G. Rasool
Rowan University,
United States
Keywords: deep learning, brain tumors, segmentation, MRI
Summary:
Radiation treatment for brain and other tumors is the standard of care in oncology clinics. The efficacy of radiation treatment relies heavily on the accurate delineation (segmentation) of the tumor region, which, in turn, serve as a target for radiation treatment. Currently, radiation oncologists and dosimetrists manually delineate tumors and healthy tissue which is a time-consuming process. Furthermore, structural changes, spatial variations, and intensity inhomogeneity in MR images could make accurate delineation very challenging. We provide accurate segmentation of brain tumors using a novel state-of-the-art U-Net algorithm. The inputs for the segmentation algorithms are multimodal MRI data, i.e., T1, T1c, T2, and FLAIR (images recorded with the field strength of 3T). We are using the Inception modules as base networks for extraction of features. The extracted features are used for the classification, localization, and segmentation of regions of the tumor, i.e., edema and three substructures of the tumor core including a non-enhancing solid core, necrotic/cystic core, and enhancing core. Our algorithm provides more than 95% of brain tumor segmentation accuracy on the multimodal brain tumor segmentation (BraTS) dataset.