479P - Radiomics to predict HER2 status in breast cancer brain metastases

Background

Significant discordance rates in HER2 status between primary breast cancer (BC) and brain metastases (BCBM) have been reported, potentially affecting clinical decision-making. Biological recharacterization of BCBMs is difficult due to inherent neurosurgery risks. We here assess the potential of radiomics to non-invasively assess HER2 status of BCBMs.

Methods

We retrospectively identified BC patients from 2 centers, for which HER2 status evaluated on BCBMs (HER2-positive versus HER2-negative by ASCO-CAP) and preoperative brain MRI were available. A total of 218 radiomics features, complying with Imaging Biomarker Standardization Initiative feature definition, were extracted from manually segmented BCBMs using Python. Robust features (selected using LASSO) associated with HER2 status in univariate analysis (p<0.05) were tested in multivariate analysis. Linear regression model (LRM) and non-linear logistic regression model (NLRM) were constructed. Machine learning (ML) algorithms (decision tree, k-nearest neighbor, and support vector machine [SVM]), were tested, by spliting the dataset into 70% for training, 20% for testing, and 10% for validation. Classifiers were trained and tested using the cross-validation approach 10 times; median values for AUC, accuracy, sensitivity and specificity were obtained. The performance of the best classifiers was evaluated on the validation dataset. Analyses were performed using MATLAB R2021b Statistics and Machine Toolbox (MathWorks).

Results

We identified 54 BC patients, for which HER2 status evaluated on 56 BCBMs (HER2+ N=27, HER2- N=29) and preoperative brain MRI was available. At univariate analysis, 9 features were significantly associated with HER2 status, but diagnostic accuracy was low (maximum 67%). Using LRM and NLRM, including the 9 significant features at univariate analysis, accuracy increased to 79%. Using the ML approach, the best classifier among the pattern recognition approaches tested was a SVM including the 9 significant features, which reached an AUC of 0.97 and an accuracy of 91% and 86% on test and validation set, respectively.

Conclusions

Radiomics using machine learning approach shows potential to non-invasively predict HER2 status of BCBMs.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

University of Padova - Department of Surgery, Oncology and Gastroenterology.

Funding

University of Padova - 2021 STARS@UniPD Grant.

Disclosure

G. Griguolo: Financial Interests, Personal, Invited Speaker: Novartis, Eli Lilly; Financial Interests, Personal, Advisory Board: Gilead; Other, Travel Support: Novartis, Amgen, Daiichi Sankyo, Eli Lilly, Gilead; Other, Trave Support: Pfizer. W. Jacot: Financial Interests, Personal, Advisory Board: AstraZeneca, Eisai, Novartis, Roche, Pfizer, Eli Lilly, MSD, BMS, Chugai, Seagen, Daiichi Sankyo; Financial Interests, Personal, Invited Speaker: AstraZeneca, Pfizer, Seagen, Daiichi Sankyo; Financial Interests, Institutional, Research Grant: AstraZeneca, Daiichi Sankyo; Financial Interests, Coordinating PI: Roche, Daiichi Sankyo; Financial Interests, Local PI: Roche, Novartis, Daiichi Sankyo. M.V. Dieci: Financial Interests, Personal, Invited Speaker: Eli Lilly, Exact sciences, Gilead, Seagen, Daiichi Sankyo, Novartis; Financial Interests, Personal, Advisory Board: Novartis, Eli Lilly, Seagen, Exact Science, Daiichi Sankyo, Gilead; Financial Interests, Personal, Other, Consultancy: Pfizer; Financial Interests, Personal, Other, Consultancy on educational project: Roche. A. Darlix: Financial Interests, Institutional, Advisory Board: Servier. V. Guarneri: Financial Interests, Personal, Invited Speaker: Eli Lilly, Novartis, Amgen, GSK; Financial Interests, Personal, Advisory Board: Eli Lilly, Novartis, MSD, Gilead, Eli Lilly, Sanofi, Merck Serono, Exact Sciences, Eisai, Olema Oncology; Financial Interests, Institutional, Local PI: Eli Lilly, Roche, BMS, Novartis, AstraZeneca, MSD, Synton Biopharmaceuticals, Merck, GSK, Daiichi Sankyo, Nerviano; Non-Financial Interests, Member: ASCO. All other authors have declared no conflicts of interest.