Abstract
Identifying the tumour’s extent is a major challenge in planning treatment for brain tumours and correctly measuring their size. Magnetic resonance imaging (MRI) has emerged as a first-line diagnostic method for brain malignancies. Manually segmenting the extent of a brain tumour from 3D MRI volumes is a time-consuming job that heavily relies on the operator’s knowledge. Computer-aided tumour detection techniques and deep learning have significantly improved machine learning. So, in this paper, we proposed a modified U-Net structure based on residual networks that use shuffling periodically at the encoder section of the original U-Net and sub-pixel convolution at the decoder section. Sub-pixel convolution has the benefit over conventional resizing convolution in that it has extra parameters and thus stronger modelling capability at the same computing complexity and avoids de-convolution overlapping. The proposed U-Net model is evaluated on two benchmark datasets such as brain tumour segmentation (BraTS) Challenge 2017, 2018, and with segmentation accuracies of 93.40% and 92.20%, respectively. Also, the tumour sub regions were classified into three categories: tumour core (TC), whole tumour (WT), and enhancing core (EC). The results of the tests revealed that the suggested U-Net outperforms the existing approaches.
•Modified encoder-decoder structure for the segmentation of brain tumours.•In the encoder section, a transfer learning ResNet-34 to encode the visual features.•U-Net model replaces ReLU with leaky ReLU, to speeds up the training process.•BraTS 2017 and BraTS 2018 are utilised to classify tumours effectively.