Segmentation of Brain Tumor from Magnetic Resonance Imaging using Handcrafted Features with BOA-based Transformer

Abstract

In order to improve patients’ chances of survival and prognosis, early detection of brain tumors is essential. This task requires the physical analysis of magnetic resonance imaging (MRI) images of brain tumors. Consequently, more accurate tumor diagnosis necessitates computational methods. Shape, volume, boundaries, size, tumor identification, segmentation, and classification evaluations continue to be tough, nonetheless. Cancer features also make correct segmentation difficult, including fuzziness, complicated backgrounds, and substantial variations in size, shape, and intensity distribution. To lecture these issues, this work proposes a new Optimizer based Semantic-Aware Transformer (OSAT) for segmenting brain tumors. In addition, features based on intensity, texture, besides shape were manually retrieved from MRI data. With less memory and computational complexity, the Bonobo optimization algorithm (BOA) fine-tunes SAT, enhancing the ability of feature representation learning. Segmentation measures were among the many evaluation metrics utilized to evaluate performance in this work across the three Brain Tumor Segmentation (BraTS) challenge datasets. A more robust and generalizable solution was also obtained by improving OSAT’s performance with the addition of handcrafted features. When it comes to efficient and accurate brain tumor segmentation, this research could have major practical implications. Exploring different feature fusion methods and adding more imaging modalities to enhance the effectiveness of the projected technique are potential areas for future research.