Hassan Bagher-Ebadian1,2, Ana deCarvalho1, Tavarekere Nagaraja1, Azimeh NV Dehkordi3, Susan Irtenkauf1, Swayamparva Panda1, Robert Knight1, and James R Ewing1,2
1Henry Ford Hospital, Detroit, MI, United States, 2Oakland University, Rochester, MI, United States, 3Shahid Beheshti University, Tehran, Iran
Synopsis
In
human glioblastoma multiforme (GBM), infiltrating cells are found in remote
locations, even in the hemisphere contralateral to the primary lesion. Although MRI allows approximation of the
extent of tumor cell infiltration, the actual extent of infiltration may be
greater or less than the edema, and there is no standard MRI practice that can
be used to assess the infiltrating tumor burden. This pilot study investigates
the feasibility of using a set of MR modalities for the development of an MRI
estimate of infiltrating tumor burden in Patient-Derived Mouse Xenografts model
of GBM using an adaptive model.Introduction:
Infiltrating Glioblastoma Multiforme (GBM) cells typically move fastest
along existing structures in brain – vessels and white-matter tracts
1. Matrix metalloproteinases (MMPs) and/or ADAMs
(A-Disintegrin-And-Metalloproteinase) are up-regulated in GBM and are employed
to overcome structural barriers in the extracellular extravascular matrix (EEM)
2,3. Lysis of the EEM by MMPs is
accompanied by inflammation, and in the brain is accompanied by alteration of
the permeability of the Blood-Brain-Barrier(BBB) to small molecules
3. However, it is a matter of common experience
that, in regions with significant infiltration, even Gd-DTPA(molecular-weight
512) does not penetrate the BBB in any detectable amount over a period of tens
of minutes. Unfortunately, in humans the
central lesion, with its avid angiogenesis and accompanying leakage of vascular
proteins and copious amounts of water, produces an extensive edema that masks
subtle changes in local vascular permeability that accompany the presence of
infiltrating tumor cells. While any one mode of MRI imaging (diffusion-T1-
and T2-weighted, post-contrast-T1-weighted,dynamic
studies) fails to discriminate tissue with infiltrating tumor cells,
preliminary investigations demonstrate a multimodal analysis to be effective in
estimating tumor burden in a Patient-Derived Mouse Xenografts (PDX) model of GBM. We suggest
that MRI sequences sensitive to the presence of tissue water, tissue water
exchange across microvascular boundaries, and magnetization exchange between
tissue structures and tissue water will generate a tissue signature
characteristic of local tumor burden. Also, MRI measures of vascular volume and
tissue metabolism will reflect local tumor burden. This pilot study
investigates the feasibility of constructing an Adaptive Model (AM) as probabilistic
predictor of infiltrating tumor burden in PDX model of GBM using multiple MRI
imaging.
Material
and Methods:
Eight mice
implanted with GBM CSC HF2927 were studied. MRI studies were performed in a
Direct-Drive-Varian (Santa-Clara, CA) 7Tesla, 20cm bore system with a 24mm
2
field of view. Gradient maximum strengths and rise times were 250mT/m
and 120µs. The following image sets were acquired: high-resolution T1- and T2-weighted
(matrix:256x192, 27slices, 0.5mm thickness, NE=1, NA=4, TE/TR=16/800 and NE=4,TE/TR = (20,40,60,80)/3000 ms);
MT-weighted fast spin-echo, matrix:256x128, 27slices (0.4,0.1 mm gap), TR/TE 2248/11.4, 32 segments, 4 echoes,
NEX=6. Magnevist (0.25mmol/kg i.p) was
injected about 5 minutes before the post-contrast T1-weighted image
set was acquired. The animal was sacrificed immediately after MRI and stained
for the presence of human GBM cells. We used the following MRI sequences to
establish a basis set for training the AM: pre- and post-contrast (Magnevist, I.P.) T1-weighted, 4-echo T2, MT, 3-direction, 3 b-value diffusion-weighted.
Maps of T2 and proton density (T2-PD) were produced by
fitting the T2 data. All MRI
images were skull-stripped manually and co-registered to the T2-PD
map using an affine transform. The histology image was warped and co-registered to the T2-PD image. MR image sets were normalized to the white-matter area of the brain, and each voxel profile extracted from the 9 image set
was normalized to the summation of two normalized T2 images (echo 2
and 3), following which the normalized profile along with the
co-registered histology was used for training and testing of an artificial
neural network (ANN)
4 with Multi-Layer-perceptron architecture to
predict the presence of local tumor burden in form of tumorous cell density. About
6000 samples were used to train the ANN. The Area-Under-the-Receiver-Operating-Characteristic (AUROC) along with K-fold-cross-validation(KFCV) with 20 folds
(about 300 samples-per-fold) technique [4] were used to validate, optimize, and
estimate the predictive power of the ANN.
Results
and Conclusions:
The optimal
ANN(10:3:1) was successfully trained and validated using the 9 MR modalities. Results
imply that the chosen MRI feature-set contains adequate information for
training the ANN. The predictive power of the ANN was ~0.81. Figure 1-A shows
the ANN’s probabilistic prediction of tumor burden using a set of MRI images in
an exemplary case. The correlation coefficient for the association between the
predicted map and histology was ~0.85. Figure 1-B illustrates MHC stain(brown) for human cells. The spread of
the implanted glioma cells from the tumor mass(black dashed-line) below the
injection site(arrow) to the contralateral side of the brain(yellow dashed-line) is evident. The histology
demonstrates that GBM cells were densely present near the injection-site
and had infiltrated broadly throughout the brain. Figure 1-C, shows the post-contrast T1-weighted image. Note the
lack of enhancement in the image. Thus, in a model of infiltrating tumor with
minimal breakdown of the BBB, we have generated a probabilistic estimate of
tumor burden. Given the success of training an ANN to predict infiltrating
tumor burden, it may be possible to identify similar or different basis set of
MRI images that can robustly predict both infiltrating and solid tumor burden
in human GBM.
Acknowledgements
No acknowledgement found.References
1. Baker GJ,
Yadav VN, Motsch S, Koschmann C, Calinescu AA, Mineharu Y, Camelo-Piragua SI,
Orringer D, Bannykh S, Nichols WS, deCarvalho AC, Mikkelsen T, Castro MG,
Lowenstein PR. Mechanisms of glioma formation: iterative perivascular glioma
growth and invasion leads to tumor progression, VEGF-independent
vascularization, and resistance to antiangiogenic therapy. Neoplasia
2014;16(7):543-561.2. Nakano A,
Tani E, Miyazaki K, Yamamoto Y, Furuyama J. Matrix metalloproteinases and
tissue inhibitors of metalloproteinases in human gliomas. J Neurosurg
1995;83(2):298-307.
3. Yong VW,
Power C, Forsyth P, Edwards DR. Metalloproteinases in biology and pathology of
the nervous system. Nat Rev Neurosci 2001;2(7):502-511.
4. Bishop C
(1997) Neural Networks for Pattern Recognition. . London-United Kingdom: Oxford
University Press