Introduction

Irinotecan (IRI) is used as salvage treatment in advanced metastatic breast cancer (mBC) patients to prolong the survival¹. Liposomal formulation of irinotecan (nal-IRI) is used to increase tumor exposure to IRI². Ferumoxytol (FMX) is an ultra-small iron oxide nanoparticle (USPIO) with unique temporal biodistribution producing in varied MRI enhancement. During the first few hours after infusion, FMX resides in the vascular space producing vascular enhancement. After approximately 24 hours, most of the FMX that extravasates from leaky vasculature is phagocytosed by local macrophages, resulting in delayed enhancement on MRI ^3,4. FMX has previously been demonstrated to colocalize with tumor-associated macrophages (TAMs) in tumors^5-7, which can modulate lesion permeability characteristic and may be mechanistically relevant to nal-IRI anti-tumor activity. Phagocytosis of FMX by TAMs results in intracellular compartmentation, which has opposing effects on longitudinal $$$(R_1)$$$ and transverse relaxivity $$$(R_2^*)$$$of FMX⁸. Mismatch between apparent ferumoxytol concentration ($$$FMX_C$$$) computed from changes in pixelwise $$$R_1$$$ and $$$R_2^*$$$ post-FMX can be exploited to identify pixels with high concentrations of phagocytosed (compartmentalized) FMX. Here we investigate such a measure of FMX compartmentation, such as may result from phagocytosis of the FMX, as a companion imaging biomarker for predicting response of breast cancer brain metastases to nal-IRI taking advantage of robust co-registration across imaging sessions.

Methods and Materials:

The study schema is shown in Figure 1. A series of 6 fat suppressed spoiled gradient echo (SPGRE) scans were acquired at 1.5 T (or 3.0 T) using the following parameters: TE = 1.5(2.2), 3.0(4.5), 4.5(6.8), 6.0(9.0), 9.0(11.3), and 13.2(13.5) milliseconds, slice thickness/slice gap = 5/0 mm, and matrix size = 256 X 256 with a field-of-view to cover the whole head along with two phantom tubes containing known concentrations of FMX.
Quantification of FMX compartmentation:
In our study, we did not acquire $$$R_1$$$ maps to limit the total scan duration for reasons of patient comfort. Instead, for each patient, we used a modified SPGRE signal equation to calculate the R1-weighed post-FMX image that would be expected for a given $$$FMX_C$$$ (eq. [3]) and compared that to the R1-weighted component that was computed from the shortest TE post-FMX image that was acquired. This approach is described in equations 1-7 and Figure 2, as follows:

$$$S\;=\;M_0*\frac{(1-e^{-TR*R_1})}{(1-cos(α)*e^{-TR*R_1})}*sin(α)*e^{-TR*R_2^*}[1]$$$

$$$R_1\;=\;R_{1,0} +FMX_{c,brain}*r_{1,FMX}\;[2] $$$

$$$FMX_C\;=\;\frac{R_{2,postFMX}^* - R_{2,preFMX}^*}{r_{2,postFMX}^*}\;[3]$$$

Combining eq. [1] and [2], the signal equation can be written in terms of $$$R_1$$$ as in eq. [4]:

$$$S_{Computed,i}\;=\;\frac{S}{sin(α)*e^{-TE*R_2^*}} = M_0 * \frac{(1-e^{-TR*(R_{1,0}+FMX_{C,i}*r_{1,FMX})})}{(1-cos(α)*e^{-TR*(R_{1,0}+FMX_{C,i}*r_{1,FMX})})}\; [4]$$$

The subscript ‘i’ in the eq. [4] corresponds to the P0, P1, and P2 imaging sessions (details in Figure 2). Receiver gain differences between FMX sessions were corrected by linear scaling correction to signal intensities from the two phantom tubes that were present on all scans.
Our model requires a pre-FMX and two post-FMX imaging scans to properly resolve FMX enhancement from the vasculature and intracellular compartmentation. Non-linear regression is used to fit for the parameters $$$R_{1,0}$$$ and $$$M_0$$$ in eq. [4]. The fitted parameters are then used to calculate fitted R1, and expected R1 weighted signal as shown in eq. [5] and [6].

$$$R_{1,I,fitted}\;=\;R_{1,0,fitted} + FMX_{c,i}*r_{1,FMX}\;[5]$$$

$$$S_{Expected,i}\;=\;M_{0,fitted}*\frac{(1-e^{-TR*R_{1,I,fitted}})}{(1-cos(α)*e^{-TR*R_{1,I,fitted}})}\;[6]$$$

A compartmentation index (CI) was computed as the ratio of $$$S_Expected,i$$$ and $$$S_Computed,i$$$ as shown in eq. [7]: values above unity would indicate some degree of extravascular intracellular compartmentation.
$$$CI = (\frac{S_{Expected,16-24 hrs PostFMX}}{S_{Computed,16-24 hrs PostFMX}})\;[7]$$$

Imaging data from a total of 19 tumors in 7 patients were analyzed. A tumor was classified as non-responding tumor (NRT; 10 tumors) if it did not exhibit PR (diameter decrease less than 30%) or exhibited progression (change from nadir is greater than 20 %⁹), and classified as responding tumor (RT; 9 tumors) otherwise.
From the CI maps, we computed two metrics for each brain tumor: (i) the fraction of pixels within a tumor VOI with CI greater than 1 $$$(FCIGT1)$$$, and, (ii) the fraction of pixels within a tumor VOI with CI greater than 1 and $$$R_2^*$$$ in 90^th and higher percentile $$$(FCIGT1R_{2,90}^*)$$$. Area-under-the-receiver-operator-curve (AUROC) for the task of discriminating RT and NRT tumors was computed. Leave-one-out-cross-validation (LOOCV) was performed.

Results and Discussion:

Figure 3 illustrates the computed and expected $$$R_1$$$ weighted maps, CI, and $$$R_2^*$$$ maps for three example tumors (rows). Qualitatively, on the CI maps, the responding tumor (top row) had a higher distribution of $$$FCIGT1$$$ voxels than a tumor that initially responded but re-grew (middle row) and a progressing tumor (middle row), and these differences were less obvious on $$$R_2^*$$$ maps. The ROC curves for the 3 putative discriminators are plotted in Figure 4, and corresponding AUROC values are tabulated in Table 1. The CI metrics provided good discrimination of RT vs. NRT tumors with RT tumors having higher pre-treatment values of $$$(FCIGT1R_{2,90}^*)$$$ than NRT tumors. AUROC of $$$FCIGT1$$$, and $$$(FCIGT1R_{2,90}^*)$$$ 0.75 and 0.86 respectively. $$$(FCIGT1R_{2,90}^*)$$$ had good accuracy on both training (84%) and LOOCV (79%).

Conclusion:

$$$R_1$$$ and $$$R_2^*$$$ mismatch on MRI acquired 24 h post-FMX, measured in terms of CI greater than 1 and hypothesized to be indicative of FMX intracellular compartmentation, was predictive of the response of breast cancer brain metastases to nal-IRI at the individual tumor level. In particular, the metric FCIGT1 is highly predictive of response in both training and LOOCV.

Acknowledgements

No acknowledgement found.