Prediction of node negative breast cancer and high disease burden through image analysis software on mammographic images and clinicopathological data

Abstract Background/Aim: Sentinel lymph node biopsy is the standard staging procedure performed on all clinically node negative breast cancer patients, although more than 70% have no lymph node metastases and would not benefit from the procedure. Thus, a preoperative method of predicting lymph node status is warranted. Recently, artificial intelligence-based breast malignancy detection systems for mammograms have been developed, detecting both lesions and microcalcification. Previous models based on clinicopathological data, proposed by Dihge et al.1, 2, received an area under the curve (AUC) of 0.74 for prediction of negative lymph node status and an AUC of 0.75 for prediction of high-burden disease. In this study we aim to predict node negativity (N0) and high-burden disease (N2) in breast cancer patients through mammographic features captured by image analysis software on mammography, added to previous prediction models. Method: This is a retrospective cohort study including 770 women with unilateral breast cancer operated at Lund University Hospital 2009-2012. Mammographic images were identified for 755 women and analyzed by two image analysis software applications, Transpara and Laboratory for Individualized Breast Radiodensity Assessment (LIBRA). Transpara findings were cross-checked with Picture Archiving and Communication System for tumor localization. Clinicopathological variables, soft tissue lesion scores, calc cluster scores, breast density, malignancy score and radiologic tumor size were collected. Prediction models were created using multivariable logistic regression. AUC assessed the performance of the models to predict N0 and N2. Results: Univariable logistic regression showed an association between axillary lymph node status and radiologic size, highest score of soft tissue lesion, highest score of calc cluster and malignancy score. Addition of highest score of soft tissue lesion and highest score of calc cluster to the previously published model1 for prediction of N0 resulted in an AUC of 0.75 (confidence interval (CI) 0.70-0.79). Addition of highest score of calc cluster to a modified version of the previously published model2 for prediction of high-burden disease (N2 versus N0 and N1) resulted in an AUC of 0.83 (CI 0.75-0.90). A comparison of pathologic and radiologic tumor size showed a strong correlation between the variables and associations with lymph node status. Mammographic density was not associated to nodal status. Conclusion: The prediction models proposed in this abstract, including radiomic features, did not significantly improve the previous clinicopathological models1,2. Nevertheless, point estimates of the AUCs were improved by 0.01 and 0.08, respectively, indicating that radiomic features could be of added value in the prediction of node-negative disease and high nodal disease burden and should be further investigated. The strong correlation between measurements of tumor size suggests that radiologic tumor size could replace pathologic size to enable preoperative prediction of nodal status. 1 Dihge L, Bendahl PO, Ryden L. Nomograms for preoperative prediction of axillary nodal status in breast cancer. Br J Surg 2017;104:1494-5. 2 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 Jun 21;19(1):610. doi: 10.1186/s12885-019-5827-6 Table: Multivariable logistic regression of prediction models.Predictions modelsPrediction of negative lymph node status. (n=560)Prediction of high axillary disease burden. (n=576)OR (95% CI)POR (95% CI)PTumor size mm (continuous)0.950 (0.926-0.974)<0.001Tumor size mm (continuous)1.062 (1.019-1.107)0.005Multifocality0.006Multifocality0.655Yes1.00 (ref)Yes1.00 (ref)No1.89 (1.20-2.98)No0.82 (0.34-1.98)Lymphovascular invasion<0.001Lymphovascular invasion<0.001Yes1.00 (ref)Yes1.00 (ref)No3.97 (2.25-7.02)No0.20 (0.08-0.48)Age years (continuous)1.022 (1.004-1.040)0.015PR status0.122Mode of detection0.019Positive1.00 (ref)Symptomatic1.00 (ref)Negative0.35 (0.09-1.33)Mammographic screening1.70 (1.09-2.66)Histological type0.857Molecular subtypes0.027aLobular1.00 (ref)Luminal A1.00 (ref)Other0.89 (0.25-3.21)Luminal B HER2 negative1.20 (0.76-1.91)0.452Tumor localization0.716aLuminal B HER2 positive1.00 (0.48-2.09)0.995Upper outer1.00 (ref)HER2 positive2.00 (0.47-8.59)0.352Central1.29 (1.13-13.0)0.830Triple negative5.68 (1.95-16.5)0.001Upper inner1.40 (0.32-6.12)0.656Highest score of soft tissue lesion0.283aLower inner1.92 (0.35-10.7)0.4561-791.00 (ref)Lower outer2.79 (0.77-10.2)0.12080-911.40 (0.65-3.03)0.393Overlapping1.96 (0.69-5.60)0.20992-951.13 (0.53-2.42)0.745Highest score of calc cluster0.456a≥960.83 (0.37-1.88)0.6611-921.00 (ref)Absence1.69 (0.74-3.87)0.216≥932.86 (0.55-14.9)0.212Highest score of calc cluster0.510aAbsence2.20 (0.46-10.4)0.3231-921.00 (ref)≥930.64 (0.31-1.36)0.246Absence0.78 (0.42-1.44)0.425Constant0.1700.026Constant0.0230.003AUC0.75 (0.70-0.79)AUC0.83 (0.75-0.90)P of ≤0.05 was considered significant. Abbreviations: OR =odds ratio, CI = confidence interval, HER2 = human epidermal growth factor receptor 2. a Test of overall effect Citation Format: Cornelia Rejmer, Looket Dihge, Pär-Ola Bendahl, Daniel Förnvik, Magnus Dustler, Lisa Rydén. Prediction of node negative breast cancer and high disease burden through image analysis software on mammographic images and clinicopathological data [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P1-01-09.