Abstract
Cancers of the central nervous system, especially brain tumors, are among the leading causes of death worldwide. Magnetic Resonance Imaging (MRI) plays a crucial role in detecting and diagnosing brain tumors due to the high degree of soft tissue contrast achievable with the technique. However, early diagnosis of brain tumors is still difficult to achieve because their growth is complex and subtle in the early stage. Challenges include data scarcity, limited image clarity and difficulties in accurate classification. To tackle these issues, this work proposes a novel method for efficient and precise detection of brain tumors. Tensor compression is used to reduce the size of high-dimensional MRI image datasets while preserving essential spatial and structural content, which increases efficiency without compromising accuracy. Furthermore, an efficient sub-pixel CNN for image super-resolution is used to improve the spatial resolution of the MRI scans, enabling important minute details to be captured better. To overcome the scarcity of medical data available and the challenge of overfitting, MixUp data augmentation is used. This technique generates synthetic training samples by combining images and their corresponding labels, improving the robustness of the model. Moreover, the External Attention Transformer (EANet) enhances accuracy of tumor localization and classification by improving global feature extraction by attending to the relevant parts of the MRI scans for focused attention. All these methods work together to deliver a highly efficient and effective brain tumor detection system.
Keywords: Brain Tumor, EANet Classification, Image Super-resolution, Magnetic Resonance Imaging (MRI), MixUp Augmentation.