Abstract

Clinics must be able to identify and diagnose brain tumours early. Hence, accurate, effective, and robust segmentation of the targeted tumour region is required. In this article, we suggest a method for automatically segmenting brain tumours using convolutional neural networks (CNNs). Conventional CNNs disregard global region features in favour of local features, which are crucial for pixel detection and classification. Also, a patient’s brain tumour may develop in any area of the brain and take on any size or shape. We created a three-stream framework called multiscale CNNs that could incorporate data from various scales of the regions surrounding a pixel and automatically find the top-three scales of the image sizes. Datasets from the MICCAI 2013-organized Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) are used for both testing and training. The T1, T1-enhanced, T2, and FLAIR MRI images’ multimodal characteristics are also combined within the multiscale CNNs architecture. Our framework exhibits improvements in brain tumour segmentation accuracy and robustness when compared to conventional CNNs and the top two techniques in BRATS 2012 and 2013.