The multimodal brain tumor image segmentation based on convolutional neural networks

In this paper, we propose an automatic brain tumor segmentation algorithm based on a 22-layers deep, three-dimensional Convolutional Neural Network (CNN) for the challenging problem of gliomas segmentation. To correct the bias field distortion of MRI images, we have added N4ITK method before intensity normalization. The use of several cascaded convolution layers with small kernels allows building a deeper CNN. During training, we add dropout to reduce overfitting, and adapt the batch normalization technique to speed up training. Our proposed method has been validated in the BRATS 2015 databases, the performance for the complete, core and enhancing regions was evaluated by the online evaluation platform with Dice matric of 0.84, 0.79, 0.75, Positive Predictive Value matric of 0.88, 0.86, 0.70 and Sensitivity matric of 0.82, 0.75, 0.86. Due to the GPU acceleration and improvement of the algorithm, our total training time is about 140 min. And through the experiments and results, we can prove that our method is time-saving and efficient.