An Automated Hybrid Transfer Learning System for Detection and Segmentation of Tumor in MRI Brain Images with UNet and VGG-19 Network

Abstract

In the face of the heterogeneous properties exhibited in the facade, amount and the shape of tumors, brain tumor segmentation in the context of medical imaging is being recognized as a critical issue. Features from the input images are extracted by using CNN (convolutional neural networks) for the purpose of classification. Physical discovery and segmentation of brain tumor or Glioma is tedious due to the nature of asymmetry of shape, lenient nature of location and potholed margins. The proposed framework aims at developing a model to perform the tasks like detection and segmentation of brain tumor based on transfer learning along with deep convolutional neural network (DCNN). Transfer Learning (TL) along with fully convolutional network on VGG-19 model is deployed to perform the tasks like detection of tumor and the segmentation in an automatic fashion. With the proposed transfer learning based model, three categories of brain tumor can be identified, namely: normal, glioma at the levels of low grade and high grade which, in turn, are respectively mentioned as LGG and HGG. BRaTS 2015 database is used in the proposed hybrid model for the purpose of evaluation, and by its exhibiting accuracy while performing training of data, validation as well as testing, it seems the proposed transfer learning based network yields better results compared to the existing techniques.