Abstract P3-03-24: Machine learning approach to predict the level of tumor-infiltrating lymphocytes of breast cancer via MRI-based radiomics

Abstract

Abstract Background: The level of tumor-Infiltrating lymphocytes (TILs) has important guiding significance for the individual treatment of breast cancer. In clinical, it is necessary to obtain tumor samples via invasive biopsy for pathological assessment. Our study aimed to investigate the feasibility of using machine learning algorithms to predict the level of TILs based on Multi-parametric MRI-radiomics in breast cancer. Materials and Methods: 207 patients with histologically-proven breast cancer (including 81 patients with HR+/HER2- subtype, 95 patients with HER2+ subtype, and 31 patients with TN subtype) who underwent breast MRI from January 2017 to March 2021 were retrospectively retrieved and analyzed. A total of 4198 quantitative radiomics features were extracted from the tumor regions of interest (ROI) on multi-parametric MRI, including T1-weighted Dynamic Contrast-Enhancement (T1-DCE) images, Fat-suppressed T2-weighted images (FS-T2WI) and Apparent Diffusion Coefficient (ADC) map using the Pyradiomics software. Feature selection was then performed using the least absolute shrinkage and selection operator (LASSO) and Boruta algorithm to identify the radiomics features most relevant to the level of TILs. Six machine learning models, including Logistic Regression (LR), Support Vector Machine (SVM), Random Forests (RF), Linear Discriminant Analysis (LDA), Gaussian Bayes (GNB) and Multilayer Perceptron (MLP) were developed with leave-one-out cross-validation (LOOCV) for predicting the level of TILs in breast cancer. The predictive performance of those classifiers for the level of TILs was evaluated through the receiver operating characteristic (ROC) curves, the area under curve (AUC), accuracy, sensitivity and specificity in the testing set. Results: Among the 207 patients, 94 (45.4%) showed high TILs levels (≥30%), while 113 (54.6%) showed low TILs levels (<30%). There was no inter-group distribution difference of clinicopathologic characteristics (including age, clinical TNM stage of tumor, ER, PR, HER2 and Ki-67 index) between the low TILs group and the high TILs group. After the feature selection step, 11 top-class features (including 5 features from T1-DCE images, 3 features from FS-T2WI, and 3 features from ADC-map) were selected to develop the radiomics models. The MLP classifier achieved the most excellent predictive performance in the testing dataset [AUC: 0.865, 95% confidence interval (CI), 0.788-0.942, with an accuracy of 82.8%]. Moreover, the Rad-score in the high TILs group was higher than that in the low TILs group (p = 0.014). We also evaluate the model performances in different molecular subtypes, and the MLP classifier had the best performance in HER2+ subtype with an AUC of 0.916 and an accuracy of 86.9%. While for HR+/HER2- subtype and TN subtype, the AUCs were 0.848 and 0.781, and the accuracies were 81.3% and 73.7%, respectively. Conclusion: Multi-parametric MRI-based radiomics model could provide a non-invasive tool to predict the TILs levels of breast cancer, which might help to guide clinical decision-making in individual treatment for patients with breast cancer. Citation Format: Yuhong Huang, Ying Lin. Machine learning approach to predict the level of tumor-infiltrating lymphocytes of breast cancer via MRI-based radiomics [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 P3-03-24.