Purpose:Accurately depicting the radiation target is critical for a high quality plan of tumor radiotherapy. The manual segmentation is a huge amount of work and has a strong artificial subjectivity. Using deep learning method to assist automatically segmenting of tumor target is the penetration and application of artificial intelligence technology in the field of medicine.
Methods:A 6-layer deep Convolution Neural Network (CNN) model has been constructed by taking advantage of different types of layers for brain tumor segmentation. This model is a 6 layer CNN model (6-CNN) composed of three convolution layers, two pool layers and one full connection layer. To obtain enough samples for 6-CNN model training, a patch-based technology has been adopted. That is to successively extract a local area from the whole image as a patch. and the center pixel value is taken as the pixel value of the whole patch. Similarly, the label of the center pixel is also taken as the label of the whole patch. Thus the 6-CNN model transforms the brain tumor image segmentation into patch classification based on the excellent classification characteristics of deep convolution neural network. The model combines the local features of patch, the information extracted from shallow network and the global features to predict the category label of the central pixel of patch.
Results:The model is validated on BRATS 2015 dataset and the result shows that the segmentation accuracy can be up to Dice Similarity Coefficient (DSC) 90%±4%.
Conclusions:An automatic deep CNN segmentation model for brain tumors based on MRI image patches has been constructed, which is expected to assist or even substitute the manual segmentation of brain tumors.