Cascading handcrafted features and Convolutional Neural Network for IoT-enabled brain tumor segmentation

Abstract

The Internet of Things (IoT) has revolutionized the medical world by facilitating data acquisition using various IoT devices. These devices generate the data in multiple forms including text, images, and videos. Given this, the extraction of accurate and useful information from the massive serge IoT generated data is a highly challenging task. Recently, the brain tumor segmentation from IoT generated images has emerged as a promising issue that requires sophisticated and efficient techniques. The accurate brain tumor segmentation is challenging due to large variations in tumor appearance. Existing methods either use handcrafted features based techniques or Convolutional Neural Network (CNN). In this paper, a novel cascading approach for fully automatic brain tumor segmentation has been proposed, which intelligently combines handcrafted features and CNN. First, three handcrafted features are computed namely mean intensity, Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) and then Support Vector Machine (SVM) is employed to perform pixel classification that results in Confidence Surface Modality (CSM). This CSM along with the given Magnetic Resonance Imaging (MRI) is fed to a novel three pathways CNN architecture. In the experiments on BRATS 2015 dataset, the proposed method achieves promising results with Dice similarity scores of 0.81, 0.76 and 0.73 on complete, core and enhancing tumor, respectively.