Chapter 22 – Multi-Modality Brain Image Classification With Deep Convolutional Neural Networks

Author: Dr. R. Rajasree, Assistant Professor, Department of Electronics and Communication Engineering, PSN Institute of Technology and Science, Tirunelveli, Tamil Nadu, India

Editors: Dr. Thirumurugan Shanmugam and Dr. Shweta A. Bansal

ISBN: 978-81-963849-7-5

DOIhttps://doi.org/10.59646/csebookc22/004

Abstract:

The brain is one of the important and complex organs that controls all the metabolic activities of human body. Tumor is formed in the brain due to unwanted growth of tissue. Multimodal Brain Tumor detection is a significant process to avoid mortality and severe illness. Multimodal brain tumor detection is a significant process to diagnose the presence of the tumor. In this work, the semantic type segmentation method is proposed to detect the multimodal brain tumor. The proposed multimodal semantic image segmentation for the classification and detection of brain tumors uses deep learning Multimodal Convolutional Neural Network (MMCNN). This network proposes a Bi-directional Long Short-Term Memory (BiLSTM) based segmentation method to test the integration of segmentation, feature extraction and classification. The effectiveness of the suggested methods to detect the brain tumor is tested with different scale sizes such as (12∗12, 24∗24, and 48∗48), (64∗64, 128∗128, 240∗240). In all the three scale sizes, the proposed method exhibits an increase in the accuracy and dice coefficient parameter. The multimodal tumor Magnetic Resonance Imaging (MRI) segmentation performance is improved by combining all the models pixel information retrieved from T1, T2, T1c and FLAIR different tumor modality images. Evaluation is done with the help of BRATS15 dataset. The proposed method attains the overall classification accuracy of 95.13% and the sensitivity of 0.9068. This is a better level of prediction that can lead to efficient semantic image segmentation over other algorithms that are utilized in handling higher data volumes. A five-fold cross-validation scheme is used in this work for validated the MRI BRATS2015 dataset.