INTRODUCTION

Different tumor micro-environments are the result of Darwinian dynamics, may be discernible by MRI, and are referred to as habitats¹. In order to study these, it is necessary to co-register MRI and histological images, which is an unresolved problem due to the complexity of the histology process². In the last few years, 3D-printed molds have been used to aid the image co-registration, but most of the work has been done in human prostate slices of 5 to 8mm^3-5. In this work, a co-registration framework was developed using 3D-printed tumor-molds created from in-vivo MRI images of five mice with implanted breast cancer tumors. Tumors were sectioned in 2mm slices to improve accuracy of the co-registration, compared to those used before^3-5. Preliminary co-registration and habitats results are shown and discussed. This work will help understanding MRI habitats to monitor cancer evolution as a means to aid treatment decisions in the future.

METHODS

All MRI experiments were done in a 7T horizontal magnet (Agilent-Technologies) and Bruker electronics (BioSpec AV3HD), using a 35mm birdcage coil(Doty Scientific). Six multi-parametric MRI images were acquired to determine tumor habitats using a Gaussian mixture model fit with an expectation maximization algorithm in MATLAB(Mathworks, Inc.). Using four clusters, the fit was performed on all pixels within the tumor volume of interest. Tumor contours were segmented in MATLAB using axial T₂-weighted images and exported as a 3D-tumor-surface (fig.1.a.b.c). Tumor molds were designed in SolidWorks (SolidWorks Corp.) by carefully aligning the 3D-tumor-surface with the corresponding cutting slots (fig.1.d). Tissue marking dyes (Cancer Diagnostics, Inc.) were used to aid tumor orientation in the mold(fig.1.e). Tumors were fixed in Pen-Fix (Thermo Scientific™) for 1 hour inside the mold and imaged in a microCT (Inveon, Siemens Corp.) to verify the accuracy of the mold (fig1.f). All tumors slices were cut at once using a designed multi-blade knife consisted of 11 bevel HD Microtome Blades (Cancer Diagnostics, Inc.) spaced by 2mm(fig.1.g). Tumor slices were stained with H&E and with a hypoxia marker (pimonidazole) by immunohistochemistry (IHC). Histological slices were scanned (Aperio AT2, Leica Biosystems) (fig.1.j) and loaded in MATLAB for co-registration with MRI(fig.1.k). The multi-blade holder and tumor cassettes were printed in a 3D-printer Object24 (Stratasys Ltd) with 0.022mm vertical resolution. Optimal 3D-alignment of MRI images and histological slices was estimated using non-monolithic optimization procedure consisted of outer and inner optimization problems. In each step, the outer optimization loop rotates and shifts the 3D-tumor surface reconstructed from MRI images and calculates contours obtained by its intersection with equally distant planes to mimic tumor slicing (fig.2.a). Those contours were used by inner optimization loop to find rotation of corresponding contours from histological slices that minimize the pixel-wise Euclidean distance between them. Both optimization loops utilized gradient-based algorithms implemented in MATLABs lsqnonlin function.

RESULTS AND DISCUSSION

The tumor molds were accurate with no apparent gaps as demonstrated by microCT imaging (fig.1.f). The multi-blade knife(fig.1.g) permitted more accurate tumor slicing comparing to previous experiments using a single blade. Three-dimensional aligning of MRI images and histological slices was performed for all five mice. The area identified as a tumor by both MRI outline and aligned histology covered on average 80.6% ± 7.6% of the total area, that is area classified as a tumor by MRI outline or aligned histology. Some slices showed better co-registration than others(fig.2.b), this is expected due to different deformations of tissue during the fixation process. Representative multi-parametric images of a tumor are shown in Figure 3 with its correspondent habitats clustering and co-registered histological images. Tumor habitats clustering (fig.3.f) were similar to the hypoxic pattern observed in DCE-images and slices stained with pimonidazole (fig.3.h). These results are encouraging, and show that there is a tradeoff in how the various parameters are normalized in order to reflect image features of each map. Normalization was performed such that habitat clustering matched histological features.

CONCLUSIONS

We have designed a complete workflow for the co-registration of multi-parametric MRI and histology. The results of automated 3D-alignment of MRI images and histological slices are promising and encourage further experiments using the presented workflow. Future improvements in co-registration will include a non-rigid algorithm using tumor features obtained from each parametric image instead of using just the tumor outline. Tumor habitats clustering showed encouraging results with similarities to the hypoxic pattern observed by IHC. This work will serve as a platform to assist the understanding of cancer MRI habitats and its correlation with histological images. Understanding imaging habitats will help monitor cancer evolution better and aid treatment decisions in the future without the need of invasive biopsies.