Key points are not available for this paper at this time.
571 Background: The recurrence risk (RS) based on transcriptomic profiles or Ki-67 level of HR+ EBC has been used for adjuvant treatment decisions and may also aid selecting patients who could benefit from novel treatments including CDK4/6 inhibitor or oral selective estrogen receptor degraders (SERDs). Artificial intelligence-powered H&E whole slide image (WSI) models can be a practical approach for this biomarker without the need of additional tissue samples. We developed ML-based models of histopathologic features from an H&E WSI to approximate RS of 21-gene assay (OncotypeDx) using various algorithms of different complexities. Methods: The development dataset was composed of 1,009 cases from TCGA-BRCA with gene expression profiles and 483 EBC cases from Samsung Medical Center with RS results. The prediction model was developed in two-stages. First, Lunit SCOPE, an AI-powered H&E WSI analyzer, detects various cell types and segments tissue areas resulting in various cell densities in tissue regions of interest. Next, four traditional ML models were applied on top of the first stage results. In addition, a deep learning model based on multiple instance learning (DL-MIL) was developed using three features: a supervised convolutional neural network, a supervised cell detection model, and an AI-based pathology profiling analyzer, for extracting semantic contents. The model performance was validated using two independent EBC test sets: 248 cases for comparison with the RS results (set 1), and 708 cases with DFS (set 2). Results: The cross-validation performance of the four traditional ML algorithms had area under the curve of receiver operating characteristics curve (AUROC) ranging from 0.728 to 0.779, where mitotic cell density, tumor cell density, and fibroblast density in cancer area were the variables with the highest importance for all four models. The DL-MIL model had cross validation performance of AUROC 0.831. In test set 1, DL-MIL model showed the highest discrimination with AUROC of 0.828 compared to traditional ML models where logistic regression showed the highest discrimination with AUROC of 0.786 (Table). The DL-MIL model showed the highest performance among the models to predict DFS with hazard ratio (HR) of 2.48 (1.47-4.18). Conclusions: Histopathomic models can accurately predict the RS of high risk as well as poor DFS from only H&E WSIs. These AI models can be a practical tool for treatment selection including emerging drugs such as SERDs. Table: see text
Moon et al. (Wed,) studied this question.