An Efficient Brain Tumor Segmentation Approach Using Cascade Convolutional Neural Networks

Abstract

Brain tumours pose a significant threat to human life, as they are a major cause of death. Early detection of brain tumours is vital to improve treatment and reduce mortality rates. Automatic segmentation using deep learning methods is crucial for clinical evaluation and treatment planning but remains challenging due to the diverse tumour locations and structures. In this work, we employed the concatenation of two different convolutional neural networks: the two-pathway architecture and the inception architecture. We also utilised a patch-based technique that combines global and local features to predict the output region. Our proposed system achieved dice scores of 0.86, 0.81, and 0.75 for the whole tumour, tumour core, and enhancing tumour on the BraTS 2018 dataset, respectively. For BraTS 2019, the dice scores were 0.85, 0.79, and 0.67, respectively. Compared to existing state-of-the-art CNN models, our proposed system significantly improves both qualitative and quantitative brain tumour segmentation results.