Abstract 466: MR prediction of tumor burden in Patient-Derived Mouse Xenografts model of glioblastoma using an adaptive model

Introduction: In human glioblastoma multiforme (GBM), infiltrating cells are found in remote locations, even in the hemisphere contralateral to the primary lesion. Magnetic Resonance Imaging (MRI) allows approximation of the extent of tumor cell infiltration. However, the actual extent of infiltration may be greater or less than the edema, and there is no standard MRI practice that allow exploring 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. Material and Methods: 8 mice implanted with GBM CSC HF2927 were studied. MRI studies were performed in a Direct Drive Varian 7 Tesla. The following image sets were acquired: high-res. T1-weighted; T2-weighted (TE/TR = (20, 40, 60, 80)/3000 ms); MT-weighted fast spin-echo. Magnevist (0.25 mmol/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. The histology image was warped and co-registered to the T2-PD image using mouse brain anatomical landmarks. 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) with Multi-Layer perceptron architecture to predict the presence of local tumor burden in form of tumorous cell density. Results and Conclusions: The ANN was successfully trained and validated using K-Fold Cross Validation technique (KFCV) and the 9 MR modalities as its input set. Results imply that the chosen MRI feature set contains adequate information content for training the ANN. The predictive power of the ANN was ∼ 0.81. The correlation coefficient for the association between the predicted map and histology was ∼ 0.85. Given the success of training an ANN to predict infiltrating tumor burden, it may be possible to ide