Accurate brain tumor classification with STN-NAM in ResNet50 using MRI

Abstract

Brain tumor is an abnormal cell growth that contains malignant and benign cells emerging from numerous cell types within brain. Magnetic resonance imaging (MRI) is utilized for brain tumor classification which provides high-resolution images. However, tumors exhibit different characteristics like shape, location, and size which make it challenging to accurately distinguish among different tumor types and accurately classify them. In this research, spatial transformer network and non-local attention mechanism (STN-NAM) is proposed in ResNet50 to accurately classify tumors. STN transforms spatial information while NAM identifies relationships among normal and lesion areas, which together accurately classify tumors. Initially, images are obtained from Figshare, Brats 2019, and Brats 2020 datasets. These images are pre-processed using a normalized median filter (NMF) to reduce salt and pepper noise. Then, normalization is performed to resize original image to a standard size which assists uniformity in image dimension. U-Net is employed to segment tumor regions and STN-NAM is performed to accurately classify tumors. In comparison to the existing techniques namely, multi-level attention network (MANet), mathematical model with 3D attention U-Net, and convolutional neural network (CNN), the STN-NAM achieves superior accuracy of 98.06%, 99.05%, and 98.66% in Figshare, Brats 2019, and Brats 2020 datasets, respectively.