Segmenting brain tumors from FLAIR MRI using fully convolutional neural networks

Abstract Background and Objective Magnetic resonance imaging (MRI) is an indispensable tool in diagnosing brain-tumorpatients. Automated tumor segmentation is being widely researched to accelerate the MRI analysis and allow clinicians to precisely plan treatment—accurate delineation of brain tumorsis a critical step in assessing their volume, shape, boundaries, and other characteristics. However, it is still a very challenging task due to inherent MR data characteristics and high variability, e.g., in tumor sizes or shapes. We present a new deep learning approach for accurate brain tumorsegmentation which can be trained from small and heterogeneous datasets annotated by a human reader (providing high-quality ground-truth segmentation is very costly in practice). Methods In this paper, we present a new deep learning technique for segmenting brain tumorsfrom fluid attenuation inversion recoveryMRI. Our technique exploits fully convolutional neural networks, and it is equipped with a battery of augmentation techniques that make the algorithm robust against low data quality, and heterogeneity of small training sets. We train our models using only positive (tumorous) examples, due to the limited amount of available data. Results Our algorithm was tested on a set of stage II-IV brain-tumorpatients (image data collected using MAGNETOM Prisma 3T, Siemens). Rigorous experiments, backed up with statistical tests, revealed that our approach outperforms the state-of-the-art approach (utilizing hand-crafted features) in terms of segmentation accuracy, offers very fast training and instant segmentation (analysis of an image takes less than a second). Building our deep model is 1.3 times faster compared with extracting features for extremely randomized trees, and this training time can be controlled. Finally, we showed that too aggressive data augmentation may lead to deteriorated performance of the model, especially in the fixed-budget training (with maximum numbers of training epochs). Conclusions Our method yields the better performance when compared with the state of the art method which utilizes hand-crafted features. In addition, our deep network can be effectively applied to difficult (small, imbalanced, and heterogeneous) datasets, offers controllable training time, and infers in real-time.