cVGG16-UNet: Multiregional Automatic Segmentation of Glioma Tumor Using Hybrid 3D Deep Convolutional Neural Network

Deep learning is gaining high popularity due to its high predictive power in medical field including cancer diagnosis. Deep learning-based segmentation of brain tumor is an evolving field to discriminate cancerous tissues of different internal regions such as enhancing, necrotic core, and edema. 2D and 3D convolution neural network (CNN) is being widely used for segmentation of intra-tumor regions. However, 3D techniques are comparatively powerful to leverage the context from adjacent slices. The proposed 3D cVGG16-UNet is hybrid model that is architected by combining VGG16 and UNet architectures. 220 patients from BRATS 2020 dataset are used for training and testing the model, consisting of 3D MRI volumes which are pre-processed to achieve the better results. The model has attained the dice similarity coefficient of 0.6799, 0.8463, and 0.7493 for necrotic tumor, enhancing tumor and edema respectively. The work has outperformed especially for enhancing tumor segmentation.