Deep Learning for Brain Tumors and Neuroimaging

We are applying state-of-the-art deep learning methods for brain tumor segmentation, prediction of molecular markers in gliomas (IDH, 1p/19q, MGMT), MRI super-resolution imaging, motion correction,  MEG artifact detection, and fMRI analyses. One of the most important recent discoveries in brain tumor biology has been the identification of the gene isocitrate dehydrogenase (IDH) as a marker for glioma therapy and prognosis.  Gliomas with this mutant enzyme have a better prognosis than tumors of the same grade with wild type IDH.  IDH-mutated tumors also have different management and therapeutic approaches than tumors with wild-type mutation status.  The only reliable way of obtaining this information is through direct tissue sampling of the tumor, requiring either a craniotomy and biopsy or a large open surgical resection. An effective non-invasive method of determining IDH-mutation status would be transformational in the management of gliomas. We have developed a deep learning classifier using the TCIA database and T2-weighted (T2w) MR images that can identify IDH mutation status with over 97% accuracy. We are evaluating the robustness of the approach, and extending it to other molecular markers.   A key advantage of our approach is in the use of standard T2-weighted images for the classification.  T2 images are routinely acquired as part of any MRI brain tumor evaluation.  These images are robust to motion, and can be obtained within 2 minutes of the standard clinical MRI study, making clinical translation straightforward with immediate impact on patient care.

Deep Learning for Brain Tumor and NeuroImaging publications