To study the effect of cut-point on biomarker performance of breast MRI for prediction of pathologic complete response (pCR) after one-cycle of neoadjuvant chemotherapy (NACT) and to optimize the cut-point by clinically-relevant breast cancer subtype.

Two hundred and thirty seven patients with tumors ≥3cm were consented and enrolled in the multi-center ACRIN 6657/I-SPY 1 clinical trial. Participants received anthracycline-cyclophosphamide (AC) alone or followed by a taxane (T)-based regimen before surgery. pCR was defined as having no invasive tumor present in either breast or axillary lymph nodes at the time of surgery. Functional tumor volume (FTV) was calculated from DCE-MRI by summing all voxels with an early enhancement exceeding 70% from pre-contrast within a manually defined volume of interest. Early FTV response (%ΔFTV₂) was defined as percent change in FTV from MRI₁ (before NACT) to MRI₂ (after one cycle of AC). % ΔFTV₂ greater/less than a given cut-point were considered test-negative/test-positive (Figure1). The range of the cut-point was from -100% (FTV2=0cc) to 100% (doubled FTV) after one cycle of AC. For this work we investigated the effect of changing cut-point on positive and negative predictive value (PPV/NPV) as described in the contingency table shown in Figure 1.

For any non-perfect test, the trade off between false negatives (FN) and false positives (FP) can be considered when choosing an optimum cut-point. This can be handled by defining and minimizing a penalty function based on the probabilities of false test results. We investigated the penalty function based on FN and FP that can be expressed as:

$$$Penalty=\frac{1}{N}(\beta \times FP+(1-\beta )\times FN) (1)$$$

where FP is the number of false positives, FN the number of false negatives, N is the number of patients in the study cohort, and β is a user-specified weight factor between 0 and 1 determining the relative importance of avoiding FP and FN results. To simulate a treatment response evaluation setting we chose β =0.2 to more strongly penalize false negative tests, which could result in unnecessary or harmful treatment changes.

Breast cancer subtype was defined by hormone receptor (HR) and human epidermal growth factor receptor 2 (HER2) status. The full cohort was divided into three subsets: HR+HER2-, HER2+, and HR- HER2- (TN = triple negative).

One hundred and forty two patients with %ΔFTV₂ measurement and pCR status were included in this study. Penalty values vs. all cut-points tested are plotted in Figure 2. The number of patients and pCR rates in the full cohort and in subtypes are listed in Figure 3, in which the optimal cut-point for the penalty calculation was shown as -22%, along with optimal cut-points for subtypes and comparison to that from the full cohort by calculating NPVs. For instance, when cut-point was set to be optimal (-77%) in the HR+HER2- subtype, the NPV was 94% (CI: 82−99%); whereas it was 95% (CI: 74−100%) when using the full cohort optimized cut-point of -22%. The cut-point of -22% led to 18 out of 50 non-pCRs being tested negative while the cut-point of -77% resulted in 44 out of 50 non-pCRs being tested negative. NPVs/PPVs and number of pCRs/non-pCRs predicted correctly using corresponding cut-points are plotted in Figure 4.The NPV of the early percent change of FTV was improved by choosing a cut-point by minimizing the penalty which incorporates proportions of false negatives and false positives. In this study, the weight β was chosen to be 0.2 after a limited number of tests. A better strategy is needed to estimate the optimal value of β. HR+HER2- showed distinctive trend and optimal cut-point from other two subtypes, which could be related with the small pCR rate (6 out of 56 patients). This study investigated the effect of varying cut-points for percent change between volume MRI₁ and MRI₂ on the ability of MRI to predict patients with residual disease. Cut-point analysis demonstrated tradeoffs between false negatives and false positives and these tradeoffs substantially differed in the HR+HER2- subtype. This work is ongoing and will consider alternative clinical outcomes such as residual cancer burden (RCB), a histopathologic endpoint that further stratifies extent of residual disease.