NIMG-73. ACCURATE 3D GBM SEGMENTATION IN 30 SECONDS

Abstract

Many clinical trials of Glioblastoma (GBM) therapies use changes in tumor burden to determine treatment response. Often, tumor size is estimated from bi-orthogonal lines placed to cover the largest cross-sectional extent of the lesion. This method is simple, rapid, and widely used. However, it is plagued by low accuracy. Consequently, many are developing more accurate algorithms to directly label, or “segment” the 3D extent of GBM tumors in MRI exams. The BRain Tumor Segmentation (BRATS) challenge allows objective comparison of these algorithms. BRATS includes anonymized MRI exams from patients with GBMs. Each exam includes a “true” tumor boundary determined from manual outlining by experts. A “leader board” of the most accurate algorithms is maintained on the BRATS website. However, leading algorithms often require long computation times (> 80 minutes in at least one case). We present a new method to segment brain tumors from multi-spectral MRI exams. Our approach leverages the speed of a graphical processing unit (GPU) level set algorithm to allow rapid and simple brain tumor segmentation. We implement a new data term in the level set speed function. This term eliminates the need to store MR exams on the GPU memory, as in prior approaches. Instead, a fixed size data and seed dependent lookup table is employed. This allows multi-spectral MR exams that are too large to fit on the GPU memory to be processed efficiently. We evaluated our method on the BRATS 2013 data. Our algorithm achieved an overall accuracy (Dice coefficient) on whole GBM segmentation of 83.5%. This is comparable to top performing algorithms. The average segmentation time was 30.8 seconds. Current limitations include the need for a user to place seed regions in each exam. Future work will focus on automating this step using machine-learning based approaches.