Transfer Learning Based Brain Tumor Detection and Segmentation using Superpixel Technique

Manual detection and segmentation of glioma is difficult because of its asymmetrical shape, flexibility in location and uneven boundaries. The proposed work presents a transfer learning-based brain tumor detection and segmentation is done using superpixel technique. In the initial phase, brain images are classified into three categories namely, normal, Low-Grade Glioma (LGG) and High-Grade Glioma (HGG) based on the availability of tumor. The proposed methodology is examined on Brain Tumor Segmentation (BraTS) 2019 challenge database. The tumor detection task is performed on VGG-19 transfer learning model. On the VGG-19 transfer learning model at epoch 6, training data yield 99.82% training accuracy, 96.32% validation accuracy and 99.30% testing accuracy. The obtained specificity is 100% and sensitivity is 97.81%, and Area Under Curve (AUC) is 0.99. In the second phase, segment the tumor within LGG and HGG images using the superpixel segmentation technique. Superpixel segmentation approach leads to an average detection dice index of 0.932 against the ground truth data. The experimental results prove the superiority of the proposed methodology in comparison to existing methods.