Brain Tumour Segmentation Using cGAN with Residual (ResNet) Technique

A brain tumour is a malformed cell growth that can be either cancerous or non-cancerous. The most dangerous disease is a brain tumour, which can be detected. The tumour can be detected easily and reliably using automated techniques on MRI images. Many researchers have proposed several methods for effective tumour detection, including efficient diagnosis and segmentation of brain tumors. We propose conditional generative adversarial networks (cGAN)-based approach for brain tumour segmentation in this paper. cGAN is made up of two networks: the generator, which generates the brain tumour segmentation map, and the discriminator, which validates the generator output. Encoder-decoder architecture is used in the proposed generator. It encodes the input brain MRI slice into a set of features using the encoder network, and then uses the decoder network to generate the brain tumour segmentation map from the encoded features. To maintain structural consistency, feature maps generated by a specific encoder are shared with the corresponding decoder via skip connections i.e. ResNet. Unlike previous encoder-decoder architectures, we ran the encoder feature maps through a convolution layer before passing them to the appropriate decoder. It aids in the refinement of the encoded features. We trained the proposed network for brain tumour segmentation using the BraTS-15 training set. Its testing set, on the other hand, is used to validate the proposed network for brain tumour segmentation. The experimental analysis compares proposed and existing methods for brain tumour segmentation using the Dice similarity coefficient and the Jaccard index. When compared to existing methods, the proposed method outperforms them all for brain tumour segmentation.