Since 1990s first methods for automated segmentation of brain tumors have been published, but all of them were validated on small, private datasets, until 2012, when MICCAI organized a Multimodal Brain Tumor Image Segmentation Challenge (BraTS) releasing publicly available dataset, consisting 4 modalities of MRI images: T1, T1 with contrast (gadolinium), T2 and FLAIR with manually segmented tumors.
Recent improvements in neural networks training performance by taking advantage of high performance CPUs, GPUs and ReLU activation function have led to usage of convolutional neural networks for sematic segmentation. The examples of successfully used architectures are: 2015 U-Net  and 2016 100-layer Tiramisu DenseNet . Especially, Tiramisu has shown great performance on sematic segmentation of urban scene benchmarks. It contains up-paths and up-paths, but also Dense blocks with skip-paths include Concatenation of feature maps from output of Convolutional layer along with its input. Each hidden layer uses ReLU activation and Batch Normalization and L2 regularization. Different variants of Tiramisu consist of various number of layers per Dense block (usually 4÷5, up to 12) and growth factor which defines how fast number of feature layers growths after each Concatenation (usually k=12÷15). Rise of these parameters causes rapid increase of trainable parameters.
We have used variant of Tiramisu Densenet with k=8 and number of layers per Dense Block = [4,5,7,10,12,15] with 9.3 mln of trainable parameters build in Keras with Tensorflow backend. Training dataset have been prepared by cropping each slice to 192x160 pixels, to eliminate redundant background reducing amount of needed memory and standardized by Z-Score. Network have been trained with categorical cross-entropy loss and RMSprop optimizer with learning rate 1e-3. Training on Nvidia GTX1080 took approximately 2 days.
Resulting internal measure of average Dice score on training dataset is exemplary: 0.890 for Whole Tumor, 0.798 for Tumor Core and 0.801 for Enhancing Tumor. NN have been also tested on whole BraTS validation dataset and its median Dice scores are: 0.882 for WT, 0. 751 for TC and 0.778 for ET. We will continue our research to improve segmentation quality, that is already at the level of the state of the art methods. We recently discovered corruption of considerable amount of images due to improper registration process, that could cause some underperformance of NN training and in effect in final segmentation. To correct bias fields, we plan to apply N4ITK algorithm. There is probably also a huge margin for hyper-parameter tuning, which we plan to apply in the coming future.
References: 1. Ronneberger O., Fischer P., Brox T., U-Net: Convolutional Networks for Biomedical Image Segmentation, Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, Vol.9351: 234--241, 2015 2. Jégou S., Drozdzal M., Vazquez D., Romero A., Bengio Y., The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation, IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017
This project has been supervised by Franciszek Binczyk (Silesian University of Technology, Institute of Automatic Control, Data Mining Group 44-100 Gliwice, ul. Akademicka 16, Poland). This work was financed by grant no. 02/010/BK17/0060/9; BK-204/RAU1/2017/9. Additionally, FB was financed by project UMO-2015/19/B/ST6/01736.