Room: ePoster Forums
Purpose: Brain tumor segmentation (BTS) plays a significant role in disease diagnosis, treatment planning and surgical guidance. The feature extraction by deep convolutional neural network (CNN) is investigated to segment brain tumors, thereby achieving automatic segmentation instead of conventional manual segmentation.
Methods: The total 200 patient images from BRATS database are used for training and testing. Four MRI sequences including T1-weighted, T1 with gadolinium enhancing contrast, T2-weighted and FLAIR, are for each patient. The matlab based Matconvnet toolbox is employed to construct the CNN model. In the CNN method, the patches centering on each pixel were extracted and considered as the training data. Meanwhile, the BTS gold standard was referenced. The extracted patches were divided into positive and negative samples by the central pixel lables. The positive sample regarded as the tumor region, whereas the negative sample denotes the normal tissue. The patches were considered as the network input firstly, and then were trained with CNN-Softmax structure or CNN structure, respectively. Finally, the obtained high-level features by the network were preprocessed and input into the classifier, thereby obtaining the probability value belonging to a tumor or normal tissue of the brain. According to the magnitude of the category probability, a segmented binary image of the tumor was achieved.
Results: The preliminary results show that the CNN-Softmax structure is superior to the only-CNN structure for training. Compared with the latter, the evaluation standard coefficients of CNN-Softmax, such as Dice Similarity Coefficient (DSV), Positive Predictive Value (PPV) and Sensitivity values are all improved. The values of DSV, PPV and Sensitivity are up to 0.90.
Conclusion: A simple and effective method based on CNN was proposed for MRI BTS.The CNN-Softmax method is closer to the contour of the tumor in Ground Truth than the CNN, and can be easily adapted to difference tumors.
Funding Support, Disclosures, and Conflict of Interest: This study was supported by the university student innovation project of Hefei University of Technology (2018CXCYS209).