Abstract
INTRODUCTION
Brain tumors are difficult to segment in MRI scans. Consequently, we have developed a new system for completely automatic brain tumor segmentation by combining a state-of-the-art 3D deep convolutional neural network (CNN) with a large collection of curated segmentations of brain tumors.
METHODS
Our brain tumor database holds 74,722 MRI series from 2,742 unique patients. Over the last 15 years our image analysis team has segmented brain tumors in 35,710 of these series. This preliminary experiment identified 741 pre-treatment studies that included a T1GD and FLAIR scan, and at least one adjudicated brain tumor segmentation. These studies were randomly assigned into 600 training, 41 validation, and 100 test cases. CNN training was performed in two stages: 1) 50 epochs on minimally modified MRI volumes; and, 2) 24 epochs to tune the CNN on skull-stripped volumes. Whole-tumor Dice coefficients (1=perfect overlap, 0=no overlap) were calculated by comparing CNN segmentations against adjudicated segmentations from trained measurers. Training was performed in-the-cloud using an Amazon Machine Instance equipped with an NVidia Tesla V100 GPU, 8 Intel Xeon processors, and 64 GB of RAM.
RESULTS
Training required 74 hours. Afterwards, our network required 800 seconds to segment 100 studies in the test set (8 seconds/study). The mean whole-tumor Dice coefficient on the test studies was 0.885. DISCUSSION: The best result on the highly cited 2017 BraTS brain tumor segmentation challenge was a whole-tumor Dice of 0.886, achieved by an ensemble of 7 CNNs. BraTS included 274 studies, each with T1GD, FLAIR, T1 and T2 contrasts. The performance of our single CNN may be due to our comparatively large training set. Our goal is to train our CNN on all series in our database. This may provide useful tools to monitor each patients journey, from diagnosis through treatment.