess model performance and a nomogram was developed. The agreement between preoperative radiological and postoperative pathological tumor size was assessed using correlation.

Results:

We proposed a preoperative prediction model for N0 status using AI detection on mammographic images and clinicopathological variables, with an AUC of 0.695 (confidence interval: 0.653–0.736). Applying the model, SLNB could be putatively omitted in 23.8% of patients if a false-negative rate of 10% was accepted. The mean difference between radiologic and pathologic tumor size was 0.4 mm and the corresponding Pearson correlation coefficient 0.62.

Conclusion:

To our knowledge, the prediction model proposed in this manuscript is the first preoperative prediction model for N0 status using AI on mammographic images and routine preoperative patients and tumor characteristics. The correlation between tumor size measurements suggests that radiologic tumor size can replace pathologic size as a predictor of N0 status. Applying this model may enable safe omission of SLNB in 23.8% of patients.

breast neoplasm

artificial intelligence

prediction model

sentinel lymph node

mammography

Sentinel lymph node biopsy (SLNB) is the recommended surgical axillary staging method in patients with clinically node-negative breast cancer, although approximately 75–80% have non-malignant axillary lymph nodes in the definitive pathology report [1–4]. The SLNB procedure does not benefit patients with negative sentinel lymph node (N0) status but could result in side-effects such as numbness, pain, lymphedema, limited range of motion and postoperative infection [5]. The ongoing Intergroup Sentinel Mamma study (INSEMA) trial studies morbidity and survival after omission from surgical axillary staging, SLNB and axillary lymph node dissection (ALND). The morbidity was lower after no SLNB vs. SLNB and SLNB vs. ALND at one, three, and 18 months [3, 5]. This warrants the de-escalation of axillary surgery and new methods for the preoperative identification of patients with N0 status for whom SLNB can be safely omitted.

Several clinicopathological models for the prediction of axillary lymph node (ALN) and sentinel lymph node (SLN) status have been published [6–9]. In 2019, Dihge et al. developed an artificial neural network (ANN) model to predict N0 status, which provided a framework for this study [10]. An issue with previous prediction models is that they were developed using mainly postoperative variables. A prediction model must be applicable preoperatively to provide a decision tool for the safe omission of SLNB. Pathological tumor size, a key predictor of N0 status, is a postoperative measure that may be replaced by radiological tumor size measured by mammography or ultrasound [6, 7, 11]. Vascular invasion and pathological multifocality are strong predictors of SLN status and are usually defined after pathological assessment of the surgical specimen, therefore, the preoperative availability of these variables is limited. Vascular invasion is an independent predictor of N0 status, however, to our knowledge, there is no preoperatively available alternative in clinical practice [12]. In contrast, multifocality, might be identified radiologically by imaging, although mammography has a lower sensitivity than other imaging modalities [13]. Additionally, tumor characteristics may differ between preoperative core needle biopsy (CNB) and postoperative analysis. The Ki67 labelling index, a measure of tumor proliferation, reveals moderate concordance between CNB and surgical specimens. A more reliable and preoperative proliferation measure may be histological grade [14, 15].

Radiologists currently assess mammographic images in clinical practice. However, research on breast cancer detection systems using artificial intelligence (AI) is rapidly growing. Systems have primarily been evaluated for their accuracy in detecting breast cancer. One of these systems is Transpara (Screenpoint Medical, Nijmegen, Netherlands), a diagnostic tool for detecting breast cancer. Several studies have evaluated and revealed Transpara to improve the sensitivity and specificity when used as a support system in radiologists’ assessment of mammograms, but not for the prediction of SLN status [16–18].

Prediction models for ALN status using mammographic images have previously been presented, using the presence of microcalcifications, breast density, and radiomic signatures, exclusively or in combination with clinicopathological variables, most of which were postoperatively captured [19–22]. Several other imaging modalities have been assessed for their nodal predictive value, such as contrast-enhanced computed tomography (CECT), magnetic resonance imaging (MRI) and ultrasound [23–27]. To our knowledge, no preoperative prediction model for N0 status currently incorporates AI detection on mammographic images and preoperatively available clinicopathological variables.

We aimed to evaluate the additional predictive value of radiomic variables through AI detection by Transpara on mammographic images to a preoperative prediction model for N0 status in patients with breast cancer with a clinically node-negative status. Furthermore, we present a comparison of the performance of the new preoperative prediction model and a previously published postoperative model.

Study population

A total of 770 patients diagnosed with primary breast cancer between January 2009 and December 2012 were prospectively included in a registry at the Department of Pathology at the Skåne University Hospital (Lund, Sweden). Inclusion and exclusion criteria were previously described by Dihge et al. [6, 10]. Patients underwent axillary staging using SLNB, and if needed ALND.

Clinicopathological data

Patient and tumor information was collected from patients’ electronic files and a pathology database, as described by Dihge et al. [6, 10]. The histological type was divided into two groups; the first group include ductal and lobular, while the second group include other and mixed types. Estrogen receptor status, progesterone receptor status and Ki67 percentage were analyzed and categorized according to guidelines [28, 29].

The SLN status was categorized as negative or positive (N0 or N+). N0 was defined as breast cancer with no lymph node metastasis or with only isolated tumor cells. N+ was defined as ≥1 SLN with micrometastases or macrometastases, defined as more than 200 cells and/or cluster size of 0,2 - 2 mm and cluster size >2 mm, respectively [30].

Mammographic images and image analysis systems

All available mammographic images from screening and diagnostic imaging assessments were included in the analyses. A modified Breast Imaging Reporting and Data System malignancy score (1–5) was used for mammographic and ultrasound images. The following information was collected from the Picture Archiving and Communication System (PACS): mammographic malignancy score (1–5), ultrasound malignancy score (1–5), radiologic tumor size in mm, and laterality. In cases where mammographic tumor size was missing, tumor size on ultrasound was used instead.

All mammographic images were cross-checked for laterality and correct tumor localization in Transpara before data collection. Transpara (version 1.7.0, Screenpoint Medical, Nijmegen, the Netherlands), uses deep learning algorithms to analyze diagnostic mammography and digital breast tomosynthesis images, and detects lesions suspicious for breast cancer, soft tissue lesions and calcifications (calc). Transpara marks the most suspicious findings on every mammographic image and assigns the suspicious region a score between 1 and 100. An overall exam score between 1 and 10 is reported, indicating the increasing likelihood that a visible cancer is presented on the mammographic images. Transpara was calibrated to categorize approximately 10% of examinations into each of the 10 risk groups and at least 90% of cancers in group 10 if run at a standard screening image material. The images used in this study were not used to train, validate, or test the software. The following data were collected from Transpara: the number of calc clusters, the highest calc cluster score, the number of soft tissue lesions and the highest soft tissue lesion score. The variable highest calc cluster score was divided into three groups: absence of visible clusters, low (1–92) and high (≥93). The variable highest soft tissue lesion score was divided into four groups: absence of visible lesions, low (1–91), medium (92–95) and high (≥96). All available mammograms were included in this assessment.

The Laboratory for Individualized Breast Radiodensity Assessment (LIBRA) (version 1.0.5) is an automated breast density estimation algorithm researchers at the University of Pennsylvania developed. It analyzes full-field digital mammography images based on gray-level values and segments them into dense and non-dense areas [31]. LIBRA has been validated by Gastouni et al. [32], among others. Studies have revealed an association between dense LIBRA area and breast cancer [32]. The following data were collected from the LIBRA: total breast area [cm²], dense area [cm²], and density [%]. Only mammographic images off craniocaudal (CC) and mediolateral oblique (MLO) projections were included. The mean values of the CC and MLO on the ipsilateral side were used.

Statistical analysis

Descriptive analyses were performed to explore the associations between clinical, pathological and radiomic variables and SLN status using the Mann–Whitney U test and Chi-square test for continuous and categorical variables, where appropriate. Tumor size comparison between pathological and mammographic variables was performed using Pearson correlation and Bland-Altman analysis. Univariable logistic regression was performed for all variables to predict N0 status. As a first model, the top ten variables included in the previously published model [10] were used by applying multivariable logistic regression (MLR). This method differs from that of the original article, where an ANN model with internal validation by five-fold cross-validation repeated ten times was used. The performance of a more pragmatic MLR model, as measured by the area under the received operation characteristics (ROC) curve (AUC), was in that study, however, only marginally inferior. The second model excludes vascular invasion and multifocality, variables which are difficult to obtain or of poor quality preoperatively. Postoperative Ki67 and pathological tumor size were exchanged for the corresponding preoperatively available variables, histological grade and radiological tumor size, in the third model. The models are presented in Figure 2. Multiple imputation with chained equations was used to handle missing data on clinicopathological predictor variables and radiological tumor size. Number of missing values per predictor is presented in Table 1-2. The multiple imputation strategy is described in detail in Supplementary, Additional File 1, and variables included in the imputation in Supplementary Table 1, Additional File 2. Briefly, 100 imputed datasets were created using linear regression for continuous variables, logistic regression for binary variables and predictive-mean-matching for categorical variables with more than two categories. A comparison of complete radiological cases and cases with missing radiological tumor size is presented in Supplementary Table 2, Additional file 2

. Model selection by stepwise backward elimination, with a p-value threshold for removal of ≥0.157 and ≥0.050, respectively, was performed to derive the fourth and final model with the new AI-derived variables as additional candidates. The measures AUC, from ROC analysis, and the Akaike information criterion (AIC) were used to discriminate between prediction models. The SLNB reduction rate was calculated with a cut-off based on a false-negative rate (FNR) of 10%, reflecting the FNR of SLNB [33, 34]. A nomogram was developed for the final model to visualize the strengths of the associations between each of the predictor variables and N0 status.

P-values in univariable analyses of associations between potential predictor variables and the outcome N0 status were not corrected for multiple comparisons due to the explorative nature of the study. All p-values are two-sided and should be interpreted as level of evidence against the null hypothesis without reference to a cut-off for significance. Stata (StataCorp. 2021. Stata Statistical Software: Release 17. College Station, TX; StataCorp LLC) was used.

Mammograms were identified in 755 of the 770 women included in the study. Upon data collection from the PACS mammograms, 30 women’s’ data were missing, and Transpara failed to analyze mammograms from three women. Inconclusive cases, according to radiologists, and cases without a clear indication of tumor location on mammographic images in PACS were excluded due to an inability to cross-check Transpara’s findings. The Transpara and LIBRA subgroups consisted of 689 and 755 patients, respectively. A total of 689 complete cases were included in the radiological prediction models, with multiple imputation of missing values for the clinicopathological variables. Transpara’s assessment of tumor localization on mammograms was correct in 96.1% of the cases. The inclusion and exclusion criteria are presented in Figure 1.

The patient and tumor characteristics are presented in Table 1. In this cohort, 64.8% of the participants were N0 and 35.2% were N+. The patient and tumor characteristics showing the strongest evidence for association with N0 status were pathological tumor size (p <0.001), multifocality (p <0.001), mode of tumor detection (p <0.001), vascular invasion (p <0.001), Ki67 (p <0.001), histological grade (p=0.007), age (p=0.027) and histological type (p=0.046). Patient and tumor characteristics for complete radiological cases and cases with missing mammographic images are presented in Supplementary Table 3, Additional File 2. Radiological variables are presented in Table 2. The variables showing the strongest evidence for association with N0 status were radiological tumor size (p <0.001), highest calc cluster score group (p <0.001), highest soft tissue lesion score continuous (p <0.001) and highest soft tissue lesion score group (p <0.001).

(Table 1)

(Table 2)

A Comparison of tumor size by pathological and mammographic assessment is presented in Supplementary Tables 3, Additional File 2. The mean tumor size pathology and radiology were 16.7 and 17.1, respectively, and the Pearson correlation coefficient was 0.62. A Bland-Altman plot illustrates that the difference in mean values between the two modalities is close to zero (Supplementary Figure 1, Additional File 2).

The univariable logistic regression analyses with radiological variables as predictors of N0 status are presented in Table 3. The model selection process is illustrated in Figure 2. The first, recreated, model had an AUC of 0.775 (95% confidence interval (CI): 0.740–0.810), including all available cases. Similar results were obtained when only the cases with complete Transpara variables were included, data not shown. The second model, without vascular invasion and multifocality, had an AUC of 0.725 (95% CI: 0.687–0.762). The third model, with radiological tumor size and histological grade instead of the postoperatively available pathological tumor size and Ki67, had an AUC of 0.686 (95% CI: 0.647–0.725). Backward elimination with p ≥0.157 of the third model resulted in a fourth model, including radiological tumor size, estrogen receptor status, age, mode of detection, histological type, and tumor localization, with an AUC of 0.679 (95% CI: 0.639–0.719) (Table 4). Backward elimination with a threshold of p ≥0.050 resulted in the same results, as presented in Supplementary Table 5, Additional File 2. The separate addition of radiological variables to the fourth model showed that the highest calc cluster score, highest soft tissue lesion score grouped, and highest soft tissue lesion score contributed to the prediction of N0 status. Backward elimination of the third model and all radiological variables contributing to the prediction resulted in the final model, including radiological tumor size, estrogen receptor status, age, mode of detection, histological type, tumor localization, highest calc cluster score grouped, and highest soft tissue lesion score grouped, with an AUC of 0.695 (CI: 0.653–0.736). The ROC curves corresponding to the first, third, and final models are illustrated in Figure 3. For comparison, AUC for the Transpara variables included in the final model, highest calc cluster score and highest soft tissue lesion score, were 0.514 (95% CI: 0.476-0.551) and 0.544 (95% CI: 0.501-0.587), respectively. Table 5 displays a possible SLNB reduction rate of 23.8% after applying the final prediction model to assign N0, corresponding to a cut-off that accepts a FNR of 10%. A nomogram visualizing the associations between the predictors and N0 status was developed (Figure 4).

(Table 3)

(Table 4)

Table 5: SLNB reduction rate using the final prediction model for positive sentinel lymph node status, corresponding to a FNR of 10%.

	TP	TN	FP	FN
No.	209	141	316	23
SLNB reduction rate	(TN + FN) / (TP + TN + FP + FN) = 23.8%

Abbreviations: sentinel lymph node biopsy (SLNB), false-negative rate (FNR), true positive (TP), true negative (TN), false-positive (FP) and false-negative (FN). In this table, positive refers to positive sentinel lymph node status and negative to negative sentinel lymph node status.

In the present study, we developed a prediction model for N0 status in primary breast cancer using AI on mammographic images and routine preoperative clinicopathological characteristics. Several radiological variables were associated with N0 status, supporting the hypothesis that mammographic features can aid in preoperatively identifying these patients. Transpara variables improved discrimination when added to an exclusively preoperative clinicopathological model. The final model had an AUC of 0.695 (CI: 0.653–0.736). If an FNR of 10% is accepted, reflecting the accepted FNR of SLNB, the proposed model could putatively support the omission of 23.8% SLNB. Moreover, preoperative radiologic tumor size was closely associated with postoperative pathologic tumor size and was the strongest predictor of N0 status.

Pathologic and radiologic measurements on tumor size were strongly associated in this material (Supplementary Table 4, Additional File 2), which is in accordance with previous research. An evaluation between pathology, MRI, ultrasound, and mammography showed that mammographic tumor size was similar to the reported pathological tumor size but tended to be slightly underestimated [11]. In this material we observed that decreasing radiologic tumor size was associated with N0 status. Pathological tumor size is a strong predictor for SLN status and often included in prediction models, although it is a postoperative variable [10, 19, 20]. These and previous results indicate that mammographic tumor size could replace pathological tumor size assessed from the surgical specimen as a predictor for N0 status.

The first model (Figure 2), including variables proposed by Dihge et al. [10], displayed similar discriminatory capacity when including all cases or only complete radiological cases, indicating that the AUC of the first model can be compared with that of the final model. The final model (AUC 0.695) had a lower AUC than the first model (AUC 0.775), which was expected as strong predictors, that could only be determined through the evaluation of the surgical specimen, were excluded from the model to make it clinically useful in a preoperative setting. The AIC was lower for the final model than for the first model. The fourth, reduced preoperative clinicopathological, model had a slightly lower AUC than the final model. The higher discriminatory capacity of the final model for N0 status indicates that it could potentially be used in clinical practice to preoperatively determine the omission of SLNB in selected patients. Importantly, the results indicate a predictive value of radiological variables in a preoperative prediction model.

When elaborating on the radiological variables used in this study, it is important to note that Transpara is not intended to be used as a tool for prediction of SLN status, although this study also indicates a potential predictive value in this setting. Additional development of Transpara in this direction could improve its predictive ability for SLN status. The potential clinical use and definitions of medicolegal regulations regarding this type of diagnostic tool are debated and yet to be determined [35, 36]. Implementation of image analysis software in clinical practice could have a dual application as the software is designed to detect breast cancer in a screening setting and could therefore be integrated both into cancer detection and prediction of SLN status [37]. Implementing a prediction model would entail additional costs associated with procurement, whereas omission of SLNB would likely reduce costs associated with surgery, as has been revealed previously for the ANN model proposed by Dihge et al. [10, 38]. Omission could improve the quality of life and reduce postoperative morbidity, as revealed in the INSEMA trial, where arm morbidity was significantly less frequent in the untreated group than in the SLNB group [3]. Therefore, a preoperative prediction model for N0 status could be of great clinical relevance and high value for individual patients.

The clinicopathological ANN model proposed by Dihge et al. which included some postoperative variables, had an AUC of 0.740. They performed a MLR including the same variables for comparison with an AUC of 0.727, which is slightly lower than that of the proposed ANN model [10]. In comparison with the first model in the present paper, the lower discriminatory capacity of the MLR model could be due to differences in methods where cross-validation and techniques to reduce overfitting were used in the original article. A limitation of previous prediction models is that key variables can only be obtained postoperatively [6, 7, 10]. In other studies, this problem was circumvented by including radiological variables from different imaging modalities exclusively or in addition to clinicopathological variables. Liu et al. proposed an exclusively radiological ANN model using CECT [23]. However, this method is not feasible for clinical implementation as all patients with breast cancer generally do not undergo CECT during the initial routine workup, an argument which also applies to models that include MRI images [24, 25]. Given the wide implementation of mammography screening programs, mammographic images are available for all patients and can be used for preoperative diagnostics. Cen et al. proposed a model that included age, postoperative pathological tumor size and microcalcification density on mammographic images, resulting in a model with an AUC of 0.70 [19]. They revealed that a microcalcification density >20 cm² was associated with a positive ALN status, similar to our findings. A negative association between the absence of calc clusters and N0 status was discovered in our study but was not included in the model proposed by Cen et al. Correlation with an unknown predictive variable could explain the association with the absence of calc clusters. Transpara’s assessment of calc clusters includes several features, such as cluster area, which can be compared to microcalcification density [39]. Yang et al. [21] created a prediction model for ALN status using a radiomic signature on mammography (n=147) and with an AUC of 0.88 in the validation cohort, indicating a promising predictive capacity of mammographic features for N0 status. A study by Hack et al. [20] evaluated mammographic density and postoperative clinicopathological variables for prediction of N0 status. No association was discovered between mammographic density and SLN status, similar to our findings.

A strength of this study was the relatively large cohort of 770 patients. All eligible cases during a four-year period were consecutively included, and the cohort should therefore be representative of the breast cancer population at Skåne University Hospital (Lund, Sweden) during the study period. This single- center approach makes it likely that all patients were examined identically, thus reducing the bias caused by differing assessment criteria between clinics. In addition, in this study, we captured comprehensive radiological data, used two AI-based systems, Transpara and LIBRA, and explored their association with SLN status. Another strength is the assessment of Transpara’s accuracy through cross-checking for the correct tumor location. All cases were included to resemble a clinical setting, regardless of the variable. Additionally, as missing data was handled by multiple imputation, all available cases were included.

A limitation of our study is that 65% of patients were N0 in this study, which is lower that the prevalence in more recent cohorts. This could be due to the difference in study time, as more breast cancers are being discovered earlier than in past decades. Another limitation is the exclusion of 45 cases due to missing images in the Transpara sub-cohort. No cause was identified despite repeated contact with technical support at the hospital and the PACS provider. The number of missing cases was relatively small; however, there were differences in patient and tumor characteristics between the missing and included cases, which could have affected the results. Other limitations include the data-driven categorization of Transpara variables and the histological type, as well as the lack of internal validation, which could have resulted in a slight overestimation of the AUC. Additionally, external validation is required to assess the actual predictive value of the final model proposed in this study. The performance of Transpara variables could also be biased owing to missing mammograms. The prediction model on which this study is based is an ANN model that captures non-linear associations and interactions, whereas the MLR model captures only linear effects on the log odds scale. This is a limitation and a strength, as the risk of overfitting is much lower with MLR. No interaction variables were included in the MLR models, as this would require a larger cohort. Considering these limitations, the results of our study should be interpreted with caution. Validation in an external cohort should be performed to further evaluate the proposed model.

Radiological variables from AI detection on mammographic images have nodal predictive capabilities and can improve clinicopathological preoperative prediction models. This study presents a preoperative prediction model for N0 status, which could enable the omission of SLNB in 23.8% of patients with primary breast cancer and supports the hypothesis that radiological tumor size could replace pathological tumor size in prediction of N0 status.

AI Artificial intelligence

AIC Akaike information criterion

ALN Axillary lymph node

ALND Axillary lymph node dissection

ANN Artificial neural network

AUC Area under the receiver operating characteristic curve

CC Craniocaudal

CECT Contrast-enhanced computed tomography

CI Confidence interval

CNB Core needle biopsy

FNR False-negative rate

INSEMA Intergroup Sentinel Mamma study

LIBRA Laboratory for Individualized Breast Radiodensity Assessment

MLO Mediolateral oblique

MLR Multivariable logistic regression

MRI Magnetic resonance imaging

N0 Negative sentinel lymph node

N+ Positive sentinel lymph node

PACS Picture Archiving and Communication System

ROC Receiver operating characteristic

SLN Sentinel lymph node

SLNB Sentinel lymph node biopsy

Ethics approval and consent to participate

This study was approved by the Regional Ethical Review Board of Lund, Sweden, 2012-08-15. Registration number: 2012-340. All patients consented to be included in the National Quality Registry for Breast Cancer. Patients had the opportunity to opt out of participation in this study after advertisement and information in a local newspaper. The local department for personal data admission at the care giver (KVB Samråd, Region Skåne) granted access to digital images.

Consent for publication

Not applicable.

Availability of data and materials

The datasets generated and analyzed during the current study are not publicly available due to sensitive information according to current data legislation but are available from the corresponding author upon reasonable request.

Competing interests

The authors declare that they have no competing interests.

Funding

The Erling Persson Family Foundation 2018/0410; The Governmental Funding of Clinical Research within the National Health Service (ALF) 2018/40304; The Governmental Funding of Clinical Research within the National Health Service (ALF), Young researcher Looket Dihge; The Swedish Cancer Society 2019/0388; The Swedish Research Council 2020/01491; The Swedish Cancer- and Allergy Fund. No funding body took part in the design of the study, collection of data, analysis and interpretation of data and in writing the manuscript.

Authors’ contributions

LR and MD conceptualized this study. Data collection was performed by CR and LD, with assistance from DF. CR drafted the manuscript. CR and POB performed analyses. The manuscript was substantially revised by DF, LD, LR, MD and POB. All authors have approved the submitted version of the manuscript. All authors have agreed to be personally accountable for their contributions and to ensure that questions related to the accuracy or integrity of any part of the work, even one in which the author was not personally involved, are appropriately investigated, resolved, and documented in the literature.

Acknowledgements

We would like to thank Editage [http://www.editage.com] for editing and reviewing this manuscript for English language.

Authors’ information

CR: MD, PhD student, Lund University, Sweden

LD: MD, PhD, Consultant at the Department of Plastic and Reconstructive Surgery, Skåne University Hospital, Malmö, Lund University, Sweden

DF: PhD, Associate Professor,

POB: PhD, Associate Professor, Statistician, Clinical Sciences, Department of Oncology, Lund University, Lund, Sweden

MD: PhD, Associate Researcher,

LR: Professor and Senior Consultant in Surgery, Lund University and Skåne University Hospital

Cardoso F, Kyriakides S, Ohno S, Penault-Llorca F, Poortmans P, Rubio IT, et al. Early breast cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Annals of Oncology. 2019;30(8):1194–220.
Brackstone M, Baldassarre FG, Perera FE, Cil T, Chavez Mac Gregor M, Dayes IS, et al. Management of the Axilla in Early-Stage Breast Cancer: Ontario Health (Cancer Care Ontario) and ASCO Guideline. Journal of Clinical Oncology. 2021;39(27):3056–82.
Reimer T, Stachs A, Veselinovic K, Polata S, Müller T, Kühn T, et al. Patient-reported outcomes for the Intergroup Sentinel Mamma study (INSEMA): A randomised trial with persistent impact of axillary surgery on arm and breast symptoms in patients with early breast cancer. eClinicalMedicine. 2023;55:101756.
McCartan D, Stempel M, Eaton A, Morrow M, Pilewskie M. Impact of Body Mass Index on Clinical Axillary Nodal Assessment in Breast Cancer Patients. Ann Surg Oncol. 2016;23(10):3324–9.
Brar P, Jain S, Singh I. Complications of Axillary Lymph Node Dissection in Treatment of Early Breast Cancer: A Comparison of MRM and BCS. Indian Journal of Surgical Oncology. 2011;2(2):126–32.
Dihge L, Bendahl P-O, Rydén L. Nomograms for preoperative prediction of axillary nodal status in breast cancer. British Journal of Surgery. 2017;104(11):1494–505.
Bevilacqua JLB, Kattan MW, Fey JV, Cody HS, III, Borgen PI, Van Zee KJ. Doctor, what are my chances of having a positive sentinel node? A validated nomogram for risk estimation. Journal of Clinical Oncology. 2007;25(24):3670–9.
Meretoja TJ, Heikkilä PS, Mansfield AS, Cserni G, Ambrozay E, Boross G, et al. A predictive tool to estimate the risk of axillary metastases in breast cancer patients with negative axillary ultrasound. Ann Surg Oncol. 2014;21(7):2229–36.
Reyal F, Rouzier R, Depont-Hazelzet B, Bollet MA, Pierga J-Y, Alran S, et al. The Molecular Subtype Classification Is a Determinant of Sentinel Node Positivity in Early Breast Carcinoma. PLOS ONE. 2011;6(5):e20297.
Dihge L, Ohlsson M, Edén P, Bendahl PO, Rydén L. Artificial neural network models to predict nodal status in clinically node-negative breast cancer. BMC Cancer. 2019;19(1):610.
Gruber IV, Rueckert M, Kagan KO, Staebler A, Siegmann KC, Hartkopf A, et al. Measurement of tumour size with mammography, sonography and magnetic resonance imaging as compared to histological tumour size in primary breast cancer. BMC Cancer. 2013;13(1):328.
Harris GC, Denley HE, Pinder SE, Lee AHS, Ellis IO, Elston CW, et al. Correlation of Histologic Prognostic Factors in Core Biopsies and Therapeutic Excisions of Invasive Breast Carcinoma. The American Journal of Surgical Pathology. 2003;27(1).
Steinhof-Radwańska K, Lorek A, Holecki M, Barczyk-Gutkowska A, Grażyńska A, Szczudło-Chraścina J, et al. Multifocality and Multicentrality in Breast Cancer: Comparison of the Efficiency of Mammography, Contrast-Enhanced Spectral Mammography, and Magnetic Resonance Imaging in a Group of Patients with Primarily Operable Breast Cancer. Curr Oncol. 2021;28(5):4016–30.
Romero Q, Bendahl P-O, Klintman M, Loman N, Ingvar C, Rydén L, et al. Ki67 proliferation in core biopsies versus surgical samples - a model for neo-adjuvant breast cancer studies. BMC Cancer. 2011;11(1):341.
Saghir H, Veerla S, Malmberg M, Rydén L, Ehinger A, Saal LH, et al. How Reliable Are Gene Expression-Based and Immunohistochemical Biomarkers Assessed on a Core-Needle Biopsy? A Study of Paired Core-Needle Biopsies and Surgical Specimens in Early Breast Cancer. Cancers. 2022;14(16):4000.
Larsen M, Aglen CF, Lee CI, Hoff SR, Lund-Hanssen H, Lång K, et al. Artificial Intelligence Evaluation of 122 969 Mammography Examinations from a Population-based Screening Program. Radiology. 2022;303(3):502–11.
Lauritzen AD, Rodríguez-Ruiz A, von Euler-Chelpin MC, Lynge E, Vejborg I, Nielsen M, et al. An Artificial Intelligence–based Mammography Screening Protocol for Breast Cancer: Outcome and Radiologist Workload. Radiology. 2022;304(1):41–9.
Romero-Martín S, Elías-Cabot E, Raya-Povedano JL, Gubern-Mérida A, Rodríguez-Ruiz A, Álvarez-Benito M. Stand-Alone Use of Artificial Intelligence for Digital Mammography and Digital Breast Tomosynthesis Screening: A Retrospective Evaluation. Radiology. 2022;302(3):535–42.
Cen D, Xu L, Zhang S, Zhou S, Huang Y, Chen Z, et al. BI-RADS 3–5 microcalcifications: prediction of lymph node metastasis of breast cancer. Oncotarget. 2017;8(18):30190–8.
Hack C, Häberle L, Geisler K, Schulz-Wendtland R, Hartmann A, Fasching P, et al. Mammographic Density and Prediction of Nodal Status in Breast Cancer Patients. Geburtshilfe und Frauenheilkunde. 2013;73(02):136–41.
Yang J, Wang T, Yang L, Wang Y, Li H, Zhou X, et al. Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer Using Mammography-Based Radiomics Method. Scientific Reports. 2019;9(1):4429.
Tan H, Wu Y, Bao F, Zhou J, Wan J, Tian J, et al. Mammography-based radiomics nomogram: a potential biomarker to predict axillary lymph node metastasis in breast cancer. The British Journal of Radiology. 2020;93(1111):20191019.
Liu Z, Ni S, Yang C, Sun W, Huang D, Su H, et al. Axillary lymph node metastasis prediction by contrast-enhanced computed tomography images for breast cancer patients based on deep learning. Computers in Biology and Medicine. 2021;136:104715.
Cattell R, Ying J, Lei L, Ding J, Chen S, Serrano Sosa M, et al. Preoperative prediction of lymph node metastasis using deep learning-based features. Visual Computing for Industry, Biomedicine, and Art. 2022;5(1).
Wang C, Chen X, Luo H, Liu Y, Meng R, Wang M, et al. Development and Internal Validation of a Preoperative Prediction Model for Sentinel Lymph Node Status in Breast Cancer: Combining Radiomics Signature and Clinical Factors. Front Oncol. 2021;11:754843.
Zhou LQ, Wu XL, Huang SY, Wu GG, Ye HR, Wei Q, et al. Lymph Node Metastasis Prediction from Primary Breast Cancer US Images Using Deep Learning. Radiology. 2020;294(1):19–28.
Qiu X, Jiang Y, Zhao Q, Yan C, Huang M, Jiang T. Could Ultrasound-Based Radiomics Noninvasively Predict Axillary Lymph Node Metastasis in Breast Cancer? J Ultrasound Med. 2020;39(10):1897–905.
Goldhirsch A, Winer EP, Coates AS, Gelber RD, Piccart-Gebhart M, Thürlimann B, et al. Personalizing the treatment of women with early breast cancer: highlights of the St Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 2013. Annals of Oncology. 2013;24(9):2206–23.
Wolff AC, Hammond MEH, Allison KH, Harvey BE, McShane LM, Dowsett M. HER2 Testing in Breast Cancer: American Society of Clinical Oncology/College of American Pathologists Clinical Practice Guideline Focused Update Summary. Journal of Oncology Practice. 2018;14(7):437–41.
Regionala cancercentrum i samverkan. Nationellt vårdprogram bröstcancer 2014 [updated 2022 Aug 23; cited 2023 Feb 10]. Available from: https://kunskapsbanken.cancercentrum.se/diagnoser/brostcancer/vardprogram/.
Keller BM, Nathan DL, Wang Y, Zheng Y, Gee JC, Conant EF, et al. Estimation of breast percent density in raw and processed full field digital mammography images via adaptive fuzzy c-means clustering and support vector machine segmentation. Medical Physics. 2012;39(8):4903–17.
Gastounioti A, Kasi CD, Scott CG, Brandt KR, Jensen MR, Hruska CB, et al. Evaluation of LIBRA Software for Fully Automated Mammographic Density Assessment in Breast Cancer Risk Prediction. Radiology. 2020;296(1):24–31.
Pesek S, Ashikaga T, Krag LE, Krag D. The False-Negative Rate of Sentinel Node Biopsy in Patients with Breast Cancer: A Meta-Analysis. World Journal of Surgery. 2012;36(9):2239–51.
Krag DN, Anderson SJ, Julian TB, Brown AM, Harlow SP, Ashikaga T, et al. Technical outcomes of sentinel-lymph-node resection and conventional axillary-lymph-node dissection in patients with clinically node-negative breast cancer: results from the NSABP B-32 randomised phase III trial. The Lancet Oncology. 2007;8(10):881–8.
Rodriguez-Ruiz A, Lång K, Gubern-Merida A, Broeders M, Gennaro G, Clauser P, et al. Stand-Alone Artificial Intelligence for Breast Cancer Detection in Mammography: Comparison With 101 Radiologists. JNCI: Journal of the National Cancer Institute. 2019;111(9):916 – 22.
Rodríguez-Ruiz A, Krupinski E, Mordang J-J, Schilling K, Heywang-Köbrunner SH, Sechopoulos I, et al. Detection of Breast Cancer with Mammography: Effect of an Artificial Intelligence Support System. Radiology. 2019;290(2):305–14.
Rodriguez-Ruiz A, Lång K, Gubern-Merida A, Teuwen J, Broeders M, Gennaro G, et al. Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? A feasibility study. Eur Radiol. 2019;29(9):4825–32.
Skarping I, Nilsson K, Dihge L, Fridhammar A, Ohlsson M, Huss L, et al. The implementation of a noninvasive lymph node staging (NILS) preoperative prediction model is cost effective in primary breast cancer. Breast Cancer Res Treat. 2022;194(3):577–86.
Bria A, Karssemeijer N, Tortorella F. Learning from unbalanced data: A cascade-based approach for detecting clustered microcalcifications. Medical Image Analysis. 2014;18(2):241–52.

Table 1: Patient and tumor characteristics stratified by sentinel lymph node status

	All (n=770)	N0 (n=501)	N+ (n=269)	P
*Age, years (continuous)^c**	64.7 (24.2-91.9)	65.9 (32.6-91.5)	64.2 (24.2-91.9)	0.027^a
*Pathological tumor size, mm (continuous)^d**	15 (0.5-90)	13 (0.5-55)	18 (0.9-90	<0.001^a
Missing	1	1	0
Mode of tumor detection^c**				<0.001
Symptomatic	321	184 (57)	137 (43)
Screening	449	317 (71)	132 (29)
Multifocality^d**				<0.001^b
Yes	186	96 (52)	90 (48)
No	584	405 (69)	179 (31)
Tumor localization^c**				0.108^b
Central	22	14 (64)	8 (36)
Upper inner	105	76 (72)	29 (28)
Lower inner	46	32 (70)	14 (30)
Upper outer	253	160 (63)	93 (37)
Lower outer	78	41 (53)	37 (47)
Overlapping	266	178 (67)	88 (33)
Histological type^c**				0.046^b
Ductal and lobular	713	457 (64)	256 (36)
Other or mixed	57	44 (77)	13 (23)
Histological grade^c**				0.007^b
I	186	137 (74)	49 (26)
II	350	224 (64)	126 (36)
III	226	133 (59)	93 (41)
Missing	8	7	1
Vascular invasion^d**				<0.001^b
Yes	91	27 (30)	64 (70)
No	526	390 (74)	136 (26)
Missing	153	84	69
Estrogen receptor^c**				0.056^b
Negative	69	53 (77)	16 (23)
Positive	699	446 (64)	253 (36)
Missing	2	0	2
Progesterone receptor^c**				0.077^b
Negative	122	89 (73)	33 (27)
Positive	646	410 (63)	236 (37)
Missing	2	0	2
HER2 status^c**				0.606^b
Negative	624	411 (66)	213 (34)
Positive	84	51 (61)	33 (39)
Missing	62	39 (63)	23 (37)
*Ki67 (continuous) ^c**	15 (0-94)	14 (0-94)	17 (1-81)	<0.001^a
Missing	43	15	28

Abbreviations: Negative sentinel lymph node (N0) status, positive sentinel lymph node (N+) status.
*Median (range)
**Number (%)
^a Mann–Whitney U test
^b Chi-square test

^c Preoperatively available

^d Postoperatively available

Table 3: Univariable logistic regression of radiologic variables for N0 status.

	N 770	OR (95% CI)	P
Radiological tumor size, mm (continuous)	591	0.968 (0.952 – 0.985)	<0.001
Missing	179
Breast density, % (continuous)	748	0.998 (0.991 – 1.005)	0.605
Missing	22
Dense area, cm² (continuous)	748	1.000 (0.996 – 1.004)	0.859
Missing	22
Highest calc cluster score (grouped)	689		0.003
Low	108	1 (reference)
High	135	0.379 (0.215 – 0.669)	0.001
Absence	446	0.564 (0.344 – 0.924)	0.023
Missing	81
Highest calc cluster score (continuous)	689	1.000 (0.996 – 1.003)	0.842
Missing	81
Highest soft tissue lesion score (grouped)	689		0.003
Low	242	1 (reference)
Medium	201	0.716 (0.478 – 1.073)	0.106
High	129	0.440 (0.282 – 0.688)	<0.001
Absence	117	0.897 (0.552 – 1.458)	0.661
Missing	81
Highest soft tissue lesion score (continuous)	689	0.997 (0.992 – 1.001)	0.152
Missing	81
Mammography malignancy score	667		0.514
1-2	34	1 (reference)
3	59	1.089 (0.440 – 2.699)	0.853
4	205	1.103 (0.507 – 2.401)	0.805
5	369	0.849 (0.401 – 1.795)	0.668
Missing	103
Ultrasound malignancy score	660		0.056
1-2	43	1 (reference)
3	32	0.582 (0.204 – 1.661)	0.312
4	132	0.706 (0.308 – 1.617)	0.410
5	453	0.453 (0.212 – 0.968)	0.041
Missing	110

Abbreviations: negative sentinel lymph node (N0) status, odds ratio (OR), confidence interval (CI).

Table 4: Prediction model for N0 Status with variable selection (n=689), threshold for removal p≥0.157

	OR (95% CI)	P
Radiological tumor size, mm (continuous)*	0.970 (0.951 – 0.989)	0.002
Age, years (continuous)	1.024 (1.008 – 1.039)	0.002
Mode of tumor detection		0.001
Symptomatic	1 (reference)
Screening	1.875 (1.301 – 2.703)
Histological type		0.068
Ductal or lobular	1 (reference)
Other or mixed	1.926 (0.953 – 3.900)
Estrogen receptor status*		0.001
Positive	1 (reference)
Negative	3.396 (1.655 – 6.968)
Tumor localization		0.289
Upper outer	1 (reference)
Central	1.010 (0.382 – 2.669)	0.984
Upper inner	1.634 (0.938 – 2.847)	0.083
Lower inner	1.013 (0.485 – 2.118)	0.973
Lower outer	0.694 (0.384 – 1.254)	0.226
Overlapping	1.141 (0.754 – 1.725)	0.533
Highest calc cluster score (grouped)		0.076
Low	1 (reference)
High	0.496 (0.270 – 0.909)	0.023
Absence	0.638 (0.377 – 1.080)	0.094
Highest soft tissue lesion score (grouped)		0.091
Low	1 (reference)
Medium	0.722 (0.470 – 1.110)	0.137
High	0.588 (0.361 – 0.958)	0.033
Absence	1.049 (0.622 – 1.769)	0.856
Constant	5.289

Abbreviations: negative sentinel lymph node (N0) status, odds ratio (OR), confidence interval (CI).
*Imputed

No competing interests reported.

Table 2 not available with this version.

A preoperative prediction model for sentinel lymph node status using artificial intelligence on mammographic images and clinicopathological variables in patients with clinically node-negative breast cancer

Status:

Version 1

Abstract

Background:

Methods:

Results:

Conclusion:

Figures

Background

Methods

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1