Room: Room 202
Purpose: To develop a machine-learning based reconstruction strategy for robust and accurate Diffusion Tensor Magnetic Resonance Imaging (DTI) that can quantitatively characterize intracranial tumors with shorter scan times than previously achievable.
Methods: Neurological DTI were acquired in healthy subjects (N=10) on a 3.0T Siemens Prisma MRI scanner (20 encoding directions, 5 averages, scan time: 4 minutes). â€œGold standardâ€? diffusion tensors were then reconstructed using linear least squares (LLS) fitting and fractional anisotropy (FA) values were calculated from these tensors at each voxel. An Artificial Neural Network (ANN) was trained to reconstruct the â€œgold standardâ€? FA at each voxel given subsets of the raw DTI signal as input (3-20 encoding directions, 1 average) which corresponded to shorter scans (9-48 seconds). Training was performed using a leave-one-out technique in which a model was trained for each of the ten subjects using the data from the other nine. ANN FA reconstructions were then compared with LLS for each under-sampled set in terms of accuracy and precision.A patient with a glioblastoma multiforme (GBM) was also scanned on a 1.5T scanner (Siemens Espree) using the protocol described above. FA maps were reconstructed using LLS and the ANN (trained from healthy subject data) with and without under-sampling. FA values were then compared in regions containing tumor, necrosis, and remote white matter.
Results: The ANN reconstruction improved FA accuracy and precision compared with LLS for fixed scan time. Improvement increased for shorter scans. The ANN also permitted equivalent reconstruction with 44% shorter scans than what is possible with LLS.In GBM, the ANN improved tumor-to-necrosis contrast-to-noise ratios by >30% and permitted effective delineation with highly accelerated scans.
Conclusion: The proposed ANN strategy achieved superior DTI reconstructions and enabled shorter scans than traditional LLS fitting. This translated to improved delineation between tumor and necrosis in a GBM.