DCE-MRI-derived microvasculature mapping and outcome prediction in patients with breast cancer.

Background: Tumor response to neoadjuvant therapy (NAT) in patients with breast cancer is related to perfusion in the tumor microenvironment and surrounding tissue. Direct and indirect measures of microvascular (MV) permeability could improve the ability to predict short- and long-term patient outcomes. Detection and quantification of circulating tumor DNA (ctDNA) is one such measurement that depends on microvasculature and has been shown to strongly correlate with both response to therapy and overall prognosis in a wide range of cancers. We hypothesized that a combination of ctDNA and numerical computations of MV permeability using pretreatment dynamic contrast-enhanced (DCE) MRI images (MV maps) can predict patient response to NAT and long-term survival outcomes. Here, we show how MV maps — computed from standard-of-care (SOC) imaging — improve upon ctDNA measurements in predicting pathologic complete response (pCR) following NAT as well as distant recurrence free survival (DRFS).

Methods: DCE-MRIs were obtained for a sub-sample of I-SPY2 patients (n = 81) whose ctDNA levels were previously reported (Magbanua et al 2021). Tumors were segmented from pretreatment images by a convolutional neural network with manual verification. MV maps were computed using a Tofts-like model to describe the transfer kinetics between microvasculature and extracellular extravascular space. Microvasculature parameters (MVPs), such as median and skewness, were extracted from MV maps to describe the distribution of kinetic parameters in the intra- and peritumoral regions. Logistic regression was used to predict pCR with leave-one-out cross validation. Univariate Cox models were used to stratify DRFS.

Results:  MVPs exhibited a moderate correlation with pretreatment ctDNA levels (Spearman rho = -0.26; p = 0.018) and with pCR status in this cohort (rho = 0.25; p = 0.026). In a series of logistic regression models, a combination of SOC clinical features and ctDNA levels correctly classified 23% of true pCRs, with an F1 score of 0.58. This predictive power increased when informed by MV maps; 45% of true pCRs were identified after the addition of MVPs, and the F1 score increased to 0.68. In univariate Cox models, 24 MVPs significantly stratified DRFS (p < 0.05) when uncorrected for multiple hypothesis testing bias. The top MVP carried a hazard ratio 3.8 (95%CI 1.6 – 9.0) compared to 11.9 (95%CI 2.7 – 53.6) for ctDNA levels.

Conclusions: DCE-MRI-derived kinetic features are calculated from SOC images with no additional clinical intervention. We found that MVPs substantially improve pCR prediction when added to models trained on ctDNA levels and clinical features. ctDNA levels and MVPs were predictive of DRFS to differing degrees of confidence. Although further validation in larger datasets is needed, this exploratory analysis uncovered a promising role for MV maps in predicting response to NAT as well as patient outcomes.

View publication here.