We present a radiomics model to discriminate between patients at low risk and those at high risk of treatment failure at long-term follow-up based on eigentumors: principal components computed from volumes encompassing tumors in washin and washout images of pre-treatment dynamic contrast-enhanced (DCE-) MR images. Eigentumors were computed from the images of 563 patients from the MARGINS study. Subsequently, a least absolute shrinkage selection operator (LASSO) selected candidates from the components that contained 90% of the variance of the data. The model for prediction of survival after treatment (median follow-up time 86 months) was based on logistic regression. Receiver operating characteristic (ROC) analysis was applied and area-under-the-curve (AUC) values were computed as measures of training and cross-validated performances. The discriminating potential of the model was confirmed using Kaplan–Meier survival curves and log-rank tests.
From the 322 principal components that explained 90% of the variance of the data, the LASSO selected 28 components. The ROC curves of the model yielded AUC values of 0.88, 0.77 and 0.73, for the training, leave-one-out cross-validated and bootstrapped performances, respectively. The bootstrapped Kaplan–Meier survival curves confirmed significant separation for all tumors (P < 0.0001). Survival analysis on immunohistochemical subgroups shows significant separation for the estrogen-receptor subtype tumors (P < 0.0001) and the triple-negative subtype tumors (P = 0.0039), but not for tumors of the HER2 subtype (P = 0.41). The results of this retrospective study show the potential of early-stage pre-treatment eigentumors for use in prediction of treatment failure of breast cancer.